Mirrors/the-algorithm
2
0
Fork 0
mirror of https://github.com/twitter/the-algorithm.git synced 2024-06-13 06:38:52 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' into main

Browse Source
This commit is contained in:
Erfan Safari 2023-04-29 00:53:57 +03:30 committed by GitHub
commit f9b7db14c6
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
614 changed files with 54035 additions and 312 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

51
README.md
View File

@ -1,22 +1,42 @@
# Twitter's Recommendation Algorithm
Twitter's Recommendation Algorithm is a set of services and jobs that are responsible for constructing and serving the
Home Timeline. For an introduction to how the algorithm works, please refer to our [engineering blog](https://blog.twitter.com/engineering/en_us/topics/open-source/2023/twitter-recommendation-algorithm). The
diagram below illustrates how major services and jobs interconnect.
Twitter's Recommendation Algorithm is a set of services and jobs that are responsible for serving feeds of Tweets and other content across all Twitter product surfaces (e.g. For You Timeline, Search, Explore). For an introduction to how the algorithm works, please refer to our [engineering blog](https://blog.twitter.com/engineering/en_us/topics/open-source/2023/twitter-recommendation-algorithm).
![](docs/system-diagram.png)
## Architecture
These are the main components of the Recommendation Algorithm included in this repository:
Product surfaces at Twitter are built on a shared set of data, models, and software frameworks. The shared components included in this repository are listed below:
| Type | Component | Description |
|------------|------------|------------|
| Data | [unified-user-actions](unified_user_actions/README.md) | Real-time stream of user actions on Twitter. |
| | [user-signal-service](user-signal-service/README.md) | Centralized platform to retrieve explicit (e.g. likes, replies) and implicit (e.g. profile visits, tweet clicks) user signals. |
| Model | [SimClusters](src/scala/com/twitter/simclusters_v2/README.md) | Community detection and sparse embeddings into those communities. |
| | [TwHIN](https://github.com/twitter/the-algorithm-ml/blob/main/projects/twhin/README.md) | Dense knowledge graph embeddings for Users and Tweets. |
| | [trust-and-safety-models](trust_and_safety_models/README.md) | Models for detecting NSFW or abusive content. |
| | [real-graph](src/scala/com/twitter/interaction_graph/README.md) | Model to predict the likelihood of a Twitter User interacting with another User. |
| | [tweepcred](src/scala/com/twitter/graph/batch/job/tweepcred/README) | Page-Rank algorithm for calculating Twitter User reputation. |
| | [recos-injector](recos-injector/README.md) | Streaming event processor for building input streams for [GraphJet](https://github.com/twitter/GraphJet) based services. |
| | [graph-feature-service](graph-feature-service/README.md) | Serves graph features for a directed pair of Users (e.g. how many of User A's following liked Tweets from User B). |
| | [topic-social-proof](topic-social-proof/README.md) | Identifies topics related to individual Tweets. |
| | [representation-scorer](representation-scorer/README.md) | Compute scores between pairs of entities (Users, Tweets, etc.) using embedding similarity. |
| Software framework | [navi](navi/README.md) | High performance, machine learning model serving written in Rust. |
| | [product-mixer](product-mixer/README.md) | Software framework for building feeds of content. |
| | [timelines-aggregation-framework](timelines/data_processing/ml_util/aggregation_framework/README.md) | Framework for generating aggregate features in batch or real time. |
| | [representation-manager](representation-manager/README.md) | Service to retrieve embeddings (i.e. SimClusers and TwHIN). |
| | [twml](twml/README.md) | Legacy machine learning framework built on TensorFlow v1. |
The product surface currently included in this repository is the For You Timeline.
### For You Timeline
The diagram below illustrates how major services and jobs interconnect to construct a For You Timeline.
![](docs/system-diagram.png)
The core components of the For You Timeline included in this repository are listed below:
| Type | Component | Description |
|------------|------------|------------|
| Feature | [SimClusters](src/scala/com/twitter/simclusters_v2/README.md) | Community detection and sparse embeddings into those communities. |
| | [TwHIN](https://github.com/twitter/the-algorithm-ml/blob/main/projects/twhin/README.md) | Dense knowledge graph embeddings for Users and Tweets. |
| | [trust-and-safety-models](trust_and_safety_models/README.md) | Models for detecting NSFW or abusive content. |
| | [real-graph](src/scala/com/twitter/interaction_graph/README.md) | Model to predict the likelihood of a Twitter User interacting with another User. |
| | [tweepcred](src/scala/com/twitter/graph/batch/job/tweepcred/README) | Page-Rank algorithm for calculating Twitter User reputation. |
| | [recos-injector](recos-injector/README.md) | Streaming event processor for building input streams for [GraphJet](https://github.com/twitter/GraphJet) based services. |
| | [graph-feature-service](graph-feature-service/README.md) | Serves graph features for a directed pair of Users (e.g. how many of User A's following liked Tweets from User B). |
| Candidate Source | [search-index](src/java/com/twitter/search/README.md) | Find and rank In-Network Tweets. ~50% of Tweets come from this candidate source. |
| | [cr-mixer](cr-mixer/README.md) | Coordination layer for fetching Out-of-Network tweet candidates from underlying compute services. |
| | [user-tweet-entity-graph](src/scala/com/twitter/recos/user_tweet_entity_graph/README.md) (UTEG)| Maintains an in memory User to Tweet interaction graph, and finds candidates based on traversals of this graph. This is built on the [GraphJet](https://github.com/twitter/GraphJet) framework. Several other GraphJet based features and candidate sources are located [here](src/scala/com/twitter/recos). |
@ -26,11 +46,10 @@ These are the main components of the Recommendation Algorithm included in this r
| Tweet mixing & filtering | [home-mixer](home-mixer/README.md) | Main service used to construct and serve the Home Timeline. Built on [product-mixer](product-mixer/README.md). |
| | [visibility-filters](visibilitylib/README.md) | Responsible for filtering Twitter content to support legal compliance, improve product quality, increase user trust, protect revenue through the use of hard-filtering, visible product treatments, and coarse-grained downranking. |
| | [timelineranker](timelineranker/README.md) | Legacy service which provides relevance-scored tweets from the Earlybird Search Index and UTEG service. |
| Software framework | [navi](navi/README.md) | High performance, machine learning model serving written in Rust. |
| | [product-mixer](product-mixer/README.md) | Software framework for building feeds of content. |
| | [twml](twml/README.md) | Legacy machine learning framework built on TensorFlow v1. |
We include Bazel BUILD files for most components, but not a top-level BUILD or WORKSPACE file.
## Build and test code
We include Bazel BUILD files for most components, but not a top-level BUILD or WORKSPACE file. We plan to add a more complete build and test system in the future.
## Contributing

51
RETREIVAL_SIGNALS.md Normal file
View File

@ -0,0 +1,51 @@
# Signals for Candidate Sources
## Overview
The candidate sourcing stage within the Twitter Recommendation algorithm serves to significantly narrow down the item size from approximately 1 billion to just a few thousand. This process utilizes Twitter user behavior as the primary input for the algorithm. This document comprehensively enumerates all the signals during the candidate sourcing phase.
| Signals | Description |
| :-------------------- | :-------------------------------------------------------------------- |
| Author Follow | The accounts which user explicit follows. |
| Author Unfollow | The accounts which user recently unfollows. |
| Author Mute | The accounts which user have muted. |
| Author Block | The accounts which user have blocked |
| Tweet Favorite | The tweets which user clicked the like botton. |
| Tweet Unfavorite | The tweets which user clicked the unlike botton. |
| Retweet | The tweets which user retweeted |
| Quote Tweet | The tweets which user retweeted with comments. |
| Tweet Reply | The tweets which user replied. |
| Tweet Share | The tweets which user clicked the share botton. |
| Tweet Bookmark | The tweets which user clicked the bookmark botton. |
| Tweet Click | The tweets which user clicked and viewed the tweet detail page. |
| Tweet Video Watch | The video tweets which user watched certain seconds or percentage. |
| Tweet Don't like | The tweets which user clicked "Not interested in this tweet" botton. |
| Tweet Report | The tweets which user clicked "Report Tweet" botton. |
| Notification Open | The push notification tweets which user opened. |
| Ntab click | The tweets which user click on the Notifications page. |
| User AddressBook | The author accounts identifiers of the user's addressbook. |
## Usage Details
Twitter uses these user signals as training labels and/or ML features in the each candidate sourcing algorithms. The following tables shows how they are used in the each components.
| Signals | USS | SimClusters | TwHin | UTEG | FRS | Light Ranking |
| :-------------------- | :----------------- | :----------------- | :----------------- | :----------------- | :----------------- | :----------------- |
| Author Follow | Features | Features / Labels | Features / Labels | Features | Features / Labels | N/A |
| Author Unfollow | Features | N/A | N/A | N/A | N/A | N/A |
| Author Mute | Features | N/A | N/A | N/A | Features | N/A |
| Author Block | Features | N/A | N/A | N/A | Features | N/A |
| Tweet Favorite | Features | Features | Features / Labels | Features | Features / Labels | Features / Labels |
| Tweet Unfavorite | Features | Features | N/A | N/A | N/A | N/A |
| Retweet | Features | N/A | Features / Labels | Features | Features / Labels | Features / Labels |
| Quote Tweet | Features | N/A | Features / Labels | Features | Features / Labels | Features / Labels |
| Tweet Reply | Features | N/A | Features | Features | Features / Labels | Features |
| Tweet Share | Features | N/A | N/A | N/A | Features | N/A |
| Tweet Bookmark | Features | N/A | N/A | N/A | N/A | N/A |
| Tweet Click | Features | N/A | N/A | N/A | Features | Labels |
| Tweet Video Watch | Features | Features | N/A | N/A | N/A | Labels |
| Tweet Don't like | Features | N/A | N/A | N/A | N/A | N/A |
| Tweet Report | Features | N/A | N/A | N/A | N/A | N/A |
| Notification Open | Features | Features | Features | N/A | Features | N/A |
| Ntab click | Features | Features | Features | N/A | Features | N/A |
| User AddressBook | N/A | N/A | N/A | N/A | Features | N/A |

5
navi/README.md
View File

@ -31,6 +31,11 @@ In navi/navi, you can run the following commands:
- `scripts/run_onnx.sh` for [Onnx](https://onnx.ai/)
Do note that you need to create a models directory and create some versions, preferably using epoch time, e.g., `1679693908377`.
so the models structure looks like:
models/
-web_click
- 1809000
- 1809010
## Build
You can adapt the above scripts to build using Cargo.

7
navi/dr_transform/Cargo.toml
View File

@ -3,7 +3,6 @@ name = "dr_transform"
version = "0.1.0"
edition = "2021"
# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
[dependencies]
serde = { version = "1.0", features = ["derive"] }
serde_json = "1.0"
@ -12,7 +11,6 @@ bpr_thrift = { path = "../thrift_bpr_adapter/thrift/"}
segdense = { path = "../segdense/"}
thrift = "0.17.0"
ndarray = "0.15"
ort = {git ="https://github.com/pykeio/ort.git", tag="v1.14.2"}
base64 = "0.20.0"
npyz = "0.7.2"
log = "0.4.17"
@ -21,6 +19,11 @@ prometheus = "0.13.1"
once_cell = "1.17.0"
rand = "0.8.5"
itertools = "0.10.5"
anyhow = "1.0.70"
[target.'cfg(not(target_os="linux"))'.dependencies]
ort = {git ="https://github.com/pykeio/ort.git", features=["profiling"], tag="v1.14.6"}
[target.'cfg(target_os="linux")'.dependencies]
ort = {git ="https://github.com/pykeio/ort.git", features=["profiling", "tensorrt", "cuda", "copy-dylibs"], tag="v1.14.6"}
[dev-dependencies]
criterion = "0.3.0"

3
navi/dr_transform/src/all_config.rs
View File

@ -44,5 +44,6 @@ pub struct RenamedFeatures {
}
pub fn parse(json_str: &str) -> Result<AllConfig, Error> {
serde_json::from_str(json_str)
let all_config: AllConfig = serde_json::from_str(json_str)?;
Ok(all_config)
}

262
navi/dr_transform/src/converter.rs
View File

@ -2,6 +2,9 @@ use std::collections::BTreeSet;
use std::fmt::{self, Debug, Display};
use std::fs;
use crate::all_config;
use crate::all_config::AllConfig;
use anyhow::{bail, Context};
use bpr_thrift::data::DataRecord;
use bpr_thrift::prediction_service::BatchPredictionRequest;
use bpr_thrift::tensor::GeneralTensor;
@ -16,8 +19,6 @@ use segdense::util;
use thrift::protocol::{TBinaryInputProtocol, TSerializable};
use thrift::transport::TBufferChannel;
use crate::{all_config, all_config::AllConfig};
pub fn log_feature_match(
dr: &DataRecord,
seg_dense_config: &DensificationTransformSpec,
@ -28,20 +29,24 @@ pub fn log_feature_match(
for (feature_id, feature_value) in dr.continuous_features.as_ref().unwrap() {
debug!(
"{dr_type} - Continuous Datarecord => Feature ID: {feature_id}, Feature value: {feature_value}"
"{} - Continous Datarecord => Feature ID: {}, Feature value: {}",
dr_type, feature_id, feature_value
);
for input_feature in &seg_dense_config.cont.input_features {
if input_feature.feature_id == *feature_id {
debug!("Matching input feature: {input_feature:?}")
debug!("Matching input feature: {:?}", input_feature)
}
}
}
for feature_id in dr.binary_features.as_ref().unwrap() {
debug!("{dr_type} - Binary Datarecord => Feature ID: {feature_id}");
debug!(
"{} - Binary Datarecord => Feature ID: {}",
dr_type, feature_id
);
for input_feature in &seg_dense_config.binary.input_features {
if input_feature.feature_id == *feature_id {
debug!("Found input feature: {input_feature:?}")
debug!("Found input feature: {:?}", input_feature)
}
}
}
@ -90,18 +95,19 @@ impl BatchPredictionRequestToTorchTensorConverter {
model_version: &str,
reporting_feature_ids: Vec<(i64, &str)>,
register_metric_fn: Option<impl Fn(&HistogramVec)>,
) -> BatchPredictionRequestToTorchTensorConverter {
let all_config_path = format!("{model_dir}/{model_version}/all_config.json");
let seg_dense_config_path =
format!("{model_dir}/{model_version}/segdense_transform_spec_home_recap_2022.json");
let seg_dense_config = util::load_config(&seg_dense_config_path);
) -> anyhow::Result<BatchPredictionRequestToTorchTensorConverter> {
let all_config_path = format!("{}/{}/all_config.json", model_dir, model_version);
let seg_dense_config_path = format!(
"{}/{}/segdense_transform_spec_home_recap_2022.json",
model_dir, model_version
);
let seg_dense_config = util::load_config(&seg_dense_config_path)?;
let all_config = all_config::parse(
&fs::read_to_string(&all_config_path)
.unwrap_or_else(|error| panic!("error loading all_config.json - {error}")),
)
.unwrap();
.with_context(|| "error loading all_config.json - ")?,
)?;
let feature_mapper = util::load_from_parsed_config_ref(&seg_dense_config);
let feature_mapper = util::load_from_parsed_config(seg_dense_config.clone())?;
let user_embedding_feature_id = Self::get_feature_id(
&all_config
@ -131,11 +137,11 @@ impl BatchPredictionRequestToTorchTensorConverter {
let (discrete_feature_metrics, continuous_feature_metrics) = METRICS.get_or_init(|| {
let discrete = HistogramVec::new(
HistogramOpts::new(":navi:feature_id:discrete", "Discrete Feature ID values")
.buckets(Vec::from([
0.0f64, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0, 80.0, 90.0, 100.0, 110.0,
.buckets(Vec::from(&[
0.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0, 80.0, 90.0, 100.0, 110.0,
120.0, 130.0, 140.0, 150.0, 160.0, 170.0, 180.0, 190.0, 200.0, 250.0,
300.0, 500.0, 1000.0, 10000.0, 100000.0,
])),
] as &'static [f64])),
&["feature_id"],
)
.expect("metric cannot be created");
@ -144,18 +150,18 @@ impl BatchPredictionRequestToTorchTensorConverter {
":navi:feature_id:continuous",
"continuous Feature ID values",
)
.buckets(Vec::from([
0.0f64, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0, 80.0, 90.0, 100.0, 110.0,
120.0, 130.0, 140.0, 150.0, 160.0, 170.0, 180.0, 190.0, 200.0, 250.0, 300.0,
500.0, 1000.0, 10000.0, 100000.0,
])),
.buckets(Vec::from(&[
0.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0, 80.0, 90.0, 100.0, 110.0, 120.0,
130.0, 140.0, 150.0, 160.0, 170.0, 180.0, 190.0, 200.0, 250.0, 300.0, 500.0,
1000.0, 10000.0, 100000.0,
] as &'static [f64])),
&["feature_id"],
)
.expect("metric cannot be created");
if let Some(r) = register_metric_fn {
register_metric_fn.map(|r| {
r(&discrete);
r(&continuous);
}
});
(discrete, continuous)
});
@ -164,13 +170,16 @@ impl BatchPredictionRequestToTorchTensorConverter {
for (feature_id, feature_type) in reporting_feature_ids.iter() {
match *feature_type {
"discrete" => discrete_features_to_report.insert(*feature_id),
"continuous" => continuous_features_to_report.insert(*feature_id),
_ => panic!("Invalid feature type {feature_type} for reporting metrics!"),
"discrete" => discrete_features_to_report.insert(feature_id.clone()),
"continuous" => continuous_features_to_report.insert(feature_id.clone()),
_ => bail!(
"Invalid feature type {} for reporting metrics!",
feature_type
),
};
}
BatchPredictionRequestToTorchTensorConverter {
Ok(BatchPredictionRequestToTorchTensorConverter {
all_config,
seg_dense_config,
all_config_path,
@ -183,7 +192,7 @@ impl BatchPredictionRequestToTorchTensorConverter {
continuous_features_to_report,
discrete_feature_metrics,
continuous_feature_metrics,
}
})
}
fn get_feature_id(feature_name: &str, seg_dense_config: &Root) -> i64 {
@ -218,43 +227,45 @@ impl BatchPredictionRequestToTorchTensorConverter {
let mut working_set = vec![0 as f32; total_size];
let mut bpr_start = 0;
for (bpr, &bpr_end) in bprs.iter().zip(batch_size) {
if bpr.common_features.is_some()
&& bpr.common_features.as_ref().unwrap().tensors.is_some()
&& bpr
.common_features
.as_ref()
.unwrap()
.tensors
.as_ref()
.unwrap()
.contains_key(&feature_id)
{
let source_tensor = bpr
.common_features
.as_ref()
.unwrap()
.tensors
.as_ref()
.unwrap()
.get(&feature_id)
.unwrap();
let tensor = match source_tensor {
GeneralTensor::FloatTensor(float_tensor) =>
//Tensor::of_slice(
if bpr.common_features.is_some() {
if bpr.common_features.as_ref().unwrap().tensors.is_some() {
if bpr
.common_features
.as_ref()
.unwrap()
.tensors
.as_ref()
.unwrap()
.contains_key(&feature_id)
{
float_tensor
.floats
.iter()
.map(|x| x.into_inner() as f32)
.collect::<Vec<_>>()
}
_ => vec![0 as f32; cols],
};
let source_tensor = bpr
.common_features
.as_ref()
.unwrap()
.tensors
.as_ref()
.unwrap()
.get(&feature_id)
.unwrap();
let tensor = match source_tensor {
GeneralTensor::FloatTensor(float_tensor) =>
//Tensor::of_slice(
{
float_tensor
.floats
.iter()
.map(|x| x.into_inner() as f32)
.collect::<Vec<_>>()
}
_ => vec![0 as f32; cols],
};
// since the tensor is found in common feature, add it in all batches
for row in bpr_start..bpr_end {
for col in 0..cols {
working_set[row * cols + col] = tensor[col];
// since the tensor is found in common feature, add it in all batches
for row in bpr_start..bpr_end {
for col in 0..cols {
working_set[row * cols + col] = tensor[col];
}
}
}
}
}
@ -298,9 +309,9 @@ impl BatchPredictionRequestToTorchTensorConverter {
// (INT64 --> INT64, DataRecord.discrete_feature)
fn get_continuous(&self, bprs: &[BatchPredictionRequest], batch_ends: &[usize]) -> InputTensor {
// These need to be part of model schema
let rows = batch_ends[batch_ends.len() - 1];
let cols = 5293;
let full_size = rows * cols;
let rows: usize = batch_ends[batch_ends.len() - 1];
let cols: usize = 5293;
let full_size: usize = rows * cols;
let default_val = f32::NAN;
let mut tensor = vec![default_val; full_size];
@ -325,15 +336,18 @@ impl BatchPredictionRequestToTorchTensorConverter {
.unwrap();
for feature in common_features {
if let Some(f_info) = self.feature_mapper.get(feature.0) {
let idx = f_info.index_within_tensor as usize;
if idx < cols {
// Set value in each row
for r in bpr_start..bpr_end {
let flat_index = r * cols + idx;
tensor[flat_index] = feature.1.into_inner() as f32;
match self.feature_mapper.get(feature.0) {
Some(f_info) => {
let idx = f_info.index_within_tensor as usize;
if idx < cols {
// Set value in each row
for r in bpr_start..bpr_end {
let flat_index: usize = r * cols + idx;
tensor[flat_index] = feature.1.into_inner() as f32;
}
}
}
None => (),
}
if self.continuous_features_to_report.contains(feature.0) {
self.continuous_feature_metrics
@ -349,24 +363,28 @@ impl BatchPredictionRequestToTorchTensorConverter {
// Process the batch of datarecords
for r in bpr_start..bpr_end {
let dr: &DataRecord = &bpr.individual_features_list[r - bpr_start];
let dr: &DataRecord =
&bpr.individual_features_list[usize::try_from(r - bpr_start).unwrap()];
if dr.continuous_features.is_some() {
for feature in dr.continuous_features.as_ref().unwrap() {
if let Some(f_info) = self.feature_mapper.get(feature.0) {
let idx = f_info.index_within_tensor as usize;
let flat_index = r * cols + idx;
if flat_index < tensor.len() && idx < cols {
tensor[flat_index] = feature.1.into_inner() as f32;
match self.feature_mapper.get(&feature.0) {
Some(f_info) => {
let idx = f_info.index_within_tensor as usize;
let flat_index: usize = r * cols + idx;
if flat_index < tensor.len() && idx < cols {
tensor[flat_index] = feature.1.into_inner() as f32;
}
}
None => (),
}
if self.continuous_features_to_report.contains(feature.0) {
self.continuous_feature_metrics
.with_label_values(&[feature.0.to_string().as_str()])
.observe(feature.1.into_inner())
.observe(feature.1.into_inner() as f64)
} else if self.discrete_features_to_report.contains(feature.0) {
self.discrete_feature_metrics
.with_label_values(&[feature.0.to_string().as_str()])
.observe(feature.1.into_inner())
.observe(feature.1.into_inner() as f64)
}
}
}
@ -383,10 +401,10 @@ impl BatchPredictionRequestToTorchTensorConverter {
fn get_binary(&self, bprs: &[BatchPredictionRequest], batch_ends: &[usize]) -> InputTensor {
// These need to be part of model schema
let rows = batch_ends[batch_ends.len() - 1];
let cols = 149;
let full_size = rows * cols;
let default_val = 0;
let rows: usize = batch_ends[batch_ends.len() - 1];
let cols: usize = 149;
let full_size: usize = rows * cols;
let default_val: i64 = 0;
let mut v = vec![default_val; full_size];
@ -410,15 +428,18 @@ impl BatchPredictionRequestToTorchTensorConverter {
.unwrap();
for feature in common_features {
if let Some(f_info) = self.feature_mapper.get(feature) {
let idx = f_info.index_within_tensor as usize;
if idx < cols {
// Set value in each row
for r in bpr_start..bpr_end {
let flat_index = r * cols + idx;
v[flat_index] = 1;
match self.feature_mapper.get(feature) {
Some(f_info) => {
let idx = f_info.index_within_tensor as usize;
if idx < cols {
// Set value in each row
for r in bpr_start..bpr_end {
let flat_index: usize = r * cols + idx;
v[flat_index] = 1;
}
}
}
None => (),
}
}
}
@ -428,10 +449,13 @@ impl BatchPredictionRequestToTorchTensorConverter {
let dr: &DataRecord = &bpr.individual_features_list[r - bpr_start];
if dr.binary_features.is_some() {
for feature in dr.binary_features.as_ref().unwrap() {
if let Some(f_info) = self.feature_mapper.get(feature) {
let idx = f_info.index_within_tensor as usize;
let flat_index = r * cols + idx;
v[flat_index] = 1;
match self.feature_mapper.get(&feature) {
Some(f_info) => {
let idx = f_info.index_within_tensor as usize;
let flat_index: usize = r * cols + idx;
v[flat_index] = 1;
}
None => (),
}
}
}
@ -448,10 +472,10 @@ impl BatchPredictionRequestToTorchTensorConverter {
#[allow(dead_code)]
fn get_discrete(&self, bprs: &[BatchPredictionRequest], batch_ends: &[usize]) -> InputTensor {
// These need to be part of model schema
let rows = batch_ends[batch_ends.len() - 1];
let cols = 320;
let full_size = rows * cols;
let default_val = 0;
let rows: usize = batch_ends[batch_ends.len() - 1];
let cols: usize = 320;
let full_size: usize = rows * cols;
let default_val: i64 = 0;
let mut v = vec![default_val; full_size];
@ -475,15 +499,18 @@ impl BatchPredictionRequestToTorchTensorConverter {
.unwrap();
for feature in common_features {
if let Some(f_info) = self.feature_mapper.get(feature.0) {
let idx = f_info.index_within_tensor as usize;
if idx < cols {
// Set value in each row
for r in bpr_start..bpr_end {
let flat_index = r * cols + idx;
v[flat_index] = *feature.1;
match self.feature_mapper.get(feature.0) {
Some(f_info) => {
let idx = f_info.index_within_tensor as usize;
if idx < cols {
// Set value in each row
for r in bpr_start..bpr_end {
let flat_index: usize = r * cols + idx;
v[flat_index] = *feature.1;
}
}
}
None => (),
}
if self.discrete_features_to_report.contains(feature.0) {
self.discrete_feature_metrics
@ -495,15 +522,18 @@ impl BatchPredictionRequestToTorchTensorConverter {
// Process the batch of datarecords
for r in bpr_start..bpr_end {
let dr: &DataRecord = &bpr.individual_features_list[r];
let dr: &DataRecord = &bpr.individual_features_list[usize::try_from(r).unwrap()];
if dr.discrete_features.is_some() {
for feature in dr.discrete_features.as_ref().unwrap() {
if let Some(f_info) = self.feature_mapper.get(feature.0) {
let idx = f_info.index_within_tensor as usize;
let flat_index = r * cols + idx;
if flat_index < v.len() && idx < cols {
v[flat_index] = *feature.1;
match self.feature_mapper.get(&feature.0) {
Some(f_info) => {
let idx = f_info.index_within_tensor as usize;
let flat_index: usize = r * cols + idx;
if flat_index < v.len() && idx < cols {
v[flat_index] = *feature.1;
}
}
None => (),
}
if self.discrete_features_to_report.contains(feature.0) {
self.discrete_feature_metrics
@ -569,7 +599,7 @@ impl Converter for BatchPredictionRequestToTorchTensorConverter {
.map(|bpr| bpr.individual_features_list.len())
.scan(0usize, |acc, e| {
//running total
*acc += e;
*acc = *acc + e;
Some(*acc)
})
.collect::<Vec<_>>();

1
navi/dr_transform/src/lib.rs
View File

@ -3,3 +3,4 @@ pub mod converter;
#[cfg(test)]
mod test;
pub mod util;
pub extern crate ort;

21
navi/navi/Cargo.toml
View File

@ -1,8 +1,7 @@
[package]
name = "navi"
version = "2.0.42"
version = "2.0.45"
edition = "2021"
# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
[[bin]]
name = "navi"
@ -16,12 +15,19 @@ required-features=["torch"]
name = "navi_onnx"
path = "src/bin/navi_onnx.rs"
required-features=["onnx"]
[[bin]]
name = "navi_onnx_test"
path = "src/bin/bin_tests/navi_onnx_test.rs"
[[bin]]
name = "navi_torch_test"
path = "src/bin/bin_tests/navi_torch_test.rs"
required-features=["torch"]
[features]
default=[]
navi_console=[]
torch=["tch"]
onnx=["ort"]
onnx=[]
tf=["tensorflow"]
[dependencies]
itertools = "0.10.5"
@ -47,6 +53,7 @@ parking_lot = "0.12.1"
rand = "0.8.5"
rand_pcg = "0.3.1"
random = "0.12.2"
x509-parser = "0.15.0"
sha256 = "1.0.3"
tonic = { version = "0.6.2", features=['compression', 'tls'] }
tokio = { version = "1.17.0", features = ["macros", "rt-multi-thread", "fs", "process"] }
@ -55,16 +62,12 @@ npyz = "0.7.3"
base64 = "0.21.0"
histogram = "0.6.9"
tch = {version = "0.10.3", optional = true}
tensorflow = { version = "0.20.0", optional = true }
tensorflow = { version = "0.18.0", optional = true }
once_cell = {version = "1.17.1"}
ndarray = "0.15"
serde = "1.0.154"
serde_json = "1.0.94"
dr_transform = { path = "../dr_transform"}
[target.'cfg(not(target_os="linux"))'.dependencies]
ort = {git ="https://github.com/pykeio/ort.git", features=["profiling"], optional = true, tag="v1.14.2"}
[target.'cfg(target_os="linux")'.dependencies]
ort = {git ="https://github.com/pykeio/ort.git", features=["profiling", "tensorrt", "cuda", "copy-dylibs"], optional = true, tag="v1.14.2"}
[build-dependencies]
tonic-build = {version = "0.6.2", features=['prost', "compression"] }
[profile.release]
@ -74,3 +77,5 @@ ndarray-rand = "0.14.0"
tokio-test = "*"
assert_cmd = "2.0"
criterion = "0.4.0"

10
navi/navi/proto/tensorflow/core/framework/full_type.proto
View File

@ -122,7 +122,7 @@ enum FullTypeId {
// TFT_TENSOR[TFT_INT32, TFT_UNKNOWN]
// is a Tensor of int32 element type and unknown shape.
//
// TODO: Define TFT_SHAPE and add more examples.
// TODO(mdan): Define TFT_SHAPE and add more examples.
TFT_TENSOR = 1000;
// Array (or tensorflow::TensorList in the variant type registry).
@ -178,7 +178,7 @@ enum FullTypeId {
// object (for now).
// The bool element type.
// TODO
// TODO(mdan): Quantized types, legacy representations (e.g. ref)
TFT_BOOL = 200;
// Integer element types.
TFT_UINT8 = 201;
@ -195,7 +195,7 @@ enum FullTypeId {
TFT_DOUBLE = 211;
TFT_BFLOAT16 = 215;
// Complex element types.
// TODO: Represent as TFT_COMPLEX[TFT_DOUBLE] instead?
// TODO(mdan): Represent as TFT_COMPLEX[TFT_DOUBLE] instead?
TFT_COMPLEX64 = 212;
TFT_COMPLEX128 = 213;
// The string element type.
@ -240,7 +240,7 @@ enum FullTypeId {
// ownership is in the true sense: "the op argument representing the lock is
// available".
// Mutex locks are the dynamic counterpart of control dependencies.
// TODO: Properly document this thing.
// TODO(mdan): Properly document this thing.
//
// Parametrization: TFT_MUTEX_LOCK[].
TFT_MUTEX_LOCK = 10202;
@ -271,6 +271,6 @@ message FullTypeDef {
oneof attr {
string s = 3;
int64 i = 4;
// TODO: list/tensor, map? Need to reconcile with TFT_RECORD, etc.
// TODO(mdan): list/tensor, map? Need to reconcile with TFT_RECORD, etc.
}
}

2
navi/navi/proto/tensorflow/core/framework/function.proto
View File

@ -23,7 +23,7 @@ message FunctionDefLibrary {
// with a value. When a GraphDef has a call to a function, it must
// have binding for every attr defined in the signature.
//
// TODO:
// TODO(zhifengc):
// * device spec, etc.
message FunctionDef {
// The definition of the function's name, arguments, return values,

2
navi/navi/proto/tensorflow/core/framework/node_def.proto
View File

@ -61,7 +61,7 @@ message NodeDef {
// one of the names from the corresponding OpDef's attr field).
// The values must have a type matching the corresponding OpDef
// attr's type field.
// TODO: Add some examples here showing best practices.
// TODO(josh11b): Add some examples here showing best practices.
map<string, AttrValue> attr = 5;
message ExperimentalDebugInfo {

4
navi/navi/proto/tensorflow/core/framework/op_def.proto
View File

@ -96,7 +96,7 @@ message OpDef {
// Human-readable description.
string description = 4;
// TODO: bool is_optional?
// TODO(josh11b): bool is_optional?
// --- Constraints ---
// These constraints are only in effect if specified. Default is no
@ -139,7 +139,7 @@ message OpDef {
// taking input from multiple devices with a tree of aggregate ops
// that aggregate locally within each device (and possibly within
// groups of nearby devices) before communicating.
// TODO: Implement that optimization.
// TODO(josh11b): Implement that optimization.
bool is_aggregate = 16; // for things like add
// Other optimizations go here, like

2
navi/navi/proto/tensorflow/core/framework/step_stats.proto
View File

@ -53,7 +53,7 @@ message MemoryStats {
// Time/size stats recorded for a single execution of a graph node.
message NodeExecStats {
// TODO: Use some more compact form of node identity than
// TODO(tucker): Use some more compact form of node identity than
// the full string name. Either all processes should agree on a
// global id (cost_id?) for each node, or we should use a hash of
// the name.

2
navi/navi/proto/tensorflow/core/framework/tensor.proto
View File

@ -16,7 +16,7 @@ option go_package = "github.com/tensorflow/tensorflow/tensorflow/go/core/framewo
message TensorProto {
DataType dtype = 1;
// Shape of the tensor. TODO: sort out the 0-rank issues.
// Shape of the tensor. TODO(touts): sort out the 0-rank issues.
TensorShapeProto tensor_shape = 2;
// Only one of the representations below is set, one of "tensor_contents" and

8
navi/navi/proto/tensorflow/core/protobuf/config.proto
View File

@ -532,7 +532,7 @@ message ConfigProto {
// We removed the flag client_handles_error_formatting. Marking the tag
// number as reserved.
// TODO: Should we just remove this tag so that it can be
// TODO(shikharagarwal): Should we just remove this tag so that it can be
// used in future for other purpose?
reserved 2;
@ -576,7 +576,7 @@ message ConfigProto {
// - If isolate_session_state is true, session states are isolated.
// - If isolate_session_state is false, session states are shared.
//
// TODO: Add a single API that consistently treats
// TODO(b/129330037): Add a single API that consistently treats
// isolate_session_state and ClusterSpec propagation.
bool share_session_state_in_clusterspec_propagation = 8;
@ -704,7 +704,7 @@ message ConfigProto {
// Options for a single Run() call.
message RunOptions {
// TODO Turn this into a TraceOptions proto which allows
// TODO(pbar) Turn this into a TraceOptions proto which allows
// tracing to be controlled in a more orthogonal manner?
enum TraceLevel {
NO_TRACE = 0;
@ -781,7 +781,7 @@ message RunMetadata {
repeated GraphDef partition_graphs = 3;
message FunctionGraphs {
// TODO: Include some sort of function/cache-key identifier?
// TODO(nareshmodi): Include some sort of function/cache-key identifier?
repeated GraphDef partition_graphs = 1;
GraphDef pre_optimization_graph = 2;

2
navi/navi/proto/tensorflow/core/protobuf/coordination_service.proto
View File

@ -194,7 +194,7 @@ service CoordinationService {
// Report error to the task. RPC sets the receiving instance of coordination
// service agent to error state permanently.
// TODO: Consider splitting this into a different RPC service.
// TODO(b/195990880): Consider splitting this into a different RPC service.
rpc ReportErrorToAgent(ReportErrorToAgentRequest)
returns (ReportErrorToAgentResponse);

2
navi/navi/proto/tensorflow/core/protobuf/debug.proto
View File

@ -46,7 +46,7 @@ message DebugTensorWatch {
// are to be debugged, the callers of Session::Run() must use distinct
// debug_urls to make sure that the streamed or dumped events do not overlap
// among the invocations.
// TODO: More visible documentation of this in g3docs.
// TODO(cais): More visible documentation of this in g3docs.
repeated string debug_urls = 4;
// Do not error out if debug op creation fails (e.g., due to dtype

6
navi/navi/proto/tensorflow/core/protobuf/debug_event.proto
View File

@ -12,7 +12,7 @@ option java_package = "org.tensorflow.util";
option go_package = "github.com/tensorflow/tensorflow/tensorflow/go/core/protobuf/for_core_protos_go_proto";
// Available modes for extracting debugging information from a Tensor.
// TODO: Document the detailed column names and semantics in a separate
// TODO(cais): Document the detailed column names and semantics in a separate
// markdown file once the implementation settles.
enum TensorDebugMode {
UNSPECIFIED = 0;
@ -223,7 +223,7 @@ message DebuggedDevice {
// A debugger-generated ID for the device. Guaranteed to be unique within
// the scope of the debugged TensorFlow program, including single-host and
// multi-host settings.
// TODO: Test the uniqueness guarantee in multi-host settings.
// TODO(cais): Test the uniqueness guarantee in multi-host settings.
int32 device_id = 2;
}
@ -264,7 +264,7 @@ message Execution {
// field with the DebuggedDevice messages.
repeated int32 output_tensor_device_ids = 9;
// TODO support, add more fields
// TODO(cais): When backporting to V1 Session.run() support, add more fields
// such as fetches and feeds.
}

2
navi/navi/proto/tensorflow/core/protobuf/distributed_runtime_payloads.proto
View File

@ -7,7 +7,7 @@ option go_package = "github.com/tensorflow/tensorflow/tensorflow/go/core/protobu
// Used to serialize and transmit tensorflow::Status payloads through
// grpc::Status `error_details` since grpc::Status lacks payload API.
// TODO: Use GRPC API once supported.
// TODO(b/204231601): Use GRPC API once supported.
message GrpcPayloadContainer {
map<string, bytes> payloads = 1;
}

2
navi/navi/proto/tensorflow/core/protobuf/eager_service.proto
View File

@ -172,7 +172,7 @@ message WaitQueueDoneRequest {
}
message WaitQueueDoneResponse {
// TODO: Consider adding NodeExecStats here to be able to
// TODO(nareshmodi): Consider adding NodeExecStats here to be able to
// propagate some stats.
}

2
navi/navi/proto/tensorflow/core/protobuf/master.proto
View File

@ -94,7 +94,7 @@ message ExtendSessionRequest {
}
message ExtendSessionResponse {
// TODO: Return something about the operation?
// TODO(mrry): Return something about the operation?
// The new version number for the extended graph, to be used in the next call
// to ExtendSession.

2
navi/navi/proto/tensorflow/core/protobuf/saved_object_graph.proto
View File

@ -176,7 +176,7 @@ message SavedBareConcreteFunction {
// allows the ConcreteFunction to be called with nest structure inputs. This
// field may not be populated. If this field is absent, the concrete function
// can only be called with flat inputs.
// TODO: support calling saved ConcreteFunction with structured
// TODO(b/169361281): support calling saved ConcreteFunction with structured
// inputs in C++ SavedModel API.
FunctionSpec function_spec = 4;
}

2
navi/navi/proto/tensorflow/core/protobuf/tensor_bundle.proto
View File

@ -17,7 +17,7 @@ option go_package = "github.com/tensorflow/tensorflow/tensorflow/go/core/protobu
// Special header that is associated with a bundle.
//
// TODO: maybe in the future, we can add information about
// TODO(zongheng,zhifengc): maybe in the future, we can add information about
// which binary produced this checkpoint, timestamp, etc. Sometime, these can be
// valuable debugging information. And if needed, these can be used as defensive
// information ensuring reader (binary version) of the checkpoint and the writer

4
navi/navi/proto/tensorflow/core/protobuf/worker.proto
View File

@ -188,7 +188,7 @@ message DeregisterGraphRequest {
}
message DeregisterGraphResponse {
// TODO: Optionally add summary stats for the graph.
// TODO(mrry): Optionally add summary stats for the graph.
}
////////////////////////////////////////////////////////////////////////////////
@ -294,7 +294,7 @@ message RunGraphResponse {
// If the request asked for execution stats, the cost graph, or the partition
// graphs, these are returned here.
// TODO: Package these in a RunMetadata instead.
// TODO(suharshs): Package these in a RunMetadata instead.
StepStats step_stats = 2;
CostGraphDef cost_graph = 3;
repeated GraphDef partition_graph = 4;

2
navi/navi/proto/tensorflow_serving/apis/logging.proto
View File

@ -13,5 +13,5 @@ message LogMetadata {
SamplingConfig sampling_config = 2;
// List of tags used to load the relevant MetaGraphDef from SavedModel.
repeated string saved_model_tags = 3;
// TODO: Add more metadata as mentioned in the bug.
// TODO(b/33279154): Add more metadata as mentioned in the bug.
}

4
navi/navi/proto/tensorflow_serving/config/file_system_storage_path_source.proto
View File

@ -58,7 +58,7 @@ message FileSystemStoragePathSourceConfig {
// A single servable name/base_path pair to monitor.
// DEPRECATED: Use 'servables' instead.
// TODO: Stop using these fields, and ultimately remove them here.
// TODO(b/30898016): Stop using these fields, and ultimately remove them here.
string servable_name = 1 [deprecated = true];
string base_path = 2 [deprecated = true];
@ -76,7 +76,7 @@ message FileSystemStoragePathSourceConfig {
// check for a version to appear later.)
// DEPRECATED: Use 'servable_versions_always_present' instead, which includes
// this behavior.
// TODO: Remove 2019-10-31 or later.
// TODO(b/30898016): Remove 2019-10-31 or later.
bool fail_if_zero_versions_at_startup = 4 [deprecated = true];
// If true, the servable is always expected to exist on the underlying

4
navi/navi/proto/tensorflow_serving/config/model_server_config.proto
View File

@ -9,7 +9,7 @@ import "tensorflow_serving/config/logging_config.proto";
option cc_enable_arenas = true;
// The type of model.
// TODO: DEPRECATED.
// TODO(b/31336131): DEPRECATED.
enum ModelType {
MODEL_TYPE_UNSPECIFIED = 0 [deprecated = true];
TENSORFLOW = 1 [deprecated = true];
@ -31,7 +31,7 @@ message ModelConfig {
string base_path = 2;
// Type of model.
// TODO: DEPRECATED. Please use 'model_platform' instead.
// TODO(b/31336131): DEPRECATED. Please use 'model_platform' instead.
ModelType model_type = 3 [deprecated = true];
// Type of model (e.g. "tensorflow").

9
navi/navi/scripts/run_onnx.sh
View File

@ -1,10 +1,9 @@
#!/bin/sh
#RUST_LOG=debug LD_LIBRARY_PATH=so/onnx/lib target/release/navi_onnx --port 30 --num-worker-threads 8 --intra-op-parallelism 8 --inter-op-parallelism 8 \
RUST_LOG=info LD_LIBRARY_PATH=so/onnx/lib cargo run --bin navi_onnx --features onnx -- \
--port 30 --num-worker-threads 8 --intra-op-parallelism 8 --inter-op-parallelism 8 \
--port 8030 --num-worker-threads 8 \
--model-check-interval-secs 30 \
--model-dir models/int8 \
--output caligrated_probabilities \
--input "" \
--modelsync-cli "echo" \
--onnx-ep-options use_arena=true
--onnx-ep-options use_arena=true \
--model-dir models/prod_home --output caligrated_probabilities --input "" --intra-op-parallelism 8 --inter-op-parallelism 8 --max-batch-size 1 --batch-time-out-millis 1 \
--model-dir models/prod_home1 --output caligrated_probabilities --input "" --intra-op-parallelism 8 --inter-op-parallelism 8 --max-batch-size 1 --batch-time-out-millis 1 \

15
navi/navi/src/bin/navi_onnx.rs
View File

@ -1,11 +1,24 @@
use anyhow::Result;
use log::info;
use navi::cli_args::{ARGS, MODEL_SPECS};
use navi::onnx_model::onnx::OnnxModel;
use navi::{bootstrap, metrics};
fn main() -> Result<()> {
env_logger::init();
assert_eq!(MODEL_SPECS.len(), ARGS.inter_op_parallelism.len());
info!("global: {:?}", ARGS.onnx_global_thread_pool_options);
let assert_session_params = if ARGS.onnx_global_thread_pool_options.is_empty() {
// std::env::set_var("OMP_NUM_THREADS", "1");
info!("now we use per session thread pool");
MODEL_SPECS.len()
}
else {
info!("now we use global thread pool");
0
};
assert_eq!(assert_session_params, ARGS.inter_op_parallelism.len());
assert_eq!(assert_session_params, ARGS.inter_op_parallelism.len());
metrics::register_custom_metrics();
bootstrap::bootstrap(OnnxModel::new)
}

29
navi/navi/src/bootstrap.rs
View File

@ -1,5 +1,6 @@
use anyhow::Result;
use log::{info, warn};
use x509_parser::{prelude::{parse_x509_pem}, parse_x509_certificate};
use std::collections::HashMap;
use tokio::time::Instant;
use tonic::{
@ -27,6 +28,7 @@ use crate::cli_args::{ARGS, INPUTS, OUTPUTS};
use crate::metrics::{
NAVI_VERSION, NUM_PREDICTIONS, NUM_REQUESTS_FAILED, NUM_REQUESTS_FAILED_BY_MODEL,
NUM_REQUESTS_RECEIVED, NUM_REQUESTS_RECEIVED_BY_MODEL, RESPONSE_TIME_COLLECTOR,
CERT_EXPIRY_EPOCH
};
use crate::predict_service::{Model, PredictService};
use crate::tf_proto::tensorflow_serving::model_spec::VersionChoice::Version;
@ -207,6 +209,9 @@ impl<T: Model> PredictionService for PredictService<T> {
PredictResult::DropDueToOverload => Err(Status::resource_exhausted("")),
PredictResult::ModelNotFound(idx) => {
Err(Status::not_found(format!("model index {}", idx)))
},
PredictResult::ModelNotReady(idx) => {
Err(Status::unavailable(format!("model index {}", idx)))
}
PredictResult::ModelVersionNotFound(idx, version) => Err(
Status::not_found(format!("model index:{}, version {}", idx, version)),
@ -230,6 +235,12 @@ impl<T: Model> PredictionService for PredictService<T> {
}
}
// A function that takes a timestamp as input and returns a ticker stream
fn report_expiry(expiry_time: i64) {
info!("Certificate expires at epoch: {:?}", expiry_time);
CERT_EXPIRY_EPOCH.set(expiry_time as i64);
}
pub fn bootstrap<T: Model>(model_factory: ModelFactory<T>) -> Result<()> {
info!("package: {}, version: {}, args: {:?}", NAME, VERSION, *ARGS);
//we follow SemVer. So here we assume MAJOR.MINOR.PATCH
@ -246,6 +257,7 @@ pub fn bootstrap<T: Model>(model_factory: ModelFactory<T>) -> Result<()> {
);
}
tokio::runtime::Builder::new_multi_thread()
.thread_name("async worker")
.worker_threads(ARGS.num_worker_threads)
@ -263,6 +275,21 @@ pub fn bootstrap<T: Model>(model_factory: ModelFactory<T>) -> Result<()> {
let mut builder = if ARGS.ssl_dir.is_empty() {
Server::builder()
} else {
// Read the pem file as a string
let pem_str = std::fs::read_to_string(format!("{}/server.crt", ARGS.ssl_dir)).unwrap();
let res = parse_x509_pem(&pem_str.as_bytes());
match res {
Ok((rem, pem_2)) => {
assert!(rem.is_empty());
assert_eq!(pem_2.label, String::from("CERTIFICATE"));
let res_x509 = parse_x509_certificate(&pem_2.contents);
info!("Certificate label: {}", pem_2.label);
assert!(res_x509.is_ok());
report_expiry(res_x509.unwrap().1.validity().not_after.timestamp());
},
_ => panic!("PEM parsing failed: {:?}", res),
}
let key = tokio::fs::read(format!("{}/server.key", ARGS.ssl_dir))
.await
.expect("can't find key file");
@ -278,7 +305,7 @@ pub fn bootstrap<T: Model>(model_factory: ModelFactory<T>) -> Result<()> {
let identity = Identity::from_pem(pem.clone(), key);
let client_ca_cert = Certificate::from_pem(pem.clone());
let tls = ServerTlsConfig::new()
.identity(identity)
.identity(identity)
.client_ca_root(client_ca_cert);
Server::builder()
.tls_config(tls)

10
navi/navi/src/cli_args.rs
View File

@ -87,13 +87,11 @@ pub struct Args {
pub intra_op_parallelism: Vec<String>,
#[clap(
long,
default_value = "14",
help = "number of threads to parallelize computations of the graph"
)]
pub inter_op_parallelism: Vec<String>,
#[clap(
long,
default_value = "serving_default",
help = "signature of a serving. only TF"
)]
pub serving_sig: Vec<String>,
@ -107,10 +105,12 @@ pub struct Args {
help = "max warmup records to use. warmup only implemented for TF"
)]
pub max_warmup_records: usize,
#[clap(long, value_parser = Args::parse_key_val::<String, String>, value_delimiter=',')]
pub onnx_global_thread_pool_options: Vec<(String, String)>,
#[clap(
long,
default_value = "true",
help = "when to use graph parallelization. only for ONNX"
long,
default_value = "true",
help = "when to use graph parallelization. only for ONNX"
)]
pub onnx_use_parallel_mode: String,
// #[clap(long, default_value = "false")]

1
navi/navi/src/lib.rs
View File

@ -144,6 +144,7 @@ pub enum PredictResult {
Ok(Vec<TensorScores>, i64),
DropDueToOverload,
ModelNotFound(usize),
ModelNotReady(usize),
ModelVersionNotFound(usize, i64),
}

7
navi/navi/src/metrics.rs
View File

@ -171,6 +171,9 @@ lazy_static! {
&["model_name"]
)
.expect("metric can be created");
pub static ref CERT_EXPIRY_EPOCH: IntGauge =
IntGauge::new(":navi:cert_expiry_epoch", "Timestamp when the current cert expires")
.expect("metric can be created");
}
pub fn register_custom_metrics() {
@ -249,6 +252,10 @@ pub fn register_custom_metrics() {
REGISTRY
.register(Box::new(CONVERTER_TIME_COLLECTOR.clone()))
.expect("collector can be registered");
REGISTRY
.register(Box::new(CERT_EXPIRY_EPOCH.clone()))
.expect("collector can be registered");
}
pub fn register_dynamic_metrics(c: &HistogramVec) {

56
navi/navi/src/onnx_model.rs
View File

@ -13,21 +13,22 @@ pub mod onnx {
use dr_transform::converter::{BatchPredictionRequestToTorchTensorConverter, Converter};
use itertools::Itertools;
use log::{debug, info};
use ort::environment::Environment;
use ort::session::Session;
use ort::tensor::InputTensor;
use ort::{ExecutionProvider, GraphOptimizationLevel, SessionBuilder};
use dr_transform::ort::environment::Environment;
use dr_transform::ort::session::Session;
use dr_transform::ort::tensor::InputTensor;
use dr_transform::ort::{ExecutionProvider, GraphOptimizationLevel, SessionBuilder};
use dr_transform::ort::LoggingLevel;
use serde_json::Value;
use std::fmt::{Debug, Display};
use std::sync::Arc;
use std::{fmt, fs};
use tokio::time::Instant;
lazy_static! {
pub static ref ENVIRONMENT: Arc<Environment> = Arc::new(
Environment::builder()
.with_name("onnx home")
.with_log_level(ort::LoggingLevel::Error)
.with_log_level(LoggingLevel::Error)
.with_global_thread_pool(ARGS.onnx_global_thread_pool_options.clone())
.build()
.unwrap()
);
@ -101,23 +102,30 @@ pub mod onnx {
let meta_info = format!("{}/{}/{}", ARGS.model_dir[idx], version, META_INFO);
let mut builder = SessionBuilder::new(&ENVIRONMENT)?
.with_optimization_level(GraphOptimizationLevel::Level3)?
.with_parallel_execution(ARGS.onnx_use_parallel_mode == "true")?
.with_inter_threads(
utils::get_config_or(
model_config,
"inter_op_parallelism",
&ARGS.inter_op_parallelism[idx],
)
.parse()?,
)?
.with_intra_threads(
utils::get_config_or(
model_config,
"intra_op_parallelism",
&ARGS.intra_op_parallelism[idx],
)
.parse()?,
)?
.with_parallel_execution(ARGS.onnx_use_parallel_mode == "true")?;
if ARGS.onnx_global_thread_pool_options.is_empty() {
builder = builder
.with_inter_threads(
utils::get_config_or(
model_config,
"inter_op_parallelism",
&ARGS.inter_op_parallelism[idx],
)
.parse()?,
)?
.with_intra_threads(
utils::get_config_or(
model_config,
"intra_op_parallelism",
&ARGS.intra_op_parallelism[idx],
)
.parse()?,
)?;
}
else {
builder = builder.with_disable_per_session_threads()?;
}
builder = builder
.with_memory_pattern(ARGS.onnx_use_memory_pattern == "true")?
.with_execution_providers(&OnnxModel::ep_choices())?;
match &ARGS.profiling {
@ -181,7 +189,7 @@ pub mod onnx {
&version,
reporting_feature_ids,
Some(metrics::register_dynamic_metrics),
)),
)?),
};
onnx_model.warmup()?;
Ok(onnx_model)

92
navi/navi/src/predict_service.rs
View File

@ -1,7 +1,7 @@
use anyhow::{anyhow, Result};
use arrayvec::ArrayVec;
use itertools::Itertools;
use log::{error, info, warn};
use log::{error, info};
use std::fmt::{Debug, Display};
use std::string::String;
use std::sync::Arc;
@ -24,7 +24,7 @@ use serde_json::{self, Value};
pub trait Model: Send + Sync + Display + Debug + 'static {
fn warmup(&self) -> Result<()>;
//TODO: refactor this to return Vec<Vec<TensorScores>>, i.e.
//TODO: refactor this to return vec<vec<TensorScores>>, i.e.
//we have the underlying runtime impl to split the response to each client.
//It will eliminate some inefficient memory copy in onnx_model.rs as well as simplify code
fn do_predict(
@ -179,17 +179,17 @@ impl<T: Model> PredictService<T> {
//initialize the latest version array
let mut cur_versions = vec!["".to_owned(); MODEL_SPECS.len()];
loop {
let config = utils::read_config(&meta_file).unwrap_or_else(|e| {
warn!("config file {} not found due to: {}", meta_file, e);
Value::Null
});
info!("***polling for models***"); //nice deliminter
info!("config:{}", config);
if let Some(ref cli) = ARGS.modelsync_cli {
if let Err(e) = call_external_modelsync(cli, &cur_versions).await {
error!("model sync cli running error:{}", e)
}
}
let config = utils::read_config(&meta_file).unwrap_or_else(|e| {
info!("config file {} not found due to: {}", meta_file, e);
Value::Null
});
info!("config:{}", config);
for (idx, cur_version) in cur_versions.iter_mut().enumerate() {
let model_dir = &ARGS.model_dir[idx];
PredictService::scan_load_latest_model_from_model_dir(
@ -222,33 +222,39 @@ impl<T: Model> PredictService<T> {
.map(|b| b.parse().unwrap())
.collect::<Vec<u64>>();
let no_msg_wait_millis = *batch_time_out_millis.iter().min().unwrap();
let mut all_model_predictors =
ArrayVec::<ArrayVec<BatchPredictor<T>, MAX_VERSIONS_PER_MODEL>, MAX_NUM_MODELS>::new();
let mut all_model_predictors: ArrayVec::<ArrayVec<BatchPredictor<T>, MAX_VERSIONS_PER_MODEL>, MAX_NUM_MODELS> =
(0 ..MAX_NUM_MODELS).map( |_| ArrayVec::<BatchPredictor<T>, MAX_VERSIONS_PER_MODEL>::new()).collect();
loop {
let msg = rx.try_recv();
let no_more_msg = match msg {
Ok(PredictMessage::Predict(model_spec_at, version, val, resp, ts)) => {
if let Some(model_predictors) = all_model_predictors.get_mut(model_spec_at) {
match version {
None => model_predictors[0].push(val, resp, ts),
Some(the_version) => match model_predictors
.iter_mut()
.find(|x| x.model.version() == the_version)
{
None => resp
.send(PredictResult::ModelVersionNotFound(
model_spec_at,
the_version,
))
.unwrap_or_else(|e| {
error!("cannot send back version error: {:?}", e)
}),
Some(predictor) => predictor.push(val, resp, ts),
},
if model_predictors.is_empty() {
resp.send(PredictResult::ModelNotReady(model_spec_at))
.unwrap_or_else(|e| error!("cannot send back model not ready error: {:?}", e));
}
else {
match version {
None => model_predictors[0].push(val, resp, ts),
Some(the_version) => match model_predictors
.iter_mut()
.find(|x| x.model.version() == the_version)
{
None => resp
.send(PredictResult::ModelVersionNotFound(
model_spec_at,
the_version,
))
.unwrap_or_else(|e| {
error!("cannot send back version error: {:?}", e)
}),
Some(predictor) => predictor.push(val, resp, ts),
},
}
}
} else {
resp.send(PredictResult::ModelNotFound(model_spec_at))
.unwrap_or_else(|e| error!("cannot send back model error: {:?}", e))
.unwrap_or_else(|e| error!("cannot send back model not found error: {:?}", e))
}
MPSC_CHANNEL_SIZE.dec();
false
@ -266,27 +272,23 @@ impl<T: Model> PredictService<T> {
queue_reset_ts: Instant::now(),
queue_earliest_rq_ts: Instant::now(),
};
if idx < all_model_predictors.len() {
metrics::NEW_MODEL_SNAPSHOT
.with_label_values(&[&MODEL_SPECS[idx]])
.inc();
assert!(idx < all_model_predictors.len());
metrics::NEW_MODEL_SNAPSHOT
.with_label_values(&[&MODEL_SPECS[idx]])
.inc();
info!("now we serve updated model: {}", predictor.model);
//we can do this since the vector is small
let predictors = &mut all_model_predictors[idx];
if predictors.len() == ARGS.versions_per_model {
predictors.remove(predictors.len() - 1);
}
predictors.insert(0, predictor);
} else {
info!("now we serve new model: {:}", predictor.model);
let mut predictors =
ArrayVec::<BatchPredictor<T>, MAX_VERSIONS_PER_MODEL>::new();
predictors.push(predictor);
all_model_predictors.push(predictors);
//check the invariant that we always push the last model to the end
assert_eq!(all_model_predictors.len(), idx + 1)
//we can do this since the vector is small
let predictors = &mut all_model_predictors[idx];
if predictors.len() == 0 {
info!("now we serve new model: {}", predictor.model);
}
else {
info!("now we serve updated model: {}", predictor.model);
}
if predictors.len() == ARGS.versions_per_model {
predictors.remove(predictors.len() - 1);
}
predictors.insert(0, predictor);
false
}
Err(TryRecvError::Empty) => true,

2
navi/segdense/Cargo.toml
View File

@ -3,9 +3,9 @@ name = "segdense"
version = "0.1.0"
edition = "2021"
# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
[dependencies]
env_logger = "0.10.0"
serde = { version = "1.0.104", features = ["derive"] }
serde_json = "1.0.48"
log = "0.4.17"

27
navi/segdense/src/error.rs
View File

@ -5,13 +5,13 @@ use std::fmt::Display;
*/
#[derive(Debug)]
pub enum SegDenseError {
IoError(std::io::Error),
Json(serde_json::Error),
JsonMissingRoot,
JsonMissingObject,
JsonMissingArray,
JsonArraySize,
JsonMissingInputFeature,
IoError(std::io::Error),
Json(serde_json::Error),
JsonMissingRoot,
JsonMissingObject,
JsonMissingArray,
JsonArraySize,
JsonMissingInputFeature,
}
impl Display for SegDenseError {
@ -25,19 +25,18 @@ impl Display for SegDenseError {
SegDenseError::JsonArraySize => write!(f, "SegDense JSON: Array size not as expected!"),
SegDenseError::JsonMissingInputFeature => write!(f, "SegDense JSON: Missing input feature!"),
}
}
}
impl std::error::Error for SegDenseError {}
impl From<std::io::Error> for SegDenseError {
fn from(err: std::io::Error) -> Self {
SegDenseError::IoError(err)
}
fn from(err: std::io::Error) -> Self {
SegDenseError::IoError(err)
}
}
impl From<serde_json::Error> for SegDenseError {
fn from(err: serde_json::Error) -> Self {
SegDenseError::Json(err)
}
fn from(err: serde_json::Error) -> Self {
SegDenseError::Json(err)
}
}

4
navi/segdense/src/lib.rs
View File

@ -1,4 +1,4 @@
pub mod error;
pub mod segdense_transform_spec_home_recap_2022;
pub mod mapper;
pub mod util;
pub mod segdense_transform_spec_home_recap_2022;
pub mod util;

23
navi/segdense/src/main.rs
View File

@ -5,19 +5,18 @@ use segdense::error::SegDenseError;
use segdense::util;
fn main() -> Result<(), SegDenseError> {
env_logger::init();
let args: Vec<String> = env::args().collect();
let schema_file_name: &str = if args.len() == 1 {
"json/compact.json"
} else {
&args[1]
};
env_logger::init();
let args: Vec<String> = env::args().collect();
let json_str = fs::read_to_string(schema_file_name)?;
let schema_file_name: &str = if args.len() == 1 {
"json/compact.json"
} else {
&args[1]
};
util::safe_load_config(&json_str)?;
let json_str = fs::read_to_string(schema_file_name)?;
Ok(())
util::safe_load_config(&json_str)?;
Ok(())
}

4
navi/segdense/src/mapper.rs
View File

@ -19,13 +19,13 @@ pub struct FeatureMapper {
impl FeatureMapper {
pub fn new() -> FeatureMapper {
FeatureMapper {
map: HashMap::new()
map: HashMap::new(),
}
}
}
pub trait MapWriter {
fn set(&mut self, feature_id: i64, info: FeatureInfo);
fn set(&mut self, feature_id: i64, info: FeatureInfo);
}
pub trait MapReader {

1
navi/segdense/src/segdense_transform_spec_home_recap_2022.rs
View File

@ -164,7 +164,6 @@ pub struct ComplexFeatureTypeTransformSpec {
pub tensor_shape: Vec<i64>,
}
#[derive(Default, Debug, Clone, PartialEq, Serialize, Deserialize)]
#[serde(rename_all = "camelCase")]
pub struct InputFeatureMapRecord {

43
navi/segdense/src/util.rs
View File

@ -1,10 +1,10 @@
use log::debug;
use std::fs;
use log::{debug};
use serde_json::{Value, Map};
use serde_json::{Map, Value};
use crate::error::SegDenseError;
use crate::mapper::{FeatureMapper, FeatureInfo, MapWriter};
use crate::mapper::{FeatureInfo, FeatureMapper, MapWriter};
use crate::segdense_transform_spec_home_recap_2022::{self as seg_dense, InputFeature};
pub fn load_config(file_name: &str) -> seg_dense::Root {
@ -42,15 +42,8 @@ pub fn safe_load_config(json_str: &str) -> Result<FeatureMapper, SegDenseError>
load_from_parsed_config(root)
}
pub fn load_from_parsed_config_ref(root: &seg_dense::Root) -> FeatureMapper {
load_from_parsed_config(root.clone()).unwrap_or_else(
|error| panic!("Error loading all_config.json - {}", error))
}
// Perf note : make 'root' un-owned
pub fn load_from_parsed_config(root: seg_dense::Root) ->
Result<FeatureMapper, SegDenseError> {
pub fn load_from_parsed_config(root: seg_dense::Root) -> Result<FeatureMapper, SegDenseError> {
let v = root.input_features_map;
// Do error check
@ -84,7 +77,7 @@ pub fn load_from_parsed_config(root: seg_dense::Root) ->
Some(info) => {
debug!("{:?}", info);
fm.set(feature_id, info)
},
}
None => (),
}
}
@ -92,19 +85,22 @@ pub fn load_from_parsed_config(root: seg_dense::Root) ->
Ok(fm)
}
#[allow(dead_code)]
fn add_feature_info_to_mapper(feature_mapper: &mut FeatureMapper, input_features: &Vec<InputFeature>) {
fn add_feature_info_to_mapper(
feature_mapper: &mut FeatureMapper,
input_features: &Vec<InputFeature>,
) {
for input_feature in input_features.iter() {
let feature_id = input_feature.feature_id;
let feature_info = to_feature_info(input_feature);
match feature_info {
Some(info) => {
debug!("{:?}", info);
feature_mapper.set(feature_id, info)
},
None => (),
let feature_id = input_feature.feature_id;
let feature_info = to_feature_info(input_feature);
match feature_info {
Some(info) => {
debug!("{:?}", info);
feature_mapper.set(feature_id, info)
}
None => (),
}
}
}
pub fn to_feature_info(input_feature: &seg_dense::InputFeature) -> Option<FeatureInfo> {
@ -137,7 +133,7 @@ pub fn to_feature_info(input_feature: &seg_dense::InputFeature) -> Option<Featur
2 => 0,
3 => 2,
_ => -1,
}
},
};
if input_feature.index < 0 {
@ -154,4 +150,3 @@ pub fn to_feature_info(input_feature: &seg_dense::InputFeature) -> Option<Featur
index_within_tensor: input_feature.index,
})
}

1
representation-manager/BUILD.bazel Normal file
View File

@ -0,0 +1 @@
# This prevents SQ query from grabbing //:all since it traverses up once to find a BUILD

4
representation-manager/README.md Normal file
View File

@ -0,0 +1,4 @@
# Representation Manager #
**Representation Manager** (RMS) serves as a centralized embedding management system, providing SimClusters or other embeddings as facade of the underlying storage or services.

4
representation-manager/bin/deploy.sh Executable file
View File

@ -0,0 +1,4 @@
#!/usr/bin/env bash
JOB=representation-manager bazel run --ui_event_filters=-info,-stdout,-stderr --noshow_progress \
//relevance-platform/src/main/python/deploy -- "$@"

17
representation-manager/client/src/main/scala/com/twitter/representation_manager/BUILD Normal file
View File

@ -0,0 +1,17 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"finatra/inject/inject-thrift-client",
"frigate/frigate-common/src/main/scala/com/twitter/frigate/common/store/strato",
"hermit/hermit-core/src/main/scala/com/twitter/hermit/store/common",
"relevance-platform/src/main/scala/com/twitter/relevance_platform/common/readablestore",
"representation-manager/client/src/main/scala/com/twitter/representation_manager/config",
"representation-manager/server/src/main/thrift:thrift-scala",
"src/scala/com/twitter/simclusters_v2/common",
"src/thrift/com/twitter/simclusters_v2:simclusters_v2-thrift-scala",
"stitch/stitch-storehaus",
"strato/src/main/scala/com/twitter/strato/client",
],
)

208
representation-manager/client/src/main/scala/com/twitter/representation_manager/StoreBuilder.scala Normal file
View File

@ -0,0 +1,208 @@
package com.twitter.representation_manager
import com.twitter.finagle.memcached.{Client => MemcachedClient}
import com.twitter.finagle.stats.StatsReceiver
import com.twitter.frigate.common.store.strato.StratoFetchableStore
import com.twitter.hermit.store.common.ObservedCachedReadableStore
import com.twitter.hermit.store.common.ObservedReadableStore
import com.twitter.representation_manager.config.ClientConfig
import com.twitter.representation_manager.config.DisabledInMemoryCacheParams
import com.twitter.representation_manager.config.EnabledInMemoryCacheParams
import com.twitter.representation_manager.thriftscala.SimClustersEmbeddingView
import com.twitter.simclusters_v2.common.SimClustersEmbedding
import com.twitter.simclusters_v2.thriftscala.InternalId
import com.twitter.simclusters_v2.thriftscala.LocaleEntityId
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.simclusters_v2.thriftscala.TopicId
import com.twitter.simclusters_v2.thriftscala.{SimClustersEmbedding => ThriftSimClustersEmbedding}
import com.twitter.storehaus.ReadableStore
import com.twitter.strato.client.{Client => StratoClient}
import com.twitter.strato.thrift.ScroogeConvImplicits._
/**
* This is the class that offers features to build readable stores for a given
* SimClustersEmbeddingView (i.e. embeddingType and modelVersion). It applies ClientConfig
* for a particular service and build ReadableStores which implement that config.
*/
class StoreBuilder(
clientConfig: ClientConfig,
stratoClient: StratoClient,
memCachedClient: MemcachedClient,
globalStats: StatsReceiver,
) {
private val stats =
globalStats.scope("representation_manager_client").scope(this.getClass.getSimpleName)
// Column consts
private val ColPathPrefix = "recommendations/representation_manager/"
private val SimclustersTweetColPath = ColPathPrefix + "simClustersEmbedding.Tweet"
private val SimclustersUserColPath = ColPathPrefix + "simClustersEmbedding.User"
private val SimclustersTopicIdColPath = ColPathPrefix + "simClustersEmbedding.TopicId"
private val SimclustersLocaleEntityIdColPath =
ColPathPrefix + "simClustersEmbedding.LocaleEntityId"
def buildSimclustersTweetEmbeddingStore(
embeddingColumnView: SimClustersEmbeddingView
): ReadableStore[Long, SimClustersEmbedding] = {
val rawStore = StratoFetchableStore
.withView[Long, SimClustersEmbeddingView, ThriftSimClustersEmbedding](
stratoClient,
SimclustersTweetColPath,
embeddingColumnView)
.mapValues(SimClustersEmbedding(_))
addCacheLayer(rawStore, embeddingColumnView)
}
def buildSimclustersUserEmbeddingStore(
embeddingColumnView: SimClustersEmbeddingView
): ReadableStore[Long, SimClustersEmbedding] = {
val rawStore = StratoFetchableStore
.withView[Long, SimClustersEmbeddingView, ThriftSimClustersEmbedding](
stratoClient,
SimclustersUserColPath,
embeddingColumnView)
.mapValues(SimClustersEmbedding(_))
addCacheLayer(rawStore, embeddingColumnView)
}
def buildSimclustersTopicIdEmbeddingStore(
embeddingColumnView: SimClustersEmbeddingView
): ReadableStore[TopicId, SimClustersEmbedding] = {
val rawStore = StratoFetchableStore
.withView[TopicId, SimClustersEmbeddingView, ThriftSimClustersEmbedding](
stratoClient,
SimclustersTopicIdColPath,
embeddingColumnView)
.mapValues(SimClustersEmbedding(_))
addCacheLayer(rawStore, embeddingColumnView)
}
def buildSimclustersLocaleEntityIdEmbeddingStore(
embeddingColumnView: SimClustersEmbeddingView
): ReadableStore[LocaleEntityId, SimClustersEmbedding] = {
val rawStore = StratoFetchableStore
.withView[LocaleEntityId, SimClustersEmbeddingView, ThriftSimClustersEmbedding](
stratoClient,
SimclustersLocaleEntityIdColPath,
embeddingColumnView)
.mapValues(SimClustersEmbedding(_))
addCacheLayer(rawStore, embeddingColumnView)
}
def buildSimclustersTweetEmbeddingStoreWithEmbeddingIdAsKey(
embeddingColumnView: SimClustersEmbeddingView
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val rawStore = StratoFetchableStore
.withView[Long, SimClustersEmbeddingView, ThriftSimClustersEmbedding](
stratoClient,
SimclustersTweetColPath,
embeddingColumnView)
.mapValues(SimClustersEmbedding(_))
val embeddingIdAsKeyStore = rawStore.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.TweetId(tweetId)) =>
tweetId
}
addCacheLayer(embeddingIdAsKeyStore, embeddingColumnView)
}
def buildSimclustersUserEmbeddingStoreWithEmbeddingIdAsKey(
embeddingColumnView: SimClustersEmbeddingView
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val rawStore = StratoFetchableStore
.withView[Long, SimClustersEmbeddingView, ThriftSimClustersEmbedding](
stratoClient,
SimclustersUserColPath,
embeddingColumnView)
.mapValues(SimClustersEmbedding(_))
val embeddingIdAsKeyStore = rawStore.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.UserId(userId)) =>
userId
}
addCacheLayer(embeddingIdAsKeyStore, embeddingColumnView)
}
def buildSimclustersTopicEmbeddingStoreWithEmbeddingIdAsKey(
embeddingColumnView: SimClustersEmbeddingView
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val rawStore = StratoFetchableStore
.withView[TopicId, SimClustersEmbeddingView, ThriftSimClustersEmbedding](
stratoClient,
SimclustersTopicIdColPath,
embeddingColumnView)
.mapValues(SimClustersEmbedding(_))
val embeddingIdAsKeyStore = rawStore.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.TopicId(topicId)) =>
topicId
}
addCacheLayer(embeddingIdAsKeyStore, embeddingColumnView)
}
def buildSimclustersTopicIdEmbeddingStoreWithEmbeddingIdAsKey(
embeddingColumnView: SimClustersEmbeddingView
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val rawStore = StratoFetchableStore
.withView[TopicId, SimClustersEmbeddingView, ThriftSimClustersEmbedding](
stratoClient,
SimclustersTopicIdColPath,
embeddingColumnView)
.mapValues(SimClustersEmbedding(_))
val embeddingIdAsKeyStore = rawStore.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.TopicId(topicId)) =>
topicId
}
addCacheLayer(embeddingIdAsKeyStore, embeddingColumnView)
}
def buildSimclustersLocaleEntityIdEmbeddingStoreWithEmbeddingIdAsKey(
embeddingColumnView: SimClustersEmbeddingView
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val rawStore = StratoFetchableStore
.withView[LocaleEntityId, SimClustersEmbeddingView, ThriftSimClustersEmbedding](
stratoClient,
SimclustersLocaleEntityIdColPath,
embeddingColumnView)
.mapValues(SimClustersEmbedding(_))
val embeddingIdAsKeyStore = rawStore.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.LocaleEntityId(localeEntityId)) =>
localeEntityId
}
addCacheLayer(embeddingIdAsKeyStore, embeddingColumnView)
}
private def addCacheLayer[K](
rawStore: ReadableStore[K, SimClustersEmbedding],
embeddingColumnView: SimClustersEmbeddingView,
): ReadableStore[K, SimClustersEmbedding] = {
// Add in-memory caching based on ClientConfig
val inMemCacheParams = clientConfig.inMemoryCacheConfig
.getCacheSetup(embeddingColumnView.embeddingType, embeddingColumnView.modelVersion)
val statsPerStore = stats
.scope(embeddingColumnView.embeddingType.name).scope(embeddingColumnView.modelVersion.name)
inMemCacheParams match {
case DisabledInMemoryCacheParams =>
ObservedReadableStore(
store = rawStore
)(statsPerStore)
case EnabledInMemoryCacheParams(ttl, maxKeys, cacheName) =>
ObservedCachedReadableStore.from[K, SimClustersEmbedding](
rawStore,
ttl = ttl,
maxKeys = maxKeys,
cacheName = cacheName,
windowSize = 10000L
)(statsPerStore)
}
}
}

12
representation-manager/client/src/main/scala/com/twitter/representation_manager/config/BUILD Normal file
View File

@ -0,0 +1,12 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"finatra/inject/inject-thrift-client",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/common",
"representation-manager/server/src/main/thrift:thrift-scala",
"src/thrift/com/twitter/simclusters_v2:simclusters_v2-thrift-scala",
"strato/src/main/scala/com/twitter/strato/client",
],
)

25
representation-manager/client/src/main/scala/com/twitter/representation_manager/config/ClientConfig.scala Normal file
View File

@ -0,0 +1,25 @@
package com.twitter.representation_manager.config
import com.twitter.simclusters_v2.thriftscala.EmbeddingType
import com.twitter.simclusters_v2.thriftscala.ModelVersion
/*
* This is RMS client config class.
* We only support setting up in memory cache params for now, but we expect to enable other
* customisations in the near future e.g. request timeout
*
* --------------------------------------------
* PLEASE NOTE:
* Having in-memory cache is not necessarily a free performance win, anyone considering it should
* investigate rather than blindly enabling it
* */
class ClientConfig(inMemCacheParamsOverrides: Map[
(EmbeddingType, ModelVersion),
InMemoryCacheParams
] = Map.empty) {
// In memory cache config per embedding
val inMemCacheParams = DefaultInMemoryCacheConfig.cacheParamsMap ++ inMemCacheParamsOverrides
val inMemoryCacheConfig = new InMemoryCacheConfig(inMemCacheParams)
}
object DefaultClientConfig extends ClientConfig

53
representation-manager/client/src/main/scala/com/twitter/representation_manager/config/InMemoryCacheConfig.scala Normal file
View File

@ -0,0 +1,53 @@
package com.twitter.representation_manager.config
import com.twitter.simclusters_v2.thriftscala.EmbeddingType
import com.twitter.simclusters_v2.thriftscala.ModelVersion
import com.twitter.util.Duration
/*
* --------------------------------------------
* PLEASE NOTE:
* Having in-memory cache is not necessarily a free performance win, anyone considering it should
* investigate rather than blindly enabling it
* --------------------------------------------
* */
sealed trait InMemoryCacheParams
/*
* This holds params that is required to set up a in-mem cache for a single embedding store
*/
case class EnabledInMemoryCacheParams(
ttl: Duration,
maxKeys: Int,
cacheName: String)
extends InMemoryCacheParams
object DisabledInMemoryCacheParams extends InMemoryCacheParams
/*
* This is the class for the in-memory cache config. Client could pass in their own cacheParamsMap to
* create a new InMemoryCacheConfig instead of using the DefaultInMemoryCacheConfig object below
* */
class InMemoryCacheConfig(
cacheParamsMap: Map[
(EmbeddingType, ModelVersion),
InMemoryCacheParams
] = Map.empty) {
def getCacheSetup(
embeddingType: EmbeddingType,
modelVersion: ModelVersion
): InMemoryCacheParams = {
// When requested embedding type doesn't exist, we return DisabledInMemoryCacheParams
cacheParamsMap.getOrElse((embeddingType, modelVersion), DisabledInMemoryCacheParams)
}
}
/*
* Default config for the in-memory cache
* Clients can directly import and use this one if they don't want to set up a customised config
* */
object DefaultInMemoryCacheConfig extends InMemoryCacheConfig {
// set default to no in-memory caching
val cacheParamsMap = Map.empty
}

21
representation-manager/server/BUILD Normal file
View File

@ -0,0 +1,21 @@
jvm_binary(
name = "bin",
basename = "representation-manager",
main = "com.twitter.representation_manager.RepresentationManagerFedServerMain",
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"finatra/inject/inject-logback/src/main/scala",
"loglens/loglens-logback/src/main/scala/com/twitter/loglens/logback",
"representation-manager/server/src/main/resources",
"representation-manager/server/src/main/scala/com/twitter/representation_manager",
"twitter-server/logback-classic/src/main/scala",
],
)
# Aurora Workflows build phase convention requires a jvm_app named with ${project-name}-app
jvm_app(
name = "representation-manager-app",
archive = "zip",
binary = ":bin",
)

7
representation-manager/server/src/main/resources/BUILD Normal file
View File

@ -0,0 +1,7 @@
resources(
sources = [
"*.xml",
"config/*.yml",
],
tags = ["bazel-compatible"],
)

219
representation-manager/server/src/main/resources/config/decider.yml Normal file
View File

@ -0,0 +1,219 @@
# ---------- traffic percentage by embedding type and model version ----------
# Decider strings are build dynamically following the rule in there
# i.e. s"enable_${embeddingType.name}_${modelVersion.name}"
# Hence this should be updated accordingly if usage is changed in the embedding stores
# Tweet embeddings
"enable_LogFavBasedTweet_Model20m145k2020":
comment: "Enable x% read traffic (0<=x<=10000, e.g. 1000=10%) for LogFavBasedTweet - Model20m145k2020. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavBasedTweet_Model20m145kUpdated":
comment: "Enable x% read traffic (0<=x<=10000, e.g. 1000=10%) for LogFavBasedTweet - Model20m145kUpdated. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavLongestL2EmbeddingTweet_Model20m145k2020":
comment: "Enable x% read traffic (0<=x<=10000, e.g. 1000=10%) for LogFavLongestL2EmbeddingTweet - Model20m145k2020. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavLongestL2EmbeddingTweet_Model20m145kUpdated":
comment: "Enable x% read traffic (0<=x<=10000, e.g. 1000=10%) for LogFavLongestL2EmbeddingTweet - Model20m145kUpdated. 0 means return EMPTY for all requests."
default_availability: 10000
# Topic embeddings
"enable_FavTfgTopic_Model20m145k2020":
comment: "Enable the read traffic to FavTfgTopic - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavBasedKgoApeTopic_Model20m145k2020":
comment: "Enable the read traffic to LogFavBasedKgoApeTopic - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
# User embeddings - KnownFor
"enable_FavBasedProducer_Model20m145kUpdated":
comment: "Enable the read traffic to FavBasedProducer - Model20m145kUpdated from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_FavBasedProducer_Model20m145k2020":
comment: "Enable the read traffic to FavBasedProducer - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_FollowBasedProducer_Model20m145k2020":
comment: "Enable the read traffic to FollowBasedProducer - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_AggregatableFavBasedProducer_Model20m145kUpdated":
comment: "Enable the read traffic to AggregatableFavBasedProducer - Model20m145kUpdated from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_AggregatableFavBasedProducer_Model20m145k2020":
comment: "Enable the read traffic to AggregatableFavBasedProducer - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_AggregatableLogFavBasedProducer_Model20m145kUpdated":
comment: "Enable the read traffic to AggregatableLogFavBasedProducer - Model20m145kUpdated from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_AggregatableLogFavBasedProducer_Model20m145k2020":
comment: "Enable the read traffic to AggregatableLogFavBasedProducer - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
enable_RelaxedAggregatableLogFavBasedProducer_Model20m145kUpdated:
comment: "Enable the read traffic to RelaxedAggregatableLogFavBasedProducer - Model20m145kUpdated from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
enable_RelaxedAggregatableLogFavBasedProducer_Model20m145k2020:
comment: "Enable the read traffic to RelaxedAggregatableLogFavBasedProducer - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
# User embeddings - InterestedIn
"enable_LogFavBasedUserInterestedInFromAPE_Model20m145k2020":
comment: "Enable the read traffic to LogFavBasedUserInterestedInFromAPE - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_FollowBasedUserInterestedInFromAPE_Model20m145k2020":
comment: "Enable the read traffic to FollowBasedUserInterestedInFromAPE - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_FavBasedUserInterestedIn_Model20m145kUpdated":
comment: "Enable the read traffic to FavBasedUserInterestedIn - Model20m145kUpdated from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_FavBasedUserInterestedIn_Model20m145k2020":
comment: "Enable the read traffic to FavBasedUserInterestedIn - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_FollowBasedUserInterestedIn_Model20m145k2020":
comment: "Enable the read traffic to FollowBasedUserInterestedIn - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavBasedUserInterestedIn_Model20m145k2020":
comment: "Enable the read traffic to LogFavBasedUserInterestedIn - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_FavBasedUserInterestedInFromPE_Model20m145kUpdated":
comment: "Enable the read traffic to FavBasedUserInterestedInFromPE - Model20m145kUpdated from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_FilteredUserInterestedIn_Model20m145kUpdated":
comment: "Enable the read traffic to FilteredUserInterestedIn - Model20m145kUpdated from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_FilteredUserInterestedIn_Model20m145k2020":
comment: "Enable the read traffic to FilteredUserInterestedIn - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_FilteredUserInterestedInFromPE_Model20m145kUpdated":
comment: "Enable the read traffic to FilteredUserInterestedInFromPE - Model20m145kUpdated from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_UnfilteredUserInterestedIn_Model20m145kUpdated":
comment: "Enable the read traffic to UnfilteredUserInterestedIn - Model20m145kUpdated from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_UnfilteredUserInterestedIn_Model20m145k2020":
comment: "Enable the read traffic to UnfilteredUserInterestedIn - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_UserNextInterestedIn_Model20m145k2020":
comment: "Enable the read traffic to UserNextInterestedIn - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavBasedUserInterestedMaxpoolingAddressBookFromIIAPE_Model20m145k2020":
comment: "Enable the read traffic to LogFavBasedUserInterestedMaxpoolingAddressBookFromIIAPE - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavBasedUserInterestedAverageAddressBookFromIIAPE_Model20m145k2020":
comment: "Enable the read traffic to LogFavBasedUserInterestedAverageAddressBookFromIIAPE - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavBasedUserInterestedBooktypeMaxpoolingAddressBookFromIIAPE_Model20m145k2020":
comment: "Enable the read traffic to LogFavBasedUserInterestedMaxpoolingAddressBookFromIIAPE - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavBasedUserInterestedLargestDimMaxpoolingAddressBookFromIIAPE_Model20m145k2020":
comment: "Enable the read traffic to LogFavBasedUserInterestedAverageAddressBookFromIIAPE - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavBasedUserInterestedLouvainMaxpoolingAddressBookFromIIAPE_Model20m145k2020":
comment: "Enable the read traffic to LogFavBasedUserInterestedMaxpoolingAddressBookFromIIAPE - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
"enable_LogFavBasedUserInterestedConnectedMaxpoolingAddressBookFromIIAPE_Model20m145k2020":
comment: "Enable the read traffic to LogFavBasedUserInterestedAverageAddressBookFromIIAPE - Model20m145k2020 from 0% to 100%. 0 means return EMPTY for all requests."
default_availability: 10000
# ---------- load shedding by caller id ----------
# To create a new decider, add here with the same format and caller's details :
# "representation-manager_load_shed_by_caller_id_twtr:{{role}}:{{name}}:{{environment}}:{{cluster}}"
# All the deciders below are generated by this script:
# ./strato/bin/fed deciders representation-manager --service-role=representation-manager --service-name=representation-manager
# If you need to run the script and paste the output, add ONLY the prod deciders here.
"representation-manager_load_shed_by_caller_id_all":
comment: "Reject all traffic from caller id: all"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:cr-mixer:cr-mixer:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:cr-mixer:cr-mixer:prod:atla"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:cr-mixer:cr-mixer:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:cr-mixer:cr-mixer:prod:pdxa"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:simclusters-ann:simclusters-ann-1:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:simclusters-ann:simclusters-ann-1:prod:atla"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:simclusters-ann:simclusters-ann-1:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:simclusters-ann:simclusters-ann-1:prod:pdxa"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:simclusters-ann:simclusters-ann-3:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:simclusters-ann:simclusters-ann-3:prod:atla"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:simclusters-ann:simclusters-ann-3:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:simclusters-ann:simclusters-ann-3:prod:pdxa"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:simclusters-ann:simclusters-ann-4:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:simclusters-ann:simclusters-ann-4:prod:atla"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:simclusters-ann:simclusters-ann-4:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:simclusters-ann:simclusters-ann-4:prod:pdxa"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:simclusters-ann:simclusters-ann-experimental:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:simclusters-ann:simclusters-ann-experimental:prod:atla"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:simclusters-ann:simclusters-ann-experimental:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:simclusters-ann:simclusters-ann-experimental:prod:pdxa"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:simclusters-ann:simclusters-ann:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:simclusters-ann:simclusters-ann:prod:atla"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:simclusters-ann:simclusters-ann:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:simclusters-ann:simclusters-ann:prod:pdxa"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:stratostore:stratoapi:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:stratostore:stratoapi:prod:atla"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:stratostore:stratoserver:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:stratostore:stratoserver:prod:atla"
default_availability: 0
"representation-manager_load_shed_by_caller_id_twtr:svc:stratostore:stratoserver:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:stratostore:stratoserver:prod:pdxa"
default_availability: 0
# ---------- Dark Traffic Proxy ----------
representation-manager_forward_dark_traffic:
comment: "Defines the percentage of traffic to forward to diffy-proxy. Set to 0 to disable dark traffic forwarding"
default_availability: 0

165
representation-manager/server/src/main/resources/logback.xml Normal file
View File

@ -0,0 +1,165 @@
<configuration>
<shutdownHook class="ch.qos.logback.core.hook.DelayingShutdownHook"/>
<!-- ===================================================== -->
<!-- Service Config -->
<!-- ===================================================== -->
<property name="DEFAULT_SERVICE_PATTERN"
value="%-16X{traceId} %-12X{clientId:--} %-16X{method} %-25logger{0} %msg"/>
<property name="DEFAULT_ACCESS_PATTERN"
value="%msg"/>
<!-- ===================================================== -->
<!-- Common Config -->
<!-- ===================================================== -->
<!-- JUL/JDK14 to Logback bridge -->
<contextListener class="ch.qos.logback.classic.jul.LevelChangePropagator">
<resetJUL>true</resetJUL>
</contextListener>
<!-- ====================================================================================== -->
<!-- NOTE: The following appenders use a simple TimeBasedRollingPolicy configuration. -->
<!-- You may want to consider using a more advanced SizeAndTimeBasedRollingPolicy. -->
<!-- See: https://logback.qos.ch/manual/appenders.html#SizeAndTimeBasedRollingPolicy -->
<!-- ====================================================================================== -->
<!-- Service Log (rollover daily, keep maximum of 21 days of gzip compressed logs) -->
<appender name="SERVICE" class="ch.qos.logback.core.rolling.RollingFileAppender">
<file>${log.service.output}</file>
<rollingPolicy class="ch.qos.logback.core.rolling.TimeBasedRollingPolicy">
<!-- daily rollover -->
<fileNamePattern>${log.service.output}.%d.gz</fileNamePattern>
<!-- the maximum total size of all the log files -->
<totalSizeCap>3GB</totalSizeCap>
<!-- keep maximum 21 days' worth of history -->
<maxHistory>21</maxHistory>
<cleanHistoryOnStart>true</cleanHistoryOnStart>
</rollingPolicy>
<encoder>
<pattern>%date %.-3level ${DEFAULT_SERVICE_PATTERN}%n</pattern>
</encoder>
</appender>
<!-- Access Log (rollover daily, keep maximum of 21 days of gzip compressed logs) -->
<appender name="ACCESS" class="ch.qos.logback.core.rolling.RollingFileAppender">
<file>${log.access.output}</file>
<rollingPolicy class="ch.qos.logback.core.rolling.TimeBasedRollingPolicy">
<!-- daily rollover -->
<fileNamePattern>${log.access.output}.%d.gz</fileNamePattern>
<!-- the maximum total size of all the log files -->
<totalSizeCap>100MB</totalSizeCap>
<!-- keep maximum 7 days' worth of history -->
<maxHistory>7</maxHistory>
<cleanHistoryOnStart>true</cleanHistoryOnStart>
</rollingPolicy>
<encoder>
<pattern>${DEFAULT_ACCESS_PATTERN}%n</pattern>
</encoder>
</appender>
<!--LogLens -->
<appender name="LOGLENS" class="com.twitter.loglens.logback.LoglensAppender">
<mdcAdditionalContext>true</mdcAdditionalContext>
<category>${log.lens.category}</category>
<index>${log.lens.index}</index>
<tag>${log.lens.tag}/service</tag>
<encoder>
<pattern>%msg</pattern>
</encoder>
</appender>
<!-- LogLens Access -->
<appender name="LOGLENS-ACCESS" class="com.twitter.loglens.logback.LoglensAppender">
<mdcAdditionalContext>true</mdcAdditionalContext>
<category>${log.lens.category}</category>
<index>${log.lens.index}</index>
<tag>${log.lens.tag}/access</tag>
<encoder>
<pattern>%msg</pattern>
</encoder>
</appender>
<!-- Pipeline Execution Logs -->
<appender name="ALLOW-LISTED-PIPELINE-EXECUTIONS" class="ch.qos.logback.core.rolling.RollingFileAppender">
<file>allow_listed_pipeline_executions.log</file>
<rollingPolicy class="ch.qos.logback.core.rolling.TimeBasedRollingPolicy">
<!-- daily rollover -->
<fileNamePattern>allow_listed_pipeline_executions.log.%d.gz</fileNamePattern>
<!-- the maximum total size of all the log files -->
<totalSizeCap>100MB</totalSizeCap>
<!-- keep maximum 7 days' worth of history -->
<maxHistory>7</maxHistory>
<cleanHistoryOnStart>true</cleanHistoryOnStart>
</rollingPolicy>
<encoder>
<pattern>%date %.-3level ${DEFAULT_SERVICE_PATTERN}%n</pattern>
</encoder>
</appender>
<!-- ===================================================== -->
<!-- Primary Async Appenders -->
<!-- ===================================================== -->
<property name="async_queue_size" value="${queue.size:-50000}"/>
<property name="async_max_flush_time" value="${max.flush.time:-0}"/>
<appender name="ASYNC-SERVICE" class="com.twitter.inject.logback.AsyncAppender">
<queueSize>${async_queue_size}</queueSize>
<maxFlushTime>${async_max_flush_time}</maxFlushTime>
<appender-ref ref="SERVICE"/>
</appender>
<appender name="ASYNC-ACCESS" class="com.twitter.inject.logback.AsyncAppender">
<queueSize>${async_queue_size}</queueSize>
<maxFlushTime>${async_max_flush_time}</maxFlushTime>
<appender-ref ref="ACCESS"/>
</appender>
<appender name="ASYNC-ALLOW-LISTED-PIPELINE-EXECUTIONS" class="com.twitter.inject.logback.AsyncAppender">
<queueSize>${async_queue_size}</queueSize>
<maxFlushTime>${async_max_flush_time}</maxFlushTime>
<appender-ref ref="ALLOW-LISTED-PIPELINE-EXECUTIONS"/>
</appender>
<appender name="ASYNC-LOGLENS" class="com.twitter.inject.logback.AsyncAppender">
<queueSize>${async_queue_size}</queueSize>
<maxFlushTime>${async_max_flush_time}</maxFlushTime>
<appender-ref ref="LOGLENS"/>
</appender>
<appender name="ASYNC-LOGLENS-ACCESS" class="com.twitter.inject.logback.AsyncAppender">
<queueSize>${async_queue_size}</queueSize>
<maxFlushTime>${async_max_flush_time}</maxFlushTime>
<appender-ref ref="LOGLENS-ACCESS"/>
</appender>
<!-- ===================================================== -->
<!-- Package Config -->
<!-- ===================================================== -->
<!-- Per-Package Config -->
<logger name="com.twitter" level="INHERITED"/>
<logger name="com.twitter.wilyns" level="INHERITED"/>
<logger name="com.twitter.configbus.client.file" level="INHERITED"/>
<logger name="com.twitter.finagle.mux" level="INHERITED"/>
<logger name="com.twitter.finagle.serverset2" level="INHERITED"/>
<logger name="com.twitter.logging.ScribeHandler" level="INHERITED"/>
<logger name="com.twitter.zookeeper.client.internal" level="INHERITED"/>
<!-- Root Config -->
<!-- For all logs except access logs, disable logging below log_level level by default. This can be overriden in the per-package loggers, and dynamically in the admin panel of individual instances. -->
<root level="${log_level:-INFO}">
<appender-ref ref="ASYNC-SERVICE"/>
<appender-ref ref="ASYNC-LOGLENS"/>
</root>
<!-- Access Logging -->
<!-- Access logs are turned off by default -->
<logger name="com.twitter.finatra.thrift.filters.AccessLoggingFilter" level="OFF" additivity="false">
<appender-ref ref="ASYNC-ACCESS"/>
<appender-ref ref="ASYNC-LOGLENS-ACCESS"/>
</logger>
</configuration>

13
representation-manager/server/src/main/scala/com/twitter/representation_manager/BUILD Normal file
View File

@ -0,0 +1,13 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"finatra/inject/inject-thrift-client",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/topic",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/tweet",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/user",
"strato/src/main/scala/com/twitter/strato/fed",
"strato/src/main/scala/com/twitter/strato/fed/server",
],
)

40
representation-manager/server/src/main/scala/com/twitter/representation_manager/RepresentationManagerFedServer.scala Normal file
View File

@ -0,0 +1,40 @@
package com.twitter.representation_manager
import com.google.inject.Module
import com.twitter.inject.thrift.modules.ThriftClientIdModule
import com.twitter.representation_manager.columns.topic.LocaleEntityIdSimClustersEmbeddingCol
import com.twitter.representation_manager.columns.topic.TopicIdSimClustersEmbeddingCol
import com.twitter.representation_manager.columns.tweet.TweetSimClustersEmbeddingCol
import com.twitter.representation_manager.columns.user.UserSimClustersEmbeddingCol
import com.twitter.representation_manager.modules.CacheModule
import com.twitter.representation_manager.modules.InterestsThriftClientModule
import com.twitter.representation_manager.modules.LegacyRMSConfigModule
import com.twitter.representation_manager.modules.StoreModule
import com.twitter.representation_manager.modules.TimerModule
import com.twitter.representation_manager.modules.UttClientModule
import com.twitter.strato.fed._
import com.twitter.strato.fed.server._
object RepresentationManagerFedServerMain extends RepresentationManagerFedServer
trait RepresentationManagerFedServer extends StratoFedServer {
override def dest: String = "/s/representation-manager/representation-manager"
override val modules: Seq[Module] =
Seq(
CacheModule,
InterestsThriftClientModule,
LegacyRMSConfigModule,
StoreModule,
ThriftClientIdModule,
TimerModule,
UttClientModule
)
override def columns: Seq[Class[_ <: StratoFed.Column]] =
Seq(
classOf[TweetSimClustersEmbeddingCol],
classOf[UserSimClustersEmbeddingCol],
classOf[TopicIdSimClustersEmbeddingCol],
classOf[LocaleEntityIdSimClustersEmbeddingCol]
)
}

9
representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/BUILD Normal file
View File

@ -0,0 +1,9 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"strato/src/main/scala/com/twitter/strato/fed",
"strato/src/main/scala/com/twitter/strato/fed/server",
],
)

26
representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/ColumnConfigBase.scala Normal file
View File

@ -0,0 +1,26 @@
package com.twitter.representation_manager.columns
import com.twitter.strato.access.Access.LdapGroup
import com.twitter.strato.config.ContactInfo
import com.twitter.strato.config.FromColumns
import com.twitter.strato.config.Has
import com.twitter.strato.config.Prefix
import com.twitter.strato.config.ServiceIdentifierPattern
object ColumnConfigBase {
/****************** Internal permissions *******************/
val recosPermissions: Seq[com.twitter.strato.config.Policy] = Seq()
/****************** External permissions *******************/
// This is used to grant limited access to members outside of RP team.
val externalPermissions: Seq[com.twitter.strato.config.Policy] = Seq()
val contactInfo: ContactInfo = ContactInfo(
description = "Please contact Relevance Platform for more details",
contactEmail = "no-reply@twitter.com",
ldapGroup = "ldap",
jiraProject = "JIRA",
links = Seq("http://go/rms-runbook")
)
}

14
representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/topic/BUILD Normal file
View File

@ -0,0 +1,14 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"finatra/inject/inject-core/src/main/scala",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/columns",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/modules",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/store",
"representation-manager/server/src/main/thrift:thrift-scala",
"strato/src/main/scala/com/twitter/strato/fed",
"strato/src/main/scala/com/twitter/strato/fed/server",
],
)

77
representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/topic/LocaleEntityIdSimClustersEmbeddingCol.scala Normal file
View File

@ -0,0 +1,77 @@
package com.twitter.representation_manager.columns.topic
import com.twitter.representation_manager.columns.ColumnConfigBase
import com.twitter.representation_manager.store.TopicSimClustersEmbeddingStore
import com.twitter.representation_manager.thriftscala.SimClustersEmbeddingView
import com.twitter.simclusters_v2.thriftscala.InternalId
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbedding
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.simclusters_v2.thriftscala.LocaleEntityId
import com.twitter.stitch
import com.twitter.stitch.Stitch
import com.twitter.stitch.storehaus.StitchOfReadableStore
import com.twitter.strato.catalog.OpMetadata
import com.twitter.strato.config.AnyOf
import com.twitter.strato.config.ContactInfo
import com.twitter.strato.config.FromColumns
import com.twitter.strato.config.Policy
import com.twitter.strato.config.Prefix
import com.twitter.strato.data.Conv
import com.twitter.strato.data.Description.PlainText
import com.twitter.strato.data.Lifecycle
import com.twitter.strato.fed._
import com.twitter.strato.thrift.ScroogeConv
import javax.inject.Inject
class LocaleEntityIdSimClustersEmbeddingCol @Inject() (
embeddingStore: TopicSimClustersEmbeddingStore)
extends StratoFed.Column(
"recommendations/representation_manager/simClustersEmbedding.LocaleEntityId")
with StratoFed.Fetch.Stitch {
private val storeStitch: SimClustersEmbeddingId => Stitch[SimClustersEmbedding] =
StitchOfReadableStore(embeddingStore.topicSimClustersEmbeddingStore.mapValues(_.toThrift))
val colPermissions: Seq[com.twitter.strato.config.Policy] =
ColumnConfigBase.recosPermissions ++ ColumnConfigBase.externalPermissions :+ FromColumns(
Set(
Prefix("ml/featureStore/simClusters"),
))
override val policy: Policy = AnyOf({
colPermissions
})
override type Key = LocaleEntityId
override type View = SimClustersEmbeddingView
override type Value = SimClustersEmbedding
override val keyConv: Conv[Key] = ScroogeConv.fromStruct[LocaleEntityId]
override val viewConv: Conv[View] = ScroogeConv.fromStruct[SimClustersEmbeddingView]
override val valueConv: Conv[Value] = ScroogeConv.fromStruct[SimClustersEmbedding]
override val contactInfo: ContactInfo = ColumnConfigBase.contactInfo
override val metadata: OpMetadata = OpMetadata(
lifecycle = Some(Lifecycle.Production),
description = Some(
PlainText(
"The Topic SimClusters Embedding Endpoint in Representation Management Service with LocaleEntityId." +
" TDD: http://go/rms-tdd"))
)
override def fetch(key: Key, view: View): Stitch[Result[Value]] = {
val embeddingId = SimClustersEmbeddingId(
view.embeddingType,
view.modelVersion,
InternalId.LocaleEntityId(key)
)
storeStitch(embeddingId)
.map(embedding => found(embedding))
.handle {
case stitch.NotFound => missing
}
}
}

74
representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/topic/TopicIdSimClustersEmbeddingCol.scala Normal file
View File

@ -0,0 +1,74 @@
package com.twitter.representation_manager.columns.topic
import com.twitter.representation_manager.columns.ColumnConfigBase
import com.twitter.representation_manager.store.TopicSimClustersEmbeddingStore
import com.twitter.representation_manager.thriftscala.SimClustersEmbeddingView
import com.twitter.simclusters_v2.thriftscala.InternalId
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbedding
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.simclusters_v2.thriftscala.TopicId
import com.twitter.stitch
import com.twitter.stitch.Stitch
import com.twitter.stitch.storehaus.StitchOfReadableStore
import com.twitter.strato.catalog.OpMetadata
import com.twitter.strato.config.AnyOf
import com.twitter.strato.config.ContactInfo
import com.twitter.strato.config.FromColumns
import com.twitter.strato.config.Policy
import com.twitter.strato.config.Prefix
import com.twitter.strato.data.Conv
import com.twitter.strato.data.Description.PlainText
import com.twitter.strato.data.Lifecycle
import com.twitter.strato.fed._
import com.twitter.strato.thrift.ScroogeConv
import javax.inject.Inject
class TopicIdSimClustersEmbeddingCol @Inject() (embeddingStore: TopicSimClustersEmbeddingStore)
extends StratoFed.Column("recommendations/representation_manager/simClustersEmbedding.TopicId")
with StratoFed.Fetch.Stitch {
private val storeStitch: SimClustersEmbeddingId => Stitch[SimClustersEmbedding] =
StitchOfReadableStore(embeddingStore.topicSimClustersEmbeddingStore.mapValues(_.toThrift))
val colPermissions: Seq[com.twitter.strato.config.Policy] =
ColumnConfigBase.recosPermissions ++ ColumnConfigBase.externalPermissions :+ FromColumns(
Set(
Prefix("ml/featureStore/simClusters"),
))
override val policy: Policy = AnyOf({
colPermissions
})
override type Key = TopicId
override type View = SimClustersEmbeddingView
override type Value = SimClustersEmbedding
override val keyConv: Conv[Key] = ScroogeConv.fromStruct[TopicId]
override val viewConv: Conv[View] = ScroogeConv.fromStruct[SimClustersEmbeddingView]
override val valueConv: Conv[Value] = ScroogeConv.fromStruct[SimClustersEmbedding]
override val contactInfo: ContactInfo = ColumnConfigBase.contactInfo
override val metadata: OpMetadata = OpMetadata(
lifecycle = Some(Lifecycle.Production),
description = Some(PlainText(
"The Topic SimClusters Embedding Endpoint in Representation Management Service with TopicId." +
" TDD: http://go/rms-tdd"))
)
override def fetch(key: Key, view: View): Stitch[Result[Value]] = {
val embeddingId = SimClustersEmbeddingId(
view.embeddingType,
view.modelVersion,
InternalId.TopicId(key)
)
storeStitch(embeddingId)
.map(embedding => found(embedding))
.handle {
case stitch.NotFound => missing
}
}
}

14
representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/tweet/BUILD Normal file
View File

@ -0,0 +1,14 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"finatra/inject/inject-core/src/main/scala",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/columns",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/modules",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/store",
"representation-manager/server/src/main/thrift:thrift-scala",
"strato/src/main/scala/com/twitter/strato/fed",
"strato/src/main/scala/com/twitter/strato/fed/server",
],
)

73
representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/tweet/TweetSimClustersEmbeddingCol.scala Normal file
View File

@ -0,0 +1,73 @@
package com.twitter.representation_manager.columns.tweet
import com.twitter.representation_manager.columns.ColumnConfigBase
import com.twitter.representation_manager.store.TweetSimClustersEmbeddingStore
import com.twitter.representation_manager.thriftscala.SimClustersEmbeddingView
import com.twitter.simclusters_v2.thriftscala.InternalId
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbedding
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.stitch
import com.twitter.stitch.Stitch
import com.twitter.stitch.storehaus.StitchOfReadableStore
import com.twitter.strato.catalog.OpMetadata
import com.twitter.strato.config.AnyOf
import com.twitter.strato.config.ContactInfo
import com.twitter.strato.config.FromColumns
import com.twitter.strato.config.Policy
import com.twitter.strato.config.Prefix
import com.twitter.strato.data.Conv
import com.twitter.strato.data.Description.PlainText
import com.twitter.strato.data.Lifecycle
import com.twitter.strato.fed._
import com.twitter.strato.thrift.ScroogeConv
import javax.inject.Inject
class TweetSimClustersEmbeddingCol @Inject() (embeddingStore: TweetSimClustersEmbeddingStore)
extends StratoFed.Column("recommendations/representation_manager/simClustersEmbedding.Tweet")
with StratoFed.Fetch.Stitch {
private val storeStitch: SimClustersEmbeddingId => Stitch[SimClustersEmbedding] =
StitchOfReadableStore(embeddingStore.tweetSimClustersEmbeddingStore.mapValues(_.toThrift))
val colPermissions: Seq[com.twitter.strato.config.Policy] =
ColumnConfigBase.recosPermissions ++ ColumnConfigBase.externalPermissions :+ FromColumns(
Set(
Prefix("ml/featureStore/simClusters"),
))
override val policy: Policy = AnyOf({
colPermissions
})
override type Key = Long // TweetId
override type View = SimClustersEmbeddingView
override type Value = SimClustersEmbedding
override val keyConv: Conv[Key] = Conv.long
override val viewConv: Conv[View] = ScroogeConv.fromStruct[SimClustersEmbeddingView]
override val valueConv: Conv[Value] = ScroogeConv.fromStruct[SimClustersEmbedding]
override val contactInfo: ContactInfo = ColumnConfigBase.contactInfo
override val metadata: OpMetadata = OpMetadata(
lifecycle = Some(Lifecycle.Production),
description = Some(
PlainText("The Tweet SimClusters Embedding Endpoint in Representation Management Service." +
" TDD: http://go/rms-tdd"))
)
override def fetch(key: Key, view: View): Stitch[Result[Value]] = {
val embeddingId = SimClustersEmbeddingId(
view.embeddingType,
view.modelVersion,
InternalId.TweetId(key)
)
storeStitch(embeddingId)
.map(embedding => found(embedding))
.handle {
case stitch.NotFound => missing
}
}
}

14
representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/user/BUILD Normal file
View File

@ -0,0 +1,14 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"finatra/inject/inject-core/src/main/scala",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/columns",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/modules",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/store",
"representation-manager/server/src/main/thrift:thrift-scala",
"strato/src/main/scala/com/twitter/strato/fed",
"strato/src/main/scala/com/twitter/strato/fed/server",
],
)

73
representation-manager/server/src/main/scala/com/twitter/representation_manager/columns/user/UserSimClustersEmbeddingCol.scala Normal file
View File

@ -0,0 +1,73 @@
package com.twitter.representation_manager.columns.user
import com.twitter.representation_manager.columns.ColumnConfigBase
import com.twitter.representation_manager.store.UserSimClustersEmbeddingStore
import com.twitter.representation_manager.thriftscala.SimClustersEmbeddingView
import com.twitter.simclusters_v2.thriftscala.InternalId
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbedding
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.stitch
import com.twitter.stitch.Stitch
import com.twitter.stitch.storehaus.StitchOfReadableStore
import com.twitter.strato.catalog.OpMetadata
import com.twitter.strato.config.AnyOf
import com.twitter.strato.config.ContactInfo
import com.twitter.strato.config.FromColumns
import com.twitter.strato.config.Policy
import com.twitter.strato.config.Prefix
import com.twitter.strato.data.Conv
import com.twitter.strato.data.Description.PlainText
import com.twitter.strato.data.Lifecycle
import com.twitter.strato.fed._
import com.twitter.strato.thrift.ScroogeConv
import javax.inject.Inject
class UserSimClustersEmbeddingCol @Inject() (embeddingStore: UserSimClustersEmbeddingStore)
extends StratoFed.Column("recommendations/representation_manager/simClustersEmbedding.User")
with StratoFed.Fetch.Stitch {
private val storeStitch: SimClustersEmbeddingId => Stitch[SimClustersEmbedding] =
StitchOfReadableStore(embeddingStore.userSimClustersEmbeddingStore.mapValues(_.toThrift))
val colPermissions: Seq[com.twitter.strato.config.Policy] =
ColumnConfigBase.recosPermissions ++ ColumnConfigBase.externalPermissions :+ FromColumns(
Set(
Prefix("ml/featureStore/simClusters"),
))
override val policy: Policy = AnyOf({
colPermissions
})
override type Key = Long // UserId
override type View = SimClustersEmbeddingView
override type Value = SimClustersEmbedding
override val keyConv: Conv[Key] = Conv.long
override val viewConv: Conv[View] = ScroogeConv.fromStruct[SimClustersEmbeddingView]
override val valueConv: Conv[Value] = ScroogeConv.fromStruct[SimClustersEmbedding]
override val contactInfo: ContactInfo = ColumnConfigBase.contactInfo
override val metadata: OpMetadata = OpMetadata(
lifecycle = Some(Lifecycle.Production),
description = Some(
PlainText("The User SimClusters Embedding Endpoint in Representation Management Service." +
" TDD: http://go/rms-tdd"))
)
override def fetch(key: Key, view: View): Stitch[Result[Value]] = {
val embeddingId = SimClustersEmbeddingId(
view.embeddingType,
view.modelVersion,
InternalId.UserId(key)
)
storeStitch(embeddingId)
.map(embedding => found(embedding))
.handle {
case stitch.NotFound => missing
}
}
}

13
representation-manager/server/src/main/scala/com/twitter/representation_manager/common/BUILD Normal file
View File

@ -0,0 +1,13 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"decider/src/main/scala",
"finagle/finagle-memcached",
"hermit/hermit-core/src/main/scala/com/twitter/hermit/store/common",
"relevance-platform/src/main/scala/com/twitter/relevance_platform/common/injection",
"src/scala/com/twitter/simclusters_v2/common",
"src/thrift/com/twitter/simclusters_v2:simclusters_v2-thrift-scala",
],
)

153
representation-manager/server/src/main/scala/com/twitter/representation_manager/common/MemCacheConfig.scala Normal file
View File

@ -0,0 +1,153 @@
package com.twitter.representation_manager.common
import com.twitter.bijection.scrooge.BinaryScalaCodec
import com.twitter.conversions.DurationOps._
import com.twitter.finagle.memcached.Client
import com.twitter.finagle.stats.StatsReceiver
import com.twitter.hashing.KeyHasher
import com.twitter.hermit.store.common.ObservedMemcachedReadableStore
import com.twitter.relevance_platform.common.injection.LZ4Injection
import com.twitter.simclusters_v2.common.SimClustersEmbedding
import com.twitter.simclusters_v2.common.SimClustersEmbeddingIdCacheKeyBuilder
import com.twitter.simclusters_v2.thriftscala.EmbeddingType
import com.twitter.simclusters_v2.thriftscala.EmbeddingType._
import com.twitter.simclusters_v2.thriftscala.ModelVersion
import com.twitter.simclusters_v2.thriftscala.ModelVersion._
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.simclusters_v2.thriftscala.{SimClustersEmbedding => ThriftSimClustersEmbedding}
import com.twitter.storehaus.ReadableStore
import com.twitter.util.Duration
/*
* NOTE - ALL the cache configs here are just placeholders, NONE of them is used anyweher in RMS yet
* */
sealed trait MemCacheParams
sealed trait MemCacheConfig
/*
* This holds params that is required to set up a memcache cache for a single embedding store
* */
case class EnabledMemCacheParams(ttl: Duration) extends MemCacheParams
object DisabledMemCacheParams extends MemCacheParams
/*
* We use this MemcacheConfig as the single source to set up the memcache for all RMS use cases
* NO OVERRIDE FROM CLIENT
* */
object MemCacheConfig {
val keyHasher: KeyHasher = KeyHasher.FNV1A_64
val hashKeyPrefix: String = "RMS"
val simclustersEmbeddingCacheKeyBuilder =
SimClustersEmbeddingIdCacheKeyBuilder(keyHasher.hashKey, hashKeyPrefix)
val cacheParamsMap: Map[
(EmbeddingType, ModelVersion),
MemCacheParams
] = Map(
// Tweet Embeddings
(LogFavBasedTweet, Model20m145kUpdated) -> EnabledMemCacheParams(ttl = 10.minutes),
(LogFavBasedTweet, Model20m145k2020) -> EnabledMemCacheParams(ttl = 10.minutes),
(LogFavLongestL2EmbeddingTweet, Model20m145kUpdated) -> EnabledMemCacheParams(ttl = 10.minutes),
(LogFavLongestL2EmbeddingTweet, Model20m145k2020) -> EnabledMemCacheParams(ttl = 10.minutes),
// User - KnownFor Embeddings
(FavBasedProducer, Model20m145kUpdated) -> EnabledMemCacheParams(ttl = 12.hours),
(FavBasedProducer, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(FollowBasedProducer, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(AggregatableLogFavBasedProducer, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(RelaxedAggregatableLogFavBasedProducer, Model20m145kUpdated) -> EnabledMemCacheParams(ttl =
12.hours),
(RelaxedAggregatableLogFavBasedProducer, Model20m145k2020) -> EnabledMemCacheParams(ttl =
12.hours),
// User - InterestedIn Embeddings
(LogFavBasedUserInterestedInFromAPE, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(FollowBasedUserInterestedInFromAPE, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(FavBasedUserInterestedIn, Model20m145kUpdated) -> EnabledMemCacheParams(ttl = 12.hours),
(FavBasedUserInterestedIn, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(FollowBasedUserInterestedIn, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(LogFavBasedUserInterestedIn, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(FavBasedUserInterestedInFromPE, Model20m145kUpdated) -> EnabledMemCacheParams(ttl = 12.hours),
(FilteredUserInterestedIn, Model20m145kUpdated) -> EnabledMemCacheParams(ttl = 12.hours),
(FilteredUserInterestedIn, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(FilteredUserInterestedInFromPE, Model20m145kUpdated) -> EnabledMemCacheParams(ttl = 12.hours),
(UnfilteredUserInterestedIn, Model20m145kUpdated) -> EnabledMemCacheParams(ttl = 12.hours),
(UnfilteredUserInterestedIn, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(UserNextInterestedIn, Model20m145k2020) -> EnabledMemCacheParams(ttl =
30.minutes), //embedding is updated every 2 hours, keeping it lower to avoid staleness
(
LogFavBasedUserInterestedMaxpoolingAddressBookFromIIAPE,
Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(
LogFavBasedUserInterestedAverageAddressBookFromIIAPE,
Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(
LogFavBasedUserInterestedBooktypeMaxpoolingAddressBookFromIIAPE,
Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(
LogFavBasedUserInterestedLargestDimMaxpoolingAddressBookFromIIAPE,
Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(
LogFavBasedUserInterestedLouvainMaxpoolingAddressBookFromIIAPE,
Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(
LogFavBasedUserInterestedConnectedMaxpoolingAddressBookFromIIAPE,
Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
// Topic Embeddings
(FavTfgTopic, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
(LogFavBasedKgoApeTopic, Model20m145k2020) -> EnabledMemCacheParams(ttl = 12.hours),
)
def getCacheSetup(
embeddingType: EmbeddingType,
modelVersion: ModelVersion
): MemCacheParams = {
// When requested (embeddingType, modelVersion) doesn't exist, we return DisabledMemCacheParams
cacheParamsMap.getOrElse((embeddingType, modelVersion), DisabledMemCacheParams)
}
def getCacheKeyPrefix(embeddingType: EmbeddingType, modelVersion: ModelVersion) =
s"${embeddingType.value}_${modelVersion.value}_"
def getStatsName(embeddingType: EmbeddingType, modelVersion: ModelVersion) =
s"${embeddingType.name}_${modelVersion.name}_mem_cache"
/**
* Build a ReadableStore based on MemCacheConfig.
*
* If memcache is disabled, it will return a normal readable store wrapper of the rawStore,
* with SimClustersEmbedding as value;
* If memcache is enabled, it will return a ObservedMemcachedReadableStore wrapper of the rawStore,
* with memcache set up according to the EnabledMemCacheParams
* */
def buildMemCacheStoreForSimClustersEmbedding(
rawStore: ReadableStore[SimClustersEmbeddingId, ThriftSimClustersEmbedding],
cacheClient: Client,
embeddingType: EmbeddingType,
modelVersion: ModelVersion,
stats: StatsReceiver
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val cacheParams = getCacheSetup(embeddingType, modelVersion)
val store = cacheParams match {
case DisabledMemCacheParams => rawStore
case EnabledMemCacheParams(ttl) =>
val memCacheKeyPrefix = MemCacheConfig.getCacheKeyPrefix(
embeddingType,
modelVersion
)
val statsName = MemCacheConfig.getStatsName(
embeddingType,
modelVersion
)
ObservedMemcachedReadableStore.fromCacheClient(
backingStore = rawStore,
cacheClient = cacheClient,
ttl = ttl
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = stats.scope(statsName),
keyToString = { k => memCacheKeyPrefix + k.toString }
)
}
store.mapValues(SimClustersEmbedding(_))
}
}

25
representation-manager/server/src/main/scala/com/twitter/representation_manager/common/RepresentationManagerDecider.scala Normal file
View File

@ -0,0 +1,25 @@
package com.twitter.representation_manager.common
import com.twitter.decider.Decider
import com.twitter.decider.RandomRecipient
import com.twitter.decider.Recipient
import com.twitter.simclusters_v2.common.DeciderGateBuilderWithIdHashing
import javax.inject.Inject
case class RepresentationManagerDecider @Inject() (decider: Decider) {
val deciderGateBuilder = new DeciderGateBuilderWithIdHashing(decider)
def isAvailable(feature: String, recipient: Option[Recipient]): Boolean = {
decider.isAvailable(feature, recipient)
}
/**
* When useRandomRecipient is set to false, the decider is either completely on or off.
* When useRandomRecipient is set to true, the decider is on for the specified % of traffic.
*/
def isAvailable(feature: String, useRandomRecipient: Boolean = true): Boolean = {
if (useRandomRecipient) isAvailable(feature, Some(RandomRecipient))
else isAvailable(feature, None)
}
}

25
representation-manager/server/src/main/scala/com/twitter/representation_manager/migration/BUILD Normal file
View File

@ -0,0 +1,25 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"content-recommender/server/src/main/scala/com/twitter/contentrecommender:representation-manager-deps",
"frigate/frigate-common/src/main/scala/com/twitter/frigate/common/store/strato",
"frigate/frigate-common/src/main/scala/com/twitter/frigate/common/util",
"hermit/hermit-core/src/main/scala/com/twitter/hermit/store/common",
"relevance-platform/src/main/scala/com/twitter/relevance_platform/common/injection",
"relevance-platform/src/main/scala/com/twitter/relevance_platform/common/readablestore",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/common",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/store",
"src/scala/com/twitter/ml/api/embedding",
"src/scala/com/twitter/simclusters_v2/common",
"src/scala/com/twitter/simclusters_v2/score",
"src/scala/com/twitter/simclusters_v2/summingbird/stores",
"src/scala/com/twitter/storehaus_internal/manhattan",
"src/scala/com/twitter/storehaus_internal/util",
"src/thrift/com/twitter/simclusters_v2:simclusters_v2-thrift-scala",
"src/thrift/com/twitter/socialgraph:thrift-scala",
"storage/clients/manhattan/client/src/main/scala",
"tweetypie/src/scala/com/twitter/tweetypie/util",
],
)

846
representation-manager/server/src/main/scala/com/twitter/representation_manager/migration/LegacyRMS.scala Normal file
View File

@ -0,0 +1,846 @@
package com.twitter.representation_manager.migration
import com.twitter.bijection.Injection
import com.twitter.bijection.scrooge.BinaryScalaCodec
import com.twitter.contentrecommender.store.ApeEntityEmbeddingStore
import com.twitter.contentrecommender.store.InterestsOptOutStore
import com.twitter.contentrecommender.store.SemanticCoreTopicSeedStore
import com.twitter.contentrecommender.twistly
import com.twitter.conversions.DurationOps._
import com.twitter.decider.Decider
import com.twitter.escherbird.util.uttclient.CacheConfigV2
import com.twitter.escherbird.util.uttclient.CachedUttClientV2
import com.twitter.escherbird.util.uttclient.UttClientCacheConfigsV2
import com.twitter.escherbird.utt.strato.thriftscala.Environment
import com.twitter.finagle.ThriftMux
import com.twitter.finagle.memcached.Client
import com.twitter.finagle.mtls.authentication.ServiceIdentifier
import com.twitter.finagle.mtls.client.MtlsStackClient.MtlsThriftMuxClientSyntax
import com.twitter.finagle.mux.ClientDiscardedRequestException
import com.twitter.finagle.service.ReqRep
import com.twitter.finagle.service.ResponseClass
import com.twitter.finagle.stats.StatsReceiver
import com.twitter.finagle.thrift.ClientId
import com.twitter.frigate.common.store.strato.StratoFetchableStore
import com.twitter.frigate.common.util.SeqLongInjection
import com.twitter.hashing.KeyHasher
import com.twitter.hermit.store.common.DeciderableReadableStore
import com.twitter.hermit.store.common.ObservedCachedReadableStore
import com.twitter.hermit.store.common.ObservedMemcachedReadableStore
import com.twitter.hermit.store.common.ObservedReadableStore
import com.twitter.interests.thriftscala.InterestsThriftService
import com.twitter.relevance_platform.common.injection.LZ4Injection
import com.twitter.relevance_platform.common.readablestore.ReadableStoreWithTimeout
import com.twitter.representation_manager.common.RepresentationManagerDecider
import com.twitter.representation_manager.store.DeciderConstants
import com.twitter.representation_manager.store.DeciderKey
import com.twitter.simclusters_v2.common.ModelVersions
import com.twitter.simclusters_v2.common.SimClustersEmbedding
import com.twitter.simclusters_v2.common.SimClustersEmbeddingIdCacheKeyBuilder
import com.twitter.simclusters_v2.stores.SimClustersEmbeddingStore
import com.twitter.simclusters_v2.summingbird.stores.PersistentTweetEmbeddingStore
import com.twitter.simclusters_v2.summingbird.stores.ProducerClusterEmbeddingReadableStores
import com.twitter.simclusters_v2.summingbird.stores.UserInterestedInReadableStore
import com.twitter.simclusters_v2.thriftscala.ClustersUserIsInterestedIn
import com.twitter.simclusters_v2.thriftscala.EmbeddingType
import com.twitter.simclusters_v2.thriftscala.EmbeddingType._
import com.twitter.simclusters_v2.thriftscala.InternalId
import com.twitter.simclusters_v2.thriftscala.ModelVersion
import com.twitter.simclusters_v2.thriftscala.ModelVersion.Model20m145k2020
import com.twitter.simclusters_v2.thriftscala.ModelVersion.Model20m145kUpdated
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.simclusters_v2.thriftscala.SimClustersMultiEmbedding
import com.twitter.simclusters_v2.thriftscala.SimClustersMultiEmbeddingId
import com.twitter.simclusters_v2.thriftscala.{SimClustersEmbedding => ThriftSimClustersEmbedding}
import com.twitter.storage.client.manhattan.kv.ManhattanKVClientMtlsParams
import com.twitter.storehaus.ReadableStore
import com.twitter.storehaus_internal.manhattan.Athena
import com.twitter.storehaus_internal.manhattan.ManhattanRO
import com.twitter.storehaus_internal.manhattan.ManhattanROConfig
import com.twitter.storehaus_internal.util.ApplicationID
import com.twitter.storehaus_internal.util.DatasetName
import com.twitter.storehaus_internal.util.HDFSPath
import com.twitter.strato.client.Strato
import com.twitter.strato.client.{Client => StratoClient}
import com.twitter.strato.thrift.ScroogeConvImplicits._
import com.twitter.tweetypie.util.UserId
import com.twitter.util.Duration
import com.twitter.util.Future
import com.twitter.util.Throw
import com.twitter.util.Timer
import javax.inject.Inject
import javax.inject.Named
import scala.reflect.ClassTag
class LegacyRMS @Inject() (
serviceIdentifier: ServiceIdentifier,
cacheClient: Client,
stats: StatsReceiver,
decider: Decider,
clientId: ClientId,
timer: Timer,
@Named("cacheHashKeyPrefix") val cacheHashKeyPrefix: String = "RMS",
@Named("useContentRecommenderConfiguration") val useContentRecommenderConfiguration: Boolean =
false) {
private val mhMtlsParams: ManhattanKVClientMtlsParams = ManhattanKVClientMtlsParams(
serviceIdentifier)
private val rmsDecider = RepresentationManagerDecider(decider)
val keyHasher: KeyHasher = KeyHasher.FNV1A_64
private val embeddingCacheKeyBuilder =
SimClustersEmbeddingIdCacheKeyBuilder(keyHasher.hashKey, cacheHashKeyPrefix)
private val statsReceiver = stats.scope("representation_management")
// Strato client, default timeout = 280ms
val stratoClient: StratoClient =
Strato.client
.withMutualTls(serviceIdentifier)
.build()
// Builds ThriftMux client builder for Content-Recommender service
private def makeThriftClientBuilder(
requestTimeout: Duration
): ThriftMux.Client = {
ThriftMux.client
.withClientId(clientId)
.withMutualTls(serviceIdentifier)
.withRequestTimeout(requestTimeout)
.withStatsReceiver(statsReceiver.scope("clnt"))
.withResponseClassifier {
case ReqRep(_, Throw(_: ClientDiscardedRequestException)) => ResponseClass.Ignorable
}
}
private def makeThriftClient[ThriftServiceType: ClassTag](
dest: String,
label: String,
requestTimeout: Duration = 450.milliseconds
): ThriftServiceType = {
makeThriftClientBuilder(requestTimeout)
.build[ThriftServiceType](dest, label)
}
/** *** SimCluster Embedding Stores ******/
implicit val simClustersEmbeddingIdInjection: Injection[SimClustersEmbeddingId, Array[Byte]] =
BinaryScalaCodec(SimClustersEmbeddingId)
implicit val simClustersEmbeddingInjection: Injection[ThriftSimClustersEmbedding, Array[Byte]] =
BinaryScalaCodec(ThriftSimClustersEmbedding)
implicit val simClustersMultiEmbeddingInjection: Injection[SimClustersMultiEmbedding, Array[
Byte
]] =
BinaryScalaCodec(SimClustersMultiEmbedding)
implicit val simClustersMultiEmbeddingIdInjection: Injection[SimClustersMultiEmbeddingId, Array[
Byte
]] =
BinaryScalaCodec(SimClustersMultiEmbeddingId)
def getEmbeddingsDataset(
mhMtlsParams: ManhattanKVClientMtlsParams,
datasetName: String
): ReadableStore[SimClustersEmbeddingId, ThriftSimClustersEmbedding] = {
ManhattanRO.getReadableStoreWithMtls[SimClustersEmbeddingId, ThriftSimClustersEmbedding](
ManhattanROConfig(
HDFSPath(""), // not needed
ApplicationID("content_recommender_athena"),
DatasetName(datasetName), // this should be correct
Athena
),
mhMtlsParams
)
}
lazy val logFavBasedLongestL2Tweet20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore =
PersistentTweetEmbeddingStore
.longestL2NormTweetEmbeddingStoreManhattan(
mhMtlsParams,
PersistentTweetEmbeddingStore.LogFavBased20m145k2020Dataset,
statsReceiver,
maxLength = 10,
).mapValues(_.toThrift)
val memcachedStore = ObservedMemcachedReadableStore.fromCacheClient(
backingStore = rawStore,
cacheClient = cacheClient,
ttl = 15.minutes
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver =
statsReceiver.scope("log_fav_based_longest_l2_tweet_embedding_20m145k2020_mem_cache"),
keyToString = { k =>
s"scez_l2:${LogFavBasedTweet}_${ModelVersions.Model20M145K2020}_$k"
}
)
val inMemoryCacheStore: ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] =
memcachedStore
.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(
LogFavLongestL2EmbeddingTweet,
Model20m145k2020,
InternalId.TweetId(tweetId)) =>
tweetId
}
.mapValues(SimClustersEmbedding(_))
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
inMemoryCacheStore,
ttl = 12.minute,
maxKeys = 1048575,
cacheName = "log_fav_based_longest_l2_tweet_embedding_20m145k2020_cache",
windowSize = 10000L
)(statsReceiver.scope("log_fav_based_longest_l2_tweet_embedding_20m145k2020_store"))
}
lazy val logFavBased20M145KUpdatedTweetEmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore =
PersistentTweetEmbeddingStore
.mostRecentTweetEmbeddingStoreManhattan(
mhMtlsParams,
PersistentTweetEmbeddingStore.LogFavBased20m145kUpdatedDataset,
statsReceiver
).mapValues(_.toThrift)
val memcachedStore = ObservedMemcachedReadableStore.fromCacheClient(
backingStore = rawStore,
cacheClient = cacheClient,
ttl = 10.minutes
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = statsReceiver.scope("log_fav_based_tweet_embedding_mem_cache"),
keyToString = { k =>
// SimClusters_embedding_LZ4/embeddingType_modelVersion_tweetId
s"scez:${LogFavBasedTweet}_${ModelVersions.Model20M145KUpdated}_$k"
}
)
val inMemoryCacheStore: ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
memcachedStore
.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(
LogFavBasedTweet,
Model20m145kUpdated,
InternalId.TweetId(tweetId)) =>
tweetId
}
.mapValues(SimClustersEmbedding(_))
}
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
inMemoryCacheStore,
ttl = 5.minute,
maxKeys = 1048575, // 200MB
cacheName = "log_fav_based_tweet_embedding_cache",
windowSize = 10000L
)(statsReceiver.scope("log_fav_based_tweet_embedding_store"))
}
lazy val logFavBased20M145K2020TweetEmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore =
PersistentTweetEmbeddingStore
.mostRecentTweetEmbeddingStoreManhattan(
mhMtlsParams,
PersistentTweetEmbeddingStore.LogFavBased20m145k2020Dataset,
statsReceiver,
maxLength = 10,
).mapValues(_.toThrift)
val memcachedStore = ObservedMemcachedReadableStore.fromCacheClient(
backingStore = rawStore,
cacheClient = cacheClient,
ttl = 15.minutes
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = statsReceiver.scope("log_fav_based_tweet_embedding_20m145k2020_mem_cache"),
keyToString = { k =>
// SimClusters_embedding_LZ4/embeddingType_modelVersion_tweetId
s"scez:${LogFavBasedTweet}_${ModelVersions.Model20M145K2020}_$k"
}
)
val inMemoryCacheStore: ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] =
memcachedStore
.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(
LogFavBasedTweet,
Model20m145k2020,
InternalId.TweetId(tweetId)) =>
tweetId
}
.mapValues(SimClustersEmbedding(_))
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
inMemoryCacheStore,
ttl = 12.minute,
maxKeys = 16777215,
cacheName = "log_fav_based_tweet_embedding_20m145k2020_cache",
windowSize = 10000L
)(statsReceiver.scope("log_fav_based_tweet_embedding_20m145k2020_store"))
}
lazy val favBasedTfgTopicEmbedding2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val stratoStore =
StratoFetchableStore
.withUnitView[SimClustersEmbeddingId, ThriftSimClustersEmbedding](
stratoClient,
"recommendations/simclusters_v2/embeddings/favBasedTFGTopic20M145K2020")
val truncatedStore = stratoStore.mapValues { embedding =>
SimClustersEmbedding(embedding, truncate = 50)
}
ObservedCachedReadableStore.from(
ObservedReadableStore(truncatedStore)(
statsReceiver.scope("fav_tfg_topic_embedding_2020_cache_backing_store")),
ttl = 12.hours,
maxKeys = 262143, // 200MB
cacheName = "fav_tfg_topic_embedding_2020_cache",
windowSize = 10000L
)(statsReceiver.scope("fav_tfg_topic_embedding_2020_cache"))
}
lazy val logFavBasedApe20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
ObservedReadableStore(
StratoFetchableStore
.withUnitView[SimClustersEmbeddingId, ThriftSimClustersEmbedding](
stratoClient,
"recommendations/simclusters_v2/embeddings/logFavBasedAPE20M145K2020")
.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(
AggregatableLogFavBasedProducer,
Model20m145k2020,
internalId) =>
SimClustersEmbeddingId(AggregatableLogFavBasedProducer, Model20m145k2020, internalId)
}
.mapValues(embedding => SimClustersEmbedding(embedding, 50))
)(statsReceiver.scope("aggregatable_producer_embeddings_by_logfav_score_2020"))
}
val interestService: InterestsThriftService.MethodPerEndpoint =
makeThriftClient[InterestsThriftService.MethodPerEndpoint](
"/s/interests-thrift-service/interests-thrift-service",
"interests_thrift_service"
)
val interestsOptOutStore: InterestsOptOutStore = InterestsOptOutStore(interestService)
// Save 2 ^ 18 UTTs. Promising 100% cache rate
lazy val defaultCacheConfigV2: CacheConfigV2 = CacheConfigV2(262143)
lazy val uttClientCacheConfigsV2: UttClientCacheConfigsV2 = UttClientCacheConfigsV2(
getTaxonomyConfig = defaultCacheConfigV2,
getUttTaxonomyConfig = defaultCacheConfigV2,
getLeafIds = defaultCacheConfigV2,
getLeafUttEntities = defaultCacheConfigV2
)
// CachedUttClient to use StratoClient
lazy val cachedUttClientV2: CachedUttClientV2 = new CachedUttClientV2(
stratoClient = stratoClient,
env = Environment.Prod,
cacheConfigs = uttClientCacheConfigsV2,
statsReceiver = statsReceiver.scope("cached_utt_client")
)
lazy val semanticCoreTopicSeedStore: ReadableStore[
SemanticCoreTopicSeedStore.Key,
Seq[UserId]
] = {
/*
Up to 1000 Long seeds per topic/language = 62.5kb per topic/language (worst case)
Assume ~10k active topic/languages ~= 650MB (worst case)
*/
val underlying = new SemanticCoreTopicSeedStore(cachedUttClientV2, interestsOptOutStore)(
statsReceiver.scope("semantic_core_topic_seed_store"))
val memcacheStore = ObservedMemcachedReadableStore.fromCacheClient(
backingStore = underlying,
cacheClient = cacheClient,
ttl = 12.hours
)(
valueInjection = SeqLongInjection,
statsReceiver = statsReceiver.scope("topic_producer_seed_store_mem_cache"),
keyToString = { k => s"tpss:${k.entityId}_${k.languageCode}" }
)
ObservedCachedReadableStore.from[SemanticCoreTopicSeedStore.Key, Seq[UserId]](
store = memcacheStore,
ttl = 6.hours,
maxKeys = 20e3.toInt,
cacheName = "topic_producer_seed_store_cache",
windowSize = 5000
)(statsReceiver.scope("topic_producer_seed_store_cache"))
}
lazy val logFavBasedApeEntity20M145K2020EmbeddingStore: ApeEntityEmbeddingStore = {
val apeStore = logFavBasedApe20M145K2020EmbeddingStore.composeKeyMapping[UserId]({ id =>
SimClustersEmbeddingId(
AggregatableLogFavBasedProducer,
Model20m145k2020,
InternalId.UserId(id))
})
new ApeEntityEmbeddingStore(
semanticCoreSeedStore = semanticCoreTopicSeedStore,
aggregatableProducerEmbeddingStore = apeStore,
statsReceiver = statsReceiver.scope("log_fav_based_ape_entity_2020_embedding_store"))
}
lazy val logFavBasedApeEntity20M145K2020EmbeddingCachedStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val truncatedStore =
logFavBasedApeEntity20M145K2020EmbeddingStore.mapValues(_.truncate(50).toThrift)
val memcachedStore = ObservedMemcachedReadableStore
.fromCacheClient(
backingStore = truncatedStore,
cacheClient = cacheClient,
ttl = 12.hours
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = statsReceiver.scope("log_fav_based_ape_entity_2020_embedding_mem_cache"),
keyToString = { k => embeddingCacheKeyBuilder.apply(k) }
).mapValues(SimClustersEmbedding(_))
val inMemoryCachedStore =
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
memcachedStore,
ttl = 6.hours,
maxKeys = 262143,
cacheName = "log_fav_based_ape_entity_2020_embedding_cache",
windowSize = 10000L
)(statsReceiver.scope("log_fav_based_ape_entity_2020_embedding_cached_store"))
DeciderableReadableStore(
inMemoryCachedStore,
rmsDecider.deciderGateBuilder.idGateWithHashing[SimClustersEmbeddingId](
DeciderKey.enableLogFavBasedApeEntity20M145K2020EmbeddingCachedStore),
statsReceiver.scope("log_fav_based_ape_entity_2020_embedding_deciderable_store")
)
}
lazy val relaxedLogFavBasedApe20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
ObservedReadableStore(
StratoFetchableStore
.withUnitView[SimClustersEmbeddingId, ThriftSimClustersEmbedding](
stratoClient,
"recommendations/simclusters_v2/embeddings/logFavBasedAPERelaxedFavEngagementThreshold20M145K2020")
.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(
RelaxedAggregatableLogFavBasedProducer,
Model20m145k2020,
internalId) =>
SimClustersEmbeddingId(
RelaxedAggregatableLogFavBasedProducer,
Model20m145k2020,
internalId)
}
.mapValues(embedding => SimClustersEmbedding(embedding).truncate(50))
)(statsReceiver.scope(
"aggregatable_producer_embeddings_by_logfav_score_relaxed_fav_engagement_threshold_2020"))
}
lazy val relaxedLogFavBasedApe20M145K2020EmbeddingCachedStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val truncatedStore =
relaxedLogFavBasedApe20M145K2020EmbeddingStore.mapValues(_.truncate(50).toThrift)
val memcachedStore = ObservedMemcachedReadableStore
.fromCacheClient(
backingStore = truncatedStore,
cacheClient = cacheClient,
ttl = 12.hours
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver =
statsReceiver.scope("relaxed_log_fav_based_ape_entity_2020_embedding_mem_cache"),
keyToString = { k: SimClustersEmbeddingId => embeddingCacheKeyBuilder.apply(k) }
).mapValues(SimClustersEmbedding(_))
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
memcachedStore,
ttl = 6.hours,
maxKeys = 262143,
cacheName = "relaxed_log_fav_based_ape_entity_2020_embedding_cache",
windowSize = 10000L
)(statsReceiver.scope("relaxed_log_fav_based_ape_entity_2020_embedding_cache_store"))
}
lazy val favBasedProducer20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val underlyingStore = ProducerClusterEmbeddingReadableStores
.getProducerTopKSimClusters2020EmbeddingsStore(
mhMtlsParams
).composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(
FavBasedProducer,
Model20m145k2020,
InternalId.UserId(userId)) =>
userId
}.mapValues { topSimClustersWithScore =>
ThriftSimClustersEmbedding(topSimClustersWithScore.topClusters.take(10))
}
// same memcache config as for favBasedUserInterestedIn20M145K2020Store
val memcachedStore = ObservedMemcachedReadableStore
.fromCacheClient(
backingStore = underlyingStore,
cacheClient = cacheClient,
ttl = 24.hours
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = statsReceiver.scope("fav_based_producer_embedding_20M_145K_2020_mem_cache"),
keyToString = { k => embeddingCacheKeyBuilder.apply(k) }
).mapValues(SimClustersEmbedding(_))
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
memcachedStore,
ttl = 12.hours,
maxKeys = 16777215,
cacheName = "fav_based_producer_embedding_20M_145K_2020_embedding_cache",
windowSize = 10000L
)(statsReceiver.scope("fav_based_producer_embedding_20M_145K_2020_embedding_store"))
}
// Production
lazy val interestedIn20M145KUpdatedStore: ReadableStore[UserId, ClustersUserIsInterestedIn] = {
UserInterestedInReadableStore.defaultStoreWithMtls(
mhMtlsParams,
modelVersion = ModelVersions.Model20M145KUpdated
)
}
// Production
lazy val interestedIn20M145K2020Store: ReadableStore[UserId, ClustersUserIsInterestedIn] = {
UserInterestedInReadableStore.defaultStoreWithMtls(
mhMtlsParams,
modelVersion = ModelVersions.Model20M145K2020
)
}
// Production
lazy val InterestedInFromPE20M145KUpdatedStore: ReadableStore[
UserId,
ClustersUserIsInterestedIn
] = {
UserInterestedInReadableStore.defaultIIPEStoreWithMtls(
mhMtlsParams,
modelVersion = ModelVersions.Model20M145KUpdated)
}
lazy val simClustersInterestedInStore: ReadableStore[
(UserId, ModelVersion),
ClustersUserIsInterestedIn
] = {
new ReadableStore[(UserId, ModelVersion), ClustersUserIsInterestedIn] {
override def get(k: (UserId, ModelVersion)): Future[Option[ClustersUserIsInterestedIn]] = {
k match {
case (userId, Model20m145kUpdated) =>
interestedIn20M145KUpdatedStore.get(userId)
case (userId, Model20m145k2020) =>
interestedIn20M145K2020Store.get(userId)
case _ =>
Future.None
}
}
}
}
lazy val simClustersInterestedInFromProducerEmbeddingsStore: ReadableStore[
(UserId, ModelVersion),
ClustersUserIsInterestedIn
] = {
new ReadableStore[(UserId, ModelVersion), ClustersUserIsInterestedIn] {
override def get(k: (UserId, ModelVersion)): Future[Option[ClustersUserIsInterestedIn]] = {
k match {
case (userId, ModelVersion.Model20m145kUpdated) =>
InterestedInFromPE20M145KUpdatedStore.get(userId)
case _ =>
Future.None
}
}
}
}
lazy val userInterestedInStore =
new twistly.interestedin.EmbeddingStore(
interestedInStore = simClustersInterestedInStore,
interestedInFromProducerEmbeddingStore = simClustersInterestedInFromProducerEmbeddingsStore,
statsReceiver = statsReceiver
)
// Production
lazy val favBasedUserInterestedIn20M145KUpdatedStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val underlyingStore =
UserInterestedInReadableStore
.defaultSimClustersEmbeddingStoreWithMtls(
mhMtlsParams,
EmbeddingType.FavBasedUserInterestedIn,
ModelVersion.Model20m145kUpdated)
.mapValues(_.toThrift)
val memcachedStore = ObservedMemcachedReadableStore
.fromCacheClient(
backingStore = underlyingStore,
cacheClient = cacheClient,
ttl = 12.hours
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = statsReceiver.scope("fav_based_user_interested_in_mem_cache"),
keyToString = { k => embeddingCacheKeyBuilder.apply(k) }
).mapValues(SimClustersEmbedding(_))
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
memcachedStore,
ttl = 6.hours,
maxKeys = 262143,
cacheName = "fav_based_user_interested_in_cache",
windowSize = 10000L
)(statsReceiver.scope("fav_based_user_interested_in_store"))
}
// Production
lazy val LogFavBasedInterestedInFromAPE20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val underlyingStore =
UserInterestedInReadableStore
.defaultIIAPESimClustersEmbeddingStoreWithMtls(
mhMtlsParams,
EmbeddingType.LogFavBasedUserInterestedInFromAPE,
ModelVersion.Model20m145k2020)
.mapValues(_.toThrift)
val memcachedStore = ObservedMemcachedReadableStore
.fromCacheClient(
backingStore = underlyingStore,
cacheClient = cacheClient,
ttl = 12.hours
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = statsReceiver.scope("log_fav_based_user_interested_in_from_ape_mem_cache"),
keyToString = { k => embeddingCacheKeyBuilder.apply(k) }
).mapValues(SimClustersEmbedding(_))
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
memcachedStore,
ttl = 6.hours,
maxKeys = 262143,
cacheName = "log_fav_based_user_interested_in_from_ape_cache",
windowSize = 10000L
)(statsReceiver.scope("log_fav_based_user_interested_in_from_ape_store"))
}
// Production
lazy val FollowBasedInterestedInFromAPE20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val underlyingStore =
UserInterestedInReadableStore
.defaultIIAPESimClustersEmbeddingStoreWithMtls(
mhMtlsParams,
EmbeddingType.FollowBasedUserInterestedInFromAPE,
ModelVersion.Model20m145k2020)
.mapValues(_.toThrift)
val memcachedStore = ObservedMemcachedReadableStore
.fromCacheClient(
backingStore = underlyingStore,
cacheClient = cacheClient,
ttl = 12.hours
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = statsReceiver.scope("follow_based_user_interested_in_from_ape_mem_cache"),
keyToString = { k => embeddingCacheKeyBuilder.apply(k) }
).mapValues(SimClustersEmbedding(_))
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
memcachedStore,
ttl = 6.hours,
maxKeys = 262143,
cacheName = "follow_based_user_interested_in_from_ape_cache",
windowSize = 10000L
)(statsReceiver.scope("follow_based_user_interested_in_from_ape_store"))
}
// production
lazy val favBasedUserInterestedIn20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val underlyingStore: ReadableStore[SimClustersEmbeddingId, ThriftSimClustersEmbedding] =
UserInterestedInReadableStore
.defaultSimClustersEmbeddingStoreWithMtls(
mhMtlsParams,
EmbeddingType.FavBasedUserInterestedIn,
ModelVersion.Model20m145k2020).mapValues(_.toThrift)
ObservedMemcachedReadableStore
.fromCacheClient(
backingStore = underlyingStore,
cacheClient = cacheClient,
ttl = 12.hours
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = statsReceiver.scope("fav_based_user_interested_in_2020_mem_cache"),
keyToString = { k => embeddingCacheKeyBuilder.apply(k) }
).mapValues(SimClustersEmbedding(_))
}
// Production
lazy val logFavBasedUserInterestedIn20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val underlyingStore =
UserInterestedInReadableStore
.defaultSimClustersEmbeddingStoreWithMtls(
mhMtlsParams,
EmbeddingType.LogFavBasedUserInterestedIn,
ModelVersion.Model20m145k2020)
val memcachedStore = ObservedMemcachedReadableStore
.fromCacheClient(
backingStore = underlyingStore.mapValues(_.toThrift),
cacheClient = cacheClient,
ttl = 12.hours
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = statsReceiver.scope("log_fav_based_user_interested_in_2020_store"),
keyToString = { k => embeddingCacheKeyBuilder.apply(k) }
).mapValues(SimClustersEmbedding(_))
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
memcachedStore,
ttl = 6.hours,
maxKeys = 262143,
cacheName = "log_fav_based_user_interested_in_2020_cache",
windowSize = 10000L
)(statsReceiver.scope("log_fav_based_user_interested_in_2020_store"))
}
// Production
lazy val favBasedUserInterestedInFromPE20M145KUpdatedStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val underlyingStore =
UserInterestedInReadableStore
.defaultIIPESimClustersEmbeddingStoreWithMtls(
mhMtlsParams,
EmbeddingType.FavBasedUserInterestedInFromPE,
ModelVersion.Model20m145kUpdated)
.mapValues(_.toThrift)
val memcachedStore = ObservedMemcachedReadableStore
.fromCacheClient(
backingStore = underlyingStore,
cacheClient = cacheClient,
ttl = 12.hours
)(
valueInjection = LZ4Injection.compose(BinaryScalaCodec(ThriftSimClustersEmbedding)),
statsReceiver = statsReceiver.scope("fav_based_user_interested_in_from_pe_mem_cache"),
keyToString = { k => embeddingCacheKeyBuilder.apply(k) }
).mapValues(SimClustersEmbedding(_))
ObservedCachedReadableStore.from[SimClustersEmbeddingId, SimClustersEmbedding](
memcachedStore,
ttl = 6.hours,
maxKeys = 262143,
cacheName = "fav_based_user_interested_in_from_pe_cache",
windowSize = 10000L
)(statsReceiver.scope("fav_based_user_interested_in_from_pe_cache"))
}
private val underlyingStores: Map[
(EmbeddingType, ModelVersion),
ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding]
] = Map(
// Tweet Embeddings
(LogFavBasedTweet, Model20m145kUpdated) -> logFavBased20M145KUpdatedTweetEmbeddingStore,
(LogFavBasedTweet, Model20m145k2020) -> logFavBased20M145K2020TweetEmbeddingStore,
(
LogFavLongestL2EmbeddingTweet,
Model20m145k2020) -> logFavBasedLongestL2Tweet20M145K2020EmbeddingStore,
// Entity Embeddings
(FavTfgTopic, Model20m145k2020) -> favBasedTfgTopicEmbedding2020Store,
(
LogFavBasedKgoApeTopic,
Model20m145k2020) -> logFavBasedApeEntity20M145K2020EmbeddingCachedStore,
// KnownFor Embeddings
(FavBasedProducer, Model20m145k2020) -> favBasedProducer20M145K2020EmbeddingStore,
(
RelaxedAggregatableLogFavBasedProducer,
Model20m145k2020) -> relaxedLogFavBasedApe20M145K2020EmbeddingCachedStore,
// InterestedIn Embeddings
(
LogFavBasedUserInterestedInFromAPE,
Model20m145k2020) -> LogFavBasedInterestedInFromAPE20M145K2020Store,
(
FollowBasedUserInterestedInFromAPE,
Model20m145k2020) -> FollowBasedInterestedInFromAPE20M145K2020Store,
(FavBasedUserInterestedIn, Model20m145kUpdated) -> favBasedUserInterestedIn20M145KUpdatedStore,
(FavBasedUserInterestedIn, Model20m145k2020) -> favBasedUserInterestedIn20M145K2020Store,
(LogFavBasedUserInterestedIn, Model20m145k2020) -> logFavBasedUserInterestedIn20M145K2020Store,
(
FavBasedUserInterestedInFromPE,
Model20m145kUpdated) -> favBasedUserInterestedInFromPE20M145KUpdatedStore,
(FilteredUserInterestedIn, Model20m145kUpdated) -> userInterestedInStore,
(FilteredUserInterestedIn, Model20m145k2020) -> userInterestedInStore,
(FilteredUserInterestedInFromPE, Model20m145kUpdated) -> userInterestedInStore,
(UnfilteredUserInterestedIn, Model20m145kUpdated) -> userInterestedInStore,
(UnfilteredUserInterestedIn, Model20m145k2020) -> userInterestedInStore,
)
val simClustersEmbeddingStore: ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val underlying: ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] =
SimClustersEmbeddingStore.buildWithDecider(
underlyingStores = underlyingStores,
decider = rmsDecider.decider,
statsReceiver = statsReceiver.scope("simClusters_embeddings_store_deciderable")
)
val underlyingWithTimeout: ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] =
new ReadableStoreWithTimeout(
rs = underlying,
decider = rmsDecider.decider,
enableTimeoutDeciderKey = DeciderConstants.enableSimClustersEmbeddingStoreTimeouts,
timeoutValueKey = DeciderConstants.simClustersEmbeddingStoreTimeoutValueMillis,
timer = timer,
statsReceiver = statsReceiver.scope("simClusters_embedding_store_timeouts")
)
ObservedReadableStore(
store = underlyingWithTimeout
)(statsReceiver.scope("simClusters_embeddings_store"))
}
}

18
representation-manager/server/src/main/scala/com/twitter/representation_manager/modules/BUILD Normal file
View File

@ -0,0 +1,18 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"finagle-internal/mtls/src/main/scala/com/twitter/finagle/mtls/authentication",
"finagle/finagle-stats",
"finatra/inject/inject-core/src/main/scala",
"frigate/frigate-common/src/main/scala/com/twitter/frigate/common/util",
"interests-service/thrift/src/main/thrift:thrift-scala",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/common",
"servo/util",
"src/scala/com/twitter/storehaus_internal/manhattan",
"src/scala/com/twitter/storehaus_internal/memcache",
"src/scala/com/twitter/storehaus_internal/util",
"strato/src/main/scala/com/twitter/strato/client",
],
)

34
representation-manager/server/src/main/scala/com/twitter/representation_manager/modules/CacheModule.scala Normal file
View File

@ -0,0 +1,34 @@
package com.twitter.representation_manager.modules
import com.google.inject.Provides
import com.twitter.finagle.memcached.Client
import javax.inject.Singleton
import com.twitter.conversions.DurationOps._
import com.twitter.inject.TwitterModule
import com.twitter.finagle.mtls.authentication.ServiceIdentifier
import com.twitter.finagle.stats.StatsReceiver
import com.twitter.storehaus_internal.memcache.MemcacheStore
import com.twitter.storehaus_internal.util.ClientName
import com.twitter.storehaus_internal.util.ZkEndPoint
object CacheModule extends TwitterModule {
private val cacheDest = flag[String]("cache_module.dest", "Path to memcache service")
private val timeout = flag[Int]("memcache.timeout", "Memcache client timeout")
private val retries = flag[Int]("memcache.retries", "Memcache timeout retries")
@Singleton
@Provides
def providesCache(
serviceIdentifier: ServiceIdentifier,
stats: StatsReceiver
): Client =
MemcacheStore.memcachedClient(
name = ClientName("memcache_representation_manager"),
dest = ZkEndPoint(cacheDest()),
timeout = timeout().milliseconds,
retries = retries(),
statsReceiver = stats.scope("cache_client"),
serviceIdentifier = serviceIdentifier
)
}

40
representation-manager/server/src/main/scala/com/twitter/representation_manager/modules/InterestsThriftClientModule.scala Normal file
View File

@ -0,0 +1,40 @@
package com.twitter.representation_manager.modules
import com.google.inject.Provides
import com.twitter.conversions.DurationOps._
import com.twitter.finagle.ThriftMux
import com.twitter.finagle.mtls.authentication.ServiceIdentifier
import com.twitter.finagle.mtls.client.MtlsStackClient.MtlsThriftMuxClientSyntax
import com.twitter.finagle.mux.ClientDiscardedRequestException
import com.twitter.finagle.service.ReqRep
import com.twitter.finagle.service.ResponseClass
import com.twitter.finagle.stats.StatsReceiver
import com.twitter.finagle.thrift.ClientId
import com.twitter.inject.TwitterModule
import com.twitter.interests.thriftscala.InterestsThriftService
import com.twitter.util.Throw
import javax.inject.Singleton
object InterestsThriftClientModule extends TwitterModule {
@Singleton
@Provides
def providesInterestsThriftClient(
clientId: ClientId,
serviceIdentifier: ServiceIdentifier,
statsReceiver: StatsReceiver
): InterestsThriftService.MethodPerEndpoint = {
ThriftMux.client
.withClientId(clientId)
.withMutualTls(serviceIdentifier)
.withRequestTimeout(450.milliseconds)
.withStatsReceiver(statsReceiver.scope("InterestsThriftClient"))
.withResponseClassifier {
case ReqRep(_, Throw(_: ClientDiscardedRequestException)) => ResponseClass.Ignorable
}
.build[InterestsThriftService.MethodPerEndpoint](
dest = "/s/interests-thrift-service/interests-thrift-service",
label = "interests_thrift_service"
)
}
}

18
representation-manager/server/src/main/scala/com/twitter/representation_manager/modules/LegacyRMSConfigModule.scala Normal file
View File

@ -0,0 +1,18 @@
package com.twitter.representation_manager.modules
import com.google.inject.Provides
import com.twitter.inject.TwitterModule
import javax.inject.Named
import javax.inject.Singleton
object LegacyRMSConfigModule extends TwitterModule {
@Singleton
@Provides
@Named("cacheHashKeyPrefix")
def providesCacheHashKeyPrefix: String = "RMS"
@Singleton
@Provides
@Named("useContentRecommenderConfiguration")
def providesUseContentRecommenderConfiguration: Boolean = false
}

24
representation-manager/server/src/main/scala/com/twitter/representation_manager/modules/StoreModule.scala Normal file
View File

@ -0,0 +1,24 @@
package com.twitter.representation_manager.modules
import com.google.inject.Provides
import javax.inject.Singleton
import com.twitter.inject.TwitterModule
import com.twitter.decider.Decider
import com.twitter.finagle.mtls.authentication.ServiceIdentifier
import com.twitter.representation_manager.common.RepresentationManagerDecider
import com.twitter.storage.client.manhattan.kv.ManhattanKVClientMtlsParams
object StoreModule extends TwitterModule {
@Singleton
@Provides
def providesMhMtlsParams(
serviceIdentifier: ServiceIdentifier
): ManhattanKVClientMtlsParams = ManhattanKVClientMtlsParams(serviceIdentifier)
@Singleton
@Provides
def providesRmsDecider(
decider: Decider
): RepresentationManagerDecider = RepresentationManagerDecider(decider)
}

13
representation-manager/server/src/main/scala/com/twitter/representation_manager/modules/TimerModule.scala Normal file
View File

@ -0,0 +1,13 @@
package com.twitter.representation_manager.modules
import com.google.inject.Provides
import com.twitter.finagle.util.DefaultTimer
import com.twitter.inject.TwitterModule
import com.twitter.util.Timer
import javax.inject.Singleton
object TimerModule extends TwitterModule {
@Singleton
@Provides
def providesTimer: Timer = DefaultTimer
}

39
representation-manager/server/src/main/scala/com/twitter/representation_manager/modules/UttClientModule.scala Normal file
View File

@ -0,0 +1,39 @@
package com.twitter.representation_manager.modules
import com.google.inject.Provides
import com.twitter.escherbird.util.uttclient.CacheConfigV2
import com.twitter.escherbird.util.uttclient.CachedUttClientV2
import com.twitter.escherbird.util.uttclient.UttClientCacheConfigsV2
import com.twitter.escherbird.utt.strato.thriftscala.Environment
import com.twitter.finagle.stats.StatsReceiver
import com.twitter.inject.TwitterModule
import com.twitter.strato.client.{Client => StratoClient}
import javax.inject.Singleton
object UttClientModule extends TwitterModule {
@Singleton
@Provides
def providesUttClient(
stratoClient: StratoClient,
statsReceiver: StatsReceiver
): CachedUttClientV2 = {
// Save 2 ^ 18 UTTs. Promising 100% cache rate
val defaultCacheConfigV2: CacheConfigV2 = CacheConfigV2(262143)
val uttClientCacheConfigsV2: UttClientCacheConfigsV2 = UttClientCacheConfigsV2(
getTaxonomyConfig = defaultCacheConfigV2,
getUttTaxonomyConfig = defaultCacheConfigV2,
getLeafIds = defaultCacheConfigV2,
getLeafUttEntities = defaultCacheConfigV2
)
// CachedUttClient to use StratoClient
new CachedUttClientV2(
stratoClient = stratoClient,
env = Environment.Prod,
cacheConfigs = uttClientCacheConfigsV2,
statsReceiver = statsReceiver.scope("cached_utt_client")
)
}
}

16
representation-manager/server/src/main/scala/com/twitter/representation_manager/store/BUILD Normal file
View File

@ -0,0 +1,16 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"content-recommender/server/src/main/scala/com/twitter/contentrecommender:representation-manager-deps",
"frigate/frigate-common/src/main/scala/com/twitter/frigate/common/util",
"hermit/hermit-core/src/main/scala/com/twitter/hermit/store/common",
"representation-manager/server/src/main/scala/com/twitter/representation_manager/common",
"src/scala/com/twitter/simclusters_v2/stores",
"src/scala/com/twitter/simclusters_v2/summingbird/stores",
"src/thrift/com/twitter/simclusters_v2:simclusters_v2-thrift-scala",
"storage/clients/manhattan/client/src/main/scala",
"tweetypie/src/scala/com/twitter/tweetypie/util",
],
)

39
representation-manager/server/src/main/scala/com/twitter/representation_manager/store/DeciderConstants.scala Normal file
View File

@ -0,0 +1,39 @@
package com.twitter.representation_manager.store
import com.twitter.servo.decider.DeciderKeyEnum
object DeciderConstants {
// Deciders inherited from CR and RSX and only used in LegacyRMS
// Their value are manipulated by CR and RSX's yml file and their decider dashboard
// We will remove them after migration completed
val enableLogFavBasedApeEntity20M145KUpdatedEmbeddingCachedStore =
"enableLogFavBasedApeEntity20M145KUpdatedEmbeddingCachedStore"
val enableLogFavBasedApeEntity20M145K2020EmbeddingCachedStore =
"enableLogFavBasedApeEntity20M145K2020EmbeddingCachedStore"
val enablelogFavBased20M145K2020TweetEmbeddingStoreTimeouts =
"enable_log_fav_based_tweet_embedding_20m145k2020_timeouts"
val logFavBased20M145K2020TweetEmbeddingStoreTimeoutValueMillis =
"log_fav_based_tweet_embedding_20m145k2020_timeout_value_millis"
val enablelogFavBased20M145KUpdatedTweetEmbeddingStoreTimeouts =
"enable_log_fav_based_tweet_embedding_20m145kUpdated_timeouts"
val logFavBased20M145KUpdatedTweetEmbeddingStoreTimeoutValueMillis =
"log_fav_based_tweet_embedding_20m145kUpdated_timeout_value_millis"
val enableSimClustersEmbeddingStoreTimeouts = "enable_sim_clusters_embedding_store_timeouts"
val simClustersEmbeddingStoreTimeoutValueMillis =
"sim_clusters_embedding_store_timeout_value_millis"
}
// Necessary for using servo Gates
object DeciderKey extends DeciderKeyEnum {
val enableLogFavBasedApeEntity20M145KUpdatedEmbeddingCachedStore: Value = Value(
DeciderConstants.enableLogFavBasedApeEntity20M145KUpdatedEmbeddingCachedStore
)
val enableLogFavBasedApeEntity20M145K2020EmbeddingCachedStore: Value = Value(
DeciderConstants.enableLogFavBasedApeEntity20M145K2020EmbeddingCachedStore
)
}

198
representation-manager/server/src/main/scala/com/twitter/representation_manager/store/TopicSimClustersEmbeddingStore.scala Normal file
View File

@ -0,0 +1,198 @@
package com.twitter.representation_manager.store
import com.twitter.contentrecommender.store.ApeEntityEmbeddingStore
import com.twitter.contentrecommender.store.InterestsOptOutStore
import com.twitter.contentrecommender.store.SemanticCoreTopicSeedStore
import com.twitter.conversions.DurationOps._
import com.twitter.escherbird.util.uttclient.CachedUttClientV2
import com.twitter.finagle.memcached.Client
import com.twitter.finagle.stats.StatsReceiver
import com.twitter.frigate.common.store.strato.StratoFetchableStore
import com.twitter.frigate.common.util.SeqLongInjection
import com.twitter.hermit.store.common.ObservedCachedReadableStore
import com.twitter.hermit.store.common.ObservedMemcachedReadableStore
import com.twitter.hermit.store.common.ObservedReadableStore
import com.twitter.interests.thriftscala.InterestsThriftService
import com.twitter.representation_manager.common.MemCacheConfig
import com.twitter.representation_manager.common.RepresentationManagerDecider
import com.twitter.simclusters_v2.common.SimClustersEmbedding
import com.twitter.simclusters_v2.stores.SimClustersEmbeddingStore
import com.twitter.simclusters_v2.thriftscala.EmbeddingType
import com.twitter.simclusters_v2.thriftscala.EmbeddingType._
import com.twitter.simclusters_v2.thriftscala.InternalId
import com.twitter.simclusters_v2.thriftscala.ModelVersion
import com.twitter.simclusters_v2.thriftscala.ModelVersion._
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.simclusters_v2.thriftscala.TopicId
import com.twitter.simclusters_v2.thriftscala.LocaleEntityId
import com.twitter.simclusters_v2.thriftscala.{SimClustersEmbedding => ThriftSimClustersEmbedding}
import com.twitter.storage.client.manhattan.kv.ManhattanKVClientMtlsParams
import com.twitter.storehaus.ReadableStore
import com.twitter.strato.client.{Client => StratoClient}
import com.twitter.tweetypie.util.UserId
import javax.inject.Inject
class TopicSimClustersEmbeddingStore @Inject() (
stratoClient: StratoClient,
cacheClient: Client,
globalStats: StatsReceiver,
mhMtlsParams: ManhattanKVClientMtlsParams,
rmsDecider: RepresentationManagerDecider,
interestService: InterestsThriftService.MethodPerEndpoint,
uttClient: CachedUttClientV2) {
private val stats = globalStats.scope(this.getClass.getSimpleName)
private val interestsOptOutStore = InterestsOptOutStore(interestService)
/**
* Note this is NOT an embedding store. It is a list of author account ids we use to represent
* topics
*/
private val semanticCoreTopicSeedStore: ReadableStore[
SemanticCoreTopicSeedStore.Key,
Seq[UserId]
] = {
/*
Up to 1000 Long seeds per topic/language = 62.5kb per topic/language (worst case)
Assume ~10k active topic/languages ~= 650MB (worst case)
*/
val underlying = new SemanticCoreTopicSeedStore(uttClient, interestsOptOutStore)(
stats.scope("semantic_core_topic_seed_store"))
val memcacheStore = ObservedMemcachedReadableStore.fromCacheClient(
backingStore = underlying,
cacheClient = cacheClient,
ttl = 12.hours)(
valueInjection = SeqLongInjection,
statsReceiver = stats.scope("topic_producer_seed_store_mem_cache"),
keyToString = { k => s"tpss:${k.entityId}_${k.languageCode}" }
)
ObservedCachedReadableStore.from[SemanticCoreTopicSeedStore.Key, Seq[UserId]](
store = memcacheStore,
ttl = 6.hours,
maxKeys = 20e3.toInt,
cacheName = "topic_producer_seed_store_cache",
windowSize = 5000
)(stats.scope("topic_producer_seed_store_cache"))
}
private val favBasedTfgTopicEmbedding20m145k2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore =
StratoFetchableStore
.withUnitView[SimClustersEmbeddingId, ThriftSimClustersEmbedding](
stratoClient,
"recommendations/simclusters_v2/embeddings/favBasedTFGTopic20M145K2020").mapValues(
embedding => SimClustersEmbedding(embedding, truncate = 50).toThrift)
.composeKeyMapping[LocaleEntityId] { localeEntityId =>
SimClustersEmbeddingId(
FavTfgTopic,
Model20m145k2020,
InternalId.LocaleEntityId(localeEntityId))
}
buildLocaleEntityIdMemCacheStore(rawStore, FavTfgTopic, Model20m145k2020)
}
private val logFavBasedApeEntity20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val apeStore = StratoFetchableStore
.withUnitView[SimClustersEmbeddingId, ThriftSimClustersEmbedding](
stratoClient,
"recommendations/simclusters_v2/embeddings/logFavBasedAPE20M145K2020")
.mapValues(embedding => SimClustersEmbedding(embedding, truncate = 50))
.composeKeyMapping[UserId]({ id =>
SimClustersEmbeddingId(
AggregatableLogFavBasedProducer,
Model20m145k2020,
InternalId.UserId(id))
})
val rawStore = new ApeEntityEmbeddingStore(
semanticCoreSeedStore = semanticCoreTopicSeedStore,
aggregatableProducerEmbeddingStore = apeStore,
statsReceiver = stats.scope("log_fav_based_ape_entity_2020_embedding_store"))
.mapValues(embedding => SimClustersEmbedding(embedding.toThrift, truncate = 50).toThrift)
.composeKeyMapping[TopicId] { topicId =>
SimClustersEmbeddingId(
LogFavBasedKgoApeTopic,
Model20m145k2020,
InternalId.TopicId(topicId))
}
buildTopicIdMemCacheStore(rawStore, LogFavBasedKgoApeTopic, Model20m145k2020)
}
private def buildTopicIdMemCacheStore(
rawStore: ReadableStore[TopicId, ThriftSimClustersEmbedding],
embeddingType: EmbeddingType,
modelVersion: ModelVersion
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val observedStore: ObservedReadableStore[TopicId, ThriftSimClustersEmbedding] =
ObservedReadableStore(
store = rawStore
)(stats.scope(embeddingType.name).scope(modelVersion.name))
val storeWithKeyMapping = observedStore.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.TopicId(topicId)) =>
topicId
}
MemCacheConfig.buildMemCacheStoreForSimClustersEmbedding(
storeWithKeyMapping,
cacheClient,
embeddingType,
modelVersion,
stats
)
}
private def buildLocaleEntityIdMemCacheStore(
rawStore: ReadableStore[LocaleEntityId, ThriftSimClustersEmbedding],
embeddingType: EmbeddingType,
modelVersion: ModelVersion
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val observedStore: ObservedReadableStore[LocaleEntityId, ThriftSimClustersEmbedding] =
ObservedReadableStore(
store = rawStore
)(stats.scope(embeddingType.name).scope(modelVersion.name))
val storeWithKeyMapping = observedStore.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.LocaleEntityId(localeEntityId)) =>
localeEntityId
}
MemCacheConfig.buildMemCacheStoreForSimClustersEmbedding(
storeWithKeyMapping,
cacheClient,
embeddingType,
modelVersion,
stats
)
}
private val underlyingStores: Map[
(EmbeddingType, ModelVersion),
ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding]
] = Map(
// Topic Embeddings
(FavTfgTopic, Model20m145k2020) -> favBasedTfgTopicEmbedding20m145k2020Store,
(LogFavBasedKgoApeTopic, Model20m145k2020) -> logFavBasedApeEntity20M145K2020EmbeddingStore,
)
val topicSimClustersEmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
SimClustersEmbeddingStore.buildWithDecider(
underlyingStores = underlyingStores,
decider = rmsDecider.decider,
statsReceiver = stats
)
}
}

141
representation-manager/server/src/main/scala/com/twitter/representation_manager/store/TweetSimClustersEmbeddingStore.scala Normal file
View File

@ -0,0 +1,141 @@
package com.twitter.representation_manager.store
import com.twitter.finagle.memcached.Client
import com.twitter.finagle.stats.StatsReceiver
import com.twitter.hermit.store.common.ObservedReadableStore
import com.twitter.representation_manager.common.MemCacheConfig
import com.twitter.representation_manager.common.RepresentationManagerDecider
import com.twitter.simclusters_v2.common.SimClustersEmbedding
import com.twitter.simclusters_v2.common.TweetId
import com.twitter.simclusters_v2.stores.SimClustersEmbeddingStore
import com.twitter.simclusters_v2.summingbird.stores.PersistentTweetEmbeddingStore
import com.twitter.simclusters_v2.thriftscala.EmbeddingType
import com.twitter.simclusters_v2.thriftscala.EmbeddingType._
import com.twitter.simclusters_v2.thriftscala.InternalId
import com.twitter.simclusters_v2.thriftscala.ModelVersion
import com.twitter.simclusters_v2.thriftscala.ModelVersion._
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.simclusters_v2.thriftscala.{SimClustersEmbedding => ThriftSimClustersEmbedding}
import com.twitter.storage.client.manhattan.kv.ManhattanKVClientMtlsParams
import com.twitter.storehaus.ReadableStore
import javax.inject.Inject
class TweetSimClustersEmbeddingStore @Inject() (
cacheClient: Client,
globalStats: StatsReceiver,
mhMtlsParams: ManhattanKVClientMtlsParams,
rmsDecider: RepresentationManagerDecider) {
private val stats = globalStats.scope(this.getClass.getSimpleName)
val logFavBasedLongestL2Tweet20M145KUpdatedEmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore =
PersistentTweetEmbeddingStore
.longestL2NormTweetEmbeddingStoreManhattan(
mhMtlsParams,
PersistentTweetEmbeddingStore.LogFavBased20m145kUpdatedDataset,
stats
).mapValues(_.toThrift)
buildMemCacheStore(rawStore, LogFavLongestL2EmbeddingTweet, Model20m145kUpdated)
}
val logFavBasedLongestL2Tweet20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore =
PersistentTweetEmbeddingStore
.longestL2NormTweetEmbeddingStoreManhattan(
mhMtlsParams,
PersistentTweetEmbeddingStore.LogFavBased20m145k2020Dataset,
stats
).mapValues(_.toThrift)
buildMemCacheStore(rawStore, LogFavLongestL2EmbeddingTweet, Model20m145k2020)
}
val logFavBased20M145KUpdatedTweetEmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore =
PersistentTweetEmbeddingStore
.mostRecentTweetEmbeddingStoreManhattan(
mhMtlsParams,
PersistentTweetEmbeddingStore.LogFavBased20m145kUpdatedDataset,
stats
).mapValues(_.toThrift)
buildMemCacheStore(rawStore, LogFavBasedTweet, Model20m145kUpdated)
}
val logFavBased20M145K2020TweetEmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore =
PersistentTweetEmbeddingStore
.mostRecentTweetEmbeddingStoreManhattan(
mhMtlsParams,
PersistentTweetEmbeddingStore.LogFavBased20m145k2020Dataset,
stats
).mapValues(_.toThrift)
buildMemCacheStore(rawStore, LogFavBasedTweet, Model20m145k2020)
}
private def buildMemCacheStore(
rawStore: ReadableStore[TweetId, ThriftSimClustersEmbedding],
embeddingType: EmbeddingType,
modelVersion: ModelVersion
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val observedStore: ObservedReadableStore[TweetId, ThriftSimClustersEmbedding] =
ObservedReadableStore(
store = rawStore
)(stats.scope(embeddingType.name).scope(modelVersion.name))
val storeWithKeyMapping = observedStore.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.TweetId(tweetId)) =>
tweetId
}
MemCacheConfig.buildMemCacheStoreForSimClustersEmbedding(
storeWithKeyMapping,
cacheClient,
embeddingType,
modelVersion,
stats
)
}
private val underlyingStores: Map[
(EmbeddingType, ModelVersion),
ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding]
] = Map(
// Tweet Embeddings
(LogFavBasedTweet, Model20m145kUpdated) -> logFavBased20M145KUpdatedTweetEmbeddingStore,
(LogFavBasedTweet, Model20m145k2020) -> logFavBased20M145K2020TweetEmbeddingStore,
(
LogFavLongestL2EmbeddingTweet,
Model20m145kUpdated) -> logFavBasedLongestL2Tweet20M145KUpdatedEmbeddingStore,
(
LogFavLongestL2EmbeddingTweet,
Model20m145k2020) -> logFavBasedLongestL2Tweet20M145K2020EmbeddingStore,
)
val tweetSimClustersEmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
SimClustersEmbeddingStore.buildWithDecider(
underlyingStores = underlyingStores,
decider = rmsDecider.decider,
statsReceiver = stats
)
}
}

602
representation-manager/server/src/main/scala/com/twitter/representation_manager/store/UserSimClustersEmbeddingStore.scala Normal file
View File

@ -0,0 +1,602 @@
package com.twitter.representation_manager.store
import com.twitter.contentrecommender.twistly
import com.twitter.finagle.memcached.Client
import com.twitter.finagle.stats.StatsReceiver
import com.twitter.frigate.common.store.strato.StratoFetchableStore
import com.twitter.hermit.store.common.ObservedReadableStore
import com.twitter.representation_manager.common.MemCacheConfig
import com.twitter.representation_manager.common.RepresentationManagerDecider
import com.twitter.simclusters_v2.common.ModelVersions
import com.twitter.simclusters_v2.common.SimClustersEmbedding
import com.twitter.simclusters_v2.stores.SimClustersEmbeddingStore
import com.twitter.simclusters_v2.summingbird.stores.ProducerClusterEmbeddingReadableStores
import com.twitter.simclusters_v2.summingbird.stores.UserInterestedInReadableStore
import com.twitter.simclusters_v2.summingbird.stores.UserInterestedInReadableStore.getStore
import com.twitter.simclusters_v2.summingbird.stores.UserInterestedInReadableStore.modelVersionToDatasetMap
import com.twitter.simclusters_v2.summingbird.stores.UserInterestedInReadableStore.knownModelVersions
import com.twitter.simclusters_v2.summingbird.stores.UserInterestedInReadableStore.toSimClustersEmbedding
import com.twitter.simclusters_v2.thriftscala.ClustersUserIsInterestedIn
import com.twitter.simclusters_v2.thriftscala.EmbeddingType
import com.twitter.simclusters_v2.thriftscala.EmbeddingType._
import com.twitter.simclusters_v2.thriftscala.InternalId
import com.twitter.simclusters_v2.thriftscala.ModelVersion
import com.twitter.simclusters_v2.thriftscala.ModelVersion._
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.simclusters_v2.thriftscala.{SimClustersEmbedding => ThriftSimClustersEmbedding}
import com.twitter.storage.client.manhattan.kv.ManhattanKVClientMtlsParams
import com.twitter.storehaus.ReadableStore
import com.twitter.storehaus_internal.manhattan.Apollo
import com.twitter.storehaus_internal.manhattan.ManhattanCluster
import com.twitter.strato.client.{Client => StratoClient}
import com.twitter.strato.thrift.ScroogeConvImplicits._
import com.twitter.tweetypie.util.UserId
import com.twitter.util.Future
import javax.inject.Inject
class UserSimClustersEmbeddingStore @Inject() (
stratoClient: StratoClient,
cacheClient: Client,
globalStats: StatsReceiver,
mhMtlsParams: ManhattanKVClientMtlsParams,
rmsDecider: RepresentationManagerDecider) {
private val stats = globalStats.scope(this.getClass.getSimpleName)
private val favBasedProducer20M145KUpdatedEmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore = ProducerClusterEmbeddingReadableStores
.getProducerTopKSimClustersEmbeddingsStore(
mhMtlsParams
).mapValues { topSimClustersWithScore =>
ThriftSimClustersEmbedding(topSimClustersWithScore.topClusters)
}.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.UserId(userId)) =>
userId
}
buildMemCacheStore(rawStore, FavBasedProducer, Model20m145kUpdated)
}
private val favBasedProducer20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore = ProducerClusterEmbeddingReadableStores
.getProducerTopKSimClusters2020EmbeddingsStore(
mhMtlsParams
).mapValues { topSimClustersWithScore =>
ThriftSimClustersEmbedding(topSimClustersWithScore.topClusters)
}.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.UserId(userId)) =>
userId
}
buildMemCacheStore(rawStore, FavBasedProducer, Model20m145k2020)
}
private val followBasedProducer20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore = ProducerClusterEmbeddingReadableStores
.getProducerTopKSimClustersEmbeddingsByFollowStore(
mhMtlsParams
).mapValues { topSimClustersWithScore =>
ThriftSimClustersEmbedding(topSimClustersWithScore.topClusters)
}.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(_, _, InternalId.UserId(userId)) =>
userId
}
buildMemCacheStore(rawStore, FollowBasedProducer, Model20m145k2020)
}
private val logFavBasedApe20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore = StratoFetchableStore
.withUnitView[SimClustersEmbeddingId, ThriftSimClustersEmbedding](
stratoClient,
"recommendations/simclusters_v2/embeddings/logFavBasedAPE20M145K2020")
.mapValues(embedding => SimClustersEmbedding(embedding, truncate = 50).toThrift)
buildMemCacheStore(rawStore, AggregatableLogFavBasedProducer, Model20m145k2020)
}
private val rawRelaxedLogFavBasedApe20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
ThriftSimClustersEmbedding
] = {
StratoFetchableStore
.withUnitView[SimClustersEmbeddingId, ThriftSimClustersEmbedding](
stratoClient,
"recommendations/simclusters_v2/embeddings/logFavBasedAPERelaxedFavEngagementThreshold20M145K2020")
.mapValues(embedding => SimClustersEmbedding(embedding, truncate = 50).toThrift)
}
private val relaxedLogFavBasedApe20M145K2020EmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildMemCacheStore(
rawRelaxedLogFavBasedApe20M145K2020EmbeddingStore,
RelaxedAggregatableLogFavBasedProducer,
Model20m145k2020)
}
private val relaxedLogFavBasedApe20m145kUpdatedEmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore = rawRelaxedLogFavBasedApe20M145K2020EmbeddingStore
.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(
RelaxedAggregatableLogFavBasedProducer,
Model20m145kUpdated,
internalId) =>
SimClustersEmbeddingId(
RelaxedAggregatableLogFavBasedProducer,
Model20m145k2020,
internalId)
}
buildMemCacheStore(rawStore, RelaxedAggregatableLogFavBasedProducer, Model20m145kUpdated)
}
private val logFavBasedInterestedInFromAPE20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildUserInterestedInStore(
UserInterestedInReadableStore.defaultIIAPESimClustersEmbeddingStoreWithMtls,
LogFavBasedUserInterestedInFromAPE,
Model20m145k2020)
}
private val followBasedInterestedInFromAPE20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildUserInterestedInStore(
UserInterestedInReadableStore.defaultIIAPESimClustersEmbeddingStoreWithMtls,
FollowBasedUserInterestedInFromAPE,
Model20m145k2020)
}
private val favBasedUserInterestedIn20M145KUpdatedStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildUserInterestedInStore(
UserInterestedInReadableStore.defaultSimClustersEmbeddingStoreWithMtls,
FavBasedUserInterestedIn,
Model20m145kUpdated)
}
private val favBasedUserInterestedIn20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildUserInterestedInStore(
UserInterestedInReadableStore.defaultSimClustersEmbeddingStoreWithMtls,
FavBasedUserInterestedIn,
Model20m145k2020)
}
private val followBasedUserInterestedIn20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildUserInterestedInStore(
UserInterestedInReadableStore.defaultSimClustersEmbeddingStoreWithMtls,
FollowBasedUserInterestedIn,
Model20m145k2020)
}
private val logFavBasedUserInterestedIn20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildUserInterestedInStore(
UserInterestedInReadableStore.defaultSimClustersEmbeddingStoreWithMtls,
LogFavBasedUserInterestedIn,
Model20m145k2020)
}
private val favBasedUserInterestedInFromPE20M145KUpdatedStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildUserInterestedInStore(
UserInterestedInReadableStore.defaultIIPESimClustersEmbeddingStoreWithMtls,
FavBasedUserInterestedInFromPE,
Model20m145kUpdated)
}
private val twistlyUserInterestedInStore: ReadableStore[
SimClustersEmbeddingId,
ThriftSimClustersEmbedding
] = {
val interestedIn20M145KUpdatedStore = {
UserInterestedInReadableStore.defaultStoreWithMtls(
mhMtlsParams,
modelVersion = ModelVersions.Model20M145KUpdated
)
}
val interestedIn20M145K2020Store = {
UserInterestedInReadableStore.defaultStoreWithMtls(
mhMtlsParams,
modelVersion = ModelVersions.Model20M145K2020
)
}
val interestedInFromPE20M145KUpdatedStore = {
UserInterestedInReadableStore.defaultIIPEStoreWithMtls(
mhMtlsParams,
modelVersion = ModelVersions.Model20M145KUpdated)
}
val simClustersInterestedInStore: ReadableStore[
(UserId, ModelVersion),
ClustersUserIsInterestedIn
] = {
new ReadableStore[(UserId, ModelVersion), ClustersUserIsInterestedIn] {
override def get(k: (UserId, ModelVersion)): Future[Option[ClustersUserIsInterestedIn]] = {
k match {
case (userId, Model20m145kUpdated) =>
interestedIn20M145KUpdatedStore.get(userId)
case (userId, Model20m145k2020) =>
interestedIn20M145K2020Store.get(userId)
case _ =>
Future.None
}
}
}
}
val simClustersInterestedInFromProducerEmbeddingsStore: ReadableStore[
(UserId, ModelVersion),
ClustersUserIsInterestedIn
] = {
new ReadableStore[(UserId, ModelVersion), ClustersUserIsInterestedIn] {
override def get(k: (UserId, ModelVersion)): Future[Option[ClustersUserIsInterestedIn]] = {
k match {
case (userId, ModelVersion.Model20m145kUpdated) =>
interestedInFromPE20M145KUpdatedStore.get(userId)
case _ =>
Future.None
}
}
}
}
new twistly.interestedin.EmbeddingStore(
interestedInStore = simClustersInterestedInStore,
interestedInFromProducerEmbeddingStore = simClustersInterestedInFromProducerEmbeddingsStore,
statsReceiver = stats
).mapValues(_.toThrift)
}
private val userNextInterestedIn20m145k2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildUserInterestedInStore(
UserInterestedInReadableStore.defaultNextInterestedInStoreWithMtls,
UserNextInterestedIn,
Model20m145k2020)
}
private val filteredUserInterestedIn20m145kUpdatedStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildMemCacheStore(twistlyUserInterestedInStore, FilteredUserInterestedIn, Model20m145kUpdated)
}
private val filteredUserInterestedIn20m145k2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildMemCacheStore(twistlyUserInterestedInStore, FilteredUserInterestedIn, Model20m145k2020)
}
private val filteredUserInterestedInFromPE20m145kUpdatedStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildMemCacheStore(
twistlyUserInterestedInStore,
FilteredUserInterestedInFromPE,
Model20m145kUpdated)
}
private val unfilteredUserInterestedIn20m145kUpdatedStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildMemCacheStore(
twistlyUserInterestedInStore,
UnfilteredUserInterestedIn,
Model20m145kUpdated)
}
private val unfilteredUserInterestedIn20m145k2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
buildMemCacheStore(twistlyUserInterestedInStore, UnfilteredUserInterestedIn, Model20m145k2020)
}
// [Experimental] User InterestedIn, generated by aggregating IIAPE embedding from AddressBook
private val logFavBasedInterestedMaxpoolingAddressBookFromIIAPE20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val datasetName = "addressbook_sims_embedding_iiape_maxpooling"
val appId = "wtf_embedding_apollo"
buildUserInterestedInStoreGeneric(
simClustersEmbeddingStoreWithMtls,
LogFavBasedUserInterestedMaxpoolingAddressBookFromIIAPE,
Model20m145k2020,
datasetName = datasetName,
appId = appId,
manhattanCluster = Apollo
)
}
private val logFavBasedInterestedAverageAddressBookFromIIAPE20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val datasetName = "addressbook_sims_embedding_iiape_average"
val appId = "wtf_embedding_apollo"
buildUserInterestedInStoreGeneric(
simClustersEmbeddingStoreWithMtls,
LogFavBasedUserInterestedAverageAddressBookFromIIAPE,
Model20m145k2020,
datasetName = datasetName,
appId = appId,
manhattanCluster = Apollo
)
}
private val logFavBasedUserInterestedBooktypeMaxpoolingAddressBookFromIIAPE20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val datasetName = "addressbook_sims_embedding_iiape_booktype_maxpooling"
val appId = "wtf_embedding_apollo"
buildUserInterestedInStoreGeneric(
simClustersEmbeddingStoreWithMtls,
LogFavBasedUserInterestedBooktypeMaxpoolingAddressBookFromIIAPE,
Model20m145k2020,
datasetName = datasetName,
appId = appId,
manhattanCluster = Apollo
)
}
private val logFavBasedUserInterestedLargestDimMaxpoolingAddressBookFromIIAPE20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val datasetName = "addressbook_sims_embedding_iiape_largestdim_maxpooling"
val appId = "wtf_embedding_apollo"
buildUserInterestedInStoreGeneric(
simClustersEmbeddingStoreWithMtls,
LogFavBasedUserInterestedLargestDimMaxpoolingAddressBookFromIIAPE,
Model20m145k2020,
datasetName = datasetName,
appId = appId,
manhattanCluster = Apollo
)
}
private val logFavBasedUserInterestedLouvainMaxpoolingAddressBookFromIIAPE20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val datasetName = "addressbook_sims_embedding_iiape_louvain_maxpooling"
val appId = "wtf_embedding_apollo"
buildUserInterestedInStoreGeneric(
simClustersEmbeddingStoreWithMtls,
LogFavBasedUserInterestedLouvainMaxpoolingAddressBookFromIIAPE,
Model20m145k2020,
datasetName = datasetName,
appId = appId,
manhattanCluster = Apollo
)
}
private val logFavBasedUserInterestedConnectedMaxpoolingAddressBookFromIIAPE20M145K2020Store: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val datasetName = "addressbook_sims_embedding_iiape_connected_maxpooling"
val appId = "wtf_embedding_apollo"
buildUserInterestedInStoreGeneric(
simClustersEmbeddingStoreWithMtls,
LogFavBasedUserInterestedConnectedMaxpoolingAddressBookFromIIAPE,
Model20m145k2020,
datasetName = datasetName,
appId = appId,
manhattanCluster = Apollo
)
}
/**
* Helper func to build a readable store for some UserInterestedIn embeddings with
* 1. A storeFunc from UserInterestedInReadableStore
* 2. EmbeddingType
* 3. ModelVersion
* 4. MemCacheConfig
* */
private def buildUserInterestedInStore(
storeFunc: (ManhattanKVClientMtlsParams, EmbeddingType, ModelVersion) => ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
],
embeddingType: EmbeddingType,
modelVersion: ModelVersion
): ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore = storeFunc(mhMtlsParams, embeddingType, modelVersion)
.mapValues(_.toThrift)
val observedStore = ObservedReadableStore(
store = rawStore
)(stats.scope(embeddingType.name).scope(modelVersion.name))
MemCacheConfig.buildMemCacheStoreForSimClustersEmbedding(
observedStore,
cacheClient,
embeddingType,
modelVersion,
stats
)
}
private def buildUserInterestedInStoreGeneric(
storeFunc: (ManhattanKVClientMtlsParams, EmbeddingType, ModelVersion, String, String,
ManhattanCluster) => ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
],
embeddingType: EmbeddingType,
modelVersion: ModelVersion,
datasetName: String,
appId: String,
manhattanCluster: ManhattanCluster
): ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
val rawStore =
storeFunc(mhMtlsParams, embeddingType, modelVersion, datasetName, appId, manhattanCluster)
.mapValues(_.toThrift)
val observedStore = ObservedReadableStore(
store = rawStore
)(stats.scope(embeddingType.name).scope(modelVersion.name))
MemCacheConfig.buildMemCacheStoreForSimClustersEmbedding(
observedStore,
cacheClient,
embeddingType,
modelVersion,
stats
)
}
private def simClustersEmbeddingStoreWithMtls(
mhMtlsParams: ManhattanKVClientMtlsParams,
embeddingType: EmbeddingType,
modelVersion: ModelVersion,
datasetName: String,
appId: String,
manhattanCluster: ManhattanCluster
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
if (!modelVersionToDatasetMap.contains(ModelVersions.toKnownForModelVersion(modelVersion))) {
throw new IllegalArgumentException(
"Unknown model version: " + modelVersion + ". Known model versions: " + knownModelVersions)
}
getStore(appId, mhMtlsParams, datasetName, manhattanCluster)
.composeKeyMapping[SimClustersEmbeddingId] {
case SimClustersEmbeddingId(theEmbeddingType, theModelVersion, InternalId.UserId(userId))
if theEmbeddingType == embeddingType && theModelVersion == modelVersion =>
userId
}.mapValues(toSimClustersEmbedding(_, embeddingType))
}
private def buildMemCacheStore(
rawStore: ReadableStore[SimClustersEmbeddingId, ThriftSimClustersEmbedding],
embeddingType: EmbeddingType,
modelVersion: ModelVersion
): ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding] = {
val observedStore = ObservedReadableStore(
store = rawStore
)(stats.scope(embeddingType.name).scope(modelVersion.name))
MemCacheConfig.buildMemCacheStoreForSimClustersEmbedding(
observedStore,
cacheClient,
embeddingType,
modelVersion,
stats
)
}
private val underlyingStores: Map[
(EmbeddingType, ModelVersion),
ReadableStore[SimClustersEmbeddingId, SimClustersEmbedding]
] = Map(
// KnownFor Embeddings
(FavBasedProducer, Model20m145kUpdated) -> favBasedProducer20M145KUpdatedEmbeddingStore,
(FavBasedProducer, Model20m145k2020) -> favBasedProducer20M145K2020EmbeddingStore,
(FollowBasedProducer, Model20m145k2020) -> followBasedProducer20M145K2020EmbeddingStore,
(AggregatableLogFavBasedProducer, Model20m145k2020) -> logFavBasedApe20M145K2020EmbeddingStore,
(
RelaxedAggregatableLogFavBasedProducer,
Model20m145kUpdated) -> relaxedLogFavBasedApe20m145kUpdatedEmbeddingStore,
(
RelaxedAggregatableLogFavBasedProducer,
Model20m145k2020) -> relaxedLogFavBasedApe20M145K2020EmbeddingStore,
// InterestedIn Embeddings
(
LogFavBasedUserInterestedInFromAPE,
Model20m145k2020) -> logFavBasedInterestedInFromAPE20M145K2020Store,
(
FollowBasedUserInterestedInFromAPE,
Model20m145k2020) -> followBasedInterestedInFromAPE20M145K2020Store,
(FavBasedUserInterestedIn, Model20m145kUpdated) -> favBasedUserInterestedIn20M145KUpdatedStore,
(FavBasedUserInterestedIn, Model20m145k2020) -> favBasedUserInterestedIn20M145K2020Store,
(FollowBasedUserInterestedIn, Model20m145k2020) -> followBasedUserInterestedIn20M145K2020Store,
(LogFavBasedUserInterestedIn, Model20m145k2020) -> logFavBasedUserInterestedIn20M145K2020Store,
(
FavBasedUserInterestedInFromPE,
Model20m145kUpdated) -> favBasedUserInterestedInFromPE20M145KUpdatedStore,
(FilteredUserInterestedIn, Model20m145kUpdated) -> filteredUserInterestedIn20m145kUpdatedStore,
(FilteredUserInterestedIn, Model20m145k2020) -> filteredUserInterestedIn20m145k2020Store,
(
FilteredUserInterestedInFromPE,
Model20m145kUpdated) -> filteredUserInterestedInFromPE20m145kUpdatedStore,
(
UnfilteredUserInterestedIn,
Model20m145kUpdated) -> unfilteredUserInterestedIn20m145kUpdatedStore,
(UnfilteredUserInterestedIn, Model20m145k2020) -> unfilteredUserInterestedIn20m145k2020Store,
(UserNextInterestedIn, Model20m145k2020) -> userNextInterestedIn20m145k2020Store,
(
LogFavBasedUserInterestedMaxpoolingAddressBookFromIIAPE,
Model20m145k2020) -> logFavBasedInterestedMaxpoolingAddressBookFromIIAPE20M145K2020Store,
(
LogFavBasedUserInterestedAverageAddressBookFromIIAPE,
Model20m145k2020) -> logFavBasedInterestedAverageAddressBookFromIIAPE20M145K2020Store,
(
LogFavBasedUserInterestedBooktypeMaxpoolingAddressBookFromIIAPE,
Model20m145k2020) -> logFavBasedUserInterestedBooktypeMaxpoolingAddressBookFromIIAPE20M145K2020Store,
(
LogFavBasedUserInterestedLargestDimMaxpoolingAddressBookFromIIAPE,
Model20m145k2020) -> logFavBasedUserInterestedLargestDimMaxpoolingAddressBookFromIIAPE20M145K2020Store,
(
LogFavBasedUserInterestedLouvainMaxpoolingAddressBookFromIIAPE,
Model20m145k2020) -> logFavBasedUserInterestedLouvainMaxpoolingAddressBookFromIIAPE20M145K2020Store,
(
LogFavBasedUserInterestedConnectedMaxpoolingAddressBookFromIIAPE,
Model20m145k2020) -> logFavBasedUserInterestedConnectedMaxpoolingAddressBookFromIIAPE20M145K2020Store,
)
val userSimClustersEmbeddingStore: ReadableStore[
SimClustersEmbeddingId,
SimClustersEmbedding
] = {
SimClustersEmbeddingStore.buildWithDecider(
underlyingStores = underlyingStores,
decider = rmsDecider.decider,
statsReceiver = stats
)
}
}

18
representation-manager/server/src/main/thrift/BUILD Normal file
View File

@ -0,0 +1,18 @@
create_thrift_libraries(
base_name = "thrift",
sources = [
"com/twitter/representation_manager/service.thrift",
],
platform = "java8",
tags = [
"bazel-compatible",
],
dependency_roots = [
"src/thrift/com/twitter/simclusters_v2:simclusters_v2-thrift",
],
generate_languages = [
"java",
"scala",
"strato",
],
)

14
representation-manager/server/src/main/thrift/com/twitter/representation_manager/service.thrift Normal file
View File

@ -0,0 +1,14 @@
namespace java com.twitter.representation_manager.thriftjava
#@namespace scala com.twitter.representation_manager.thriftscala
#@namespace strato com.twitter.representation_manager
include "com/twitter/simclusters_v2/online_store.thrift"
include "com/twitter/simclusters_v2/identifier.thrift"
/**
* A uniform column view for all kinds of SimClusters based embeddings.
**/
struct SimClustersEmbeddingView {
1: required identifier.EmbeddingType embeddingType
2: required online_store.ModelVersion modelVersion
}(persisted = 'false', hasPersonalData = 'false')

1
representation-scorer/BUILD.bazel Normal file
View File

@ -0,0 +1 @@
# This prevents SQ query from grabbing //:all since it traverses up once to find a BUILD

5
representation-scorer/README.md Normal file
View File

@ -0,0 +1,5 @@
# Representation Scorer #
**Representation Scorer** (RSX) serves as a centralized scoring system, offering SimClusters or other embedding-based scoring solutions as machine learning features.
The Representation Scorer acquires user behavior data from the User Signal Service (USS) and extracts embeddings from the Representation Manager (RMS). It then calculates both pairwise and listwise features. These features are used at various stages, including candidate retrieval and ranking.

8
representation-scorer/bin/canary-check.sh Executable file
View File

@ -0,0 +1,8 @@
#!/bin/bash
export CANARY_CHECK_ROLE="representation-scorer"
export CANARY_CHECK_NAME="representation-scorer"
export CANARY_CHECK_INSTANCES="0-19"
python3 relevance-platform/tools/canary_check.py "$@"

4
representation-scorer/bin/deploy.sh Executable file
View File

@ -0,0 +1,4 @@
#!/usr/bin/env bash
JOB=representation-scorer bazel run --ui_event_filters=-info,-stdout,-stderr --noshow_progress \
//relevance-platform/src/main/python/deploy -- "$@"

66
representation-scorer/bin/remote-debug-tunnel.sh Executable file
View File

@ -0,0 +1,66 @@
#!/bin/bash
set -o nounset
set -eu
DC="atla"
ROLE="$USER"
SERVICE="representation-scorer"
INSTANCE="0"
KEY="$DC/$ROLE/devel/$SERVICE/$INSTANCE"
while test $# -gt 0; do
case "$1" in
-h|--help)
echo "$0 Set up an ssh tunnel for $SERVICE remote debugging and disable aurora health checks"
echo " "
echo "See representation-scorer/README.md for details of how to use this script, and go/remote-debug for"
echo "general information about remote debugging in Aurora"
echo " "
echo "Default instance if called with no args:"
echo " $KEY"
echo " "
echo "Positional args:"
echo " $0 [datacentre] [role] [service_name] [instance]"
echo " "
echo "Options:"
echo " -h, --help show brief help"
exit 0
;;
*)
break
;;
esac
done
if [ -n "${1-}" ]; then
DC="$1"
fi
if [ -n "${2-}" ]; then
ROLE="$2"
fi
if [ -n "${3-}" ]; then
SERVICE="$3"
fi
if [ -n "${4-}" ]; then
INSTANCE="$4"
fi
KEY="$DC/$ROLE/devel/$SERVICE/$INSTANCE"
read -p "Set up remote debugger tunnel for $KEY? (y/n) " -r CONFIRM
if [[ ! $CONFIRM =~ ^[Yy]$ ]]; then
echo "Exiting, tunnel not created"
exit 1
fi
echo "Disabling health check and opening tunnel. Exit with control-c when you're finished"
CMD="aurora task ssh $KEY -c 'touch .healthchecksnooze' && aurora task ssh $KEY -L '5005:debug' --ssh-options '-N -S none -v '"
echo "Running $CMD"
eval "$CMD"

39
representation-scorer/docs/index.rst Normal file
View File

@ -0,0 +1,39 @@
Representation Scorer (RSX)
###########################
Overview
========
Representation Scorer (RSX) is a StratoFed service which serves scores for pairs of entities (User, Tweet, Topic...) based on some representation of those entities. For example, it serves User-Tweet scores based on the cosine similarity of SimClusters embeddings for each of these. It aims to provide these with low latency and at high scale, to support applications such as scoring for ANN candidate generation and feature hydration via feature store.
Current use cases
-----------------
RSX currently serves traffic for the following use cases:
- User-Tweet similarity scores for Home ranking, using SimClusters embedding dot product
- Topic-Tweet similarity scores for topical tweet candidate generation and topic social proof, using SimClusters embedding cosine similarity and CERTO scores
- Tweet-Tweet and User-Tweet similarity scores for ANN candidate generation, using SimClusters embedding cosine similarity
- (in development) User-Tweet similarity scores for Home ranking, based on various aggregations of similarities with recent faves, retweets and follows performed by the user
Getting Started
===============
Fetching scores
---------------
Scores are served from the recommendations/representation_scorer/score column.
Using RSX for your application
------------------------------
RSX may be a good fit for your application if you need scores based on combinations of SimCluster embeddings for core nouns. We also plan to support other embeddings and scoring approaches in the future.
.. toctree::
:maxdepth: 2
:hidden:
index

22
representation-scorer/server/BUILD Normal file
View File

@ -0,0 +1,22 @@
jvm_binary(
name = "bin",
basename = "representation-scorer",
main = "com.twitter.representationscorer.RepresentationScorerFedServerMain",
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"finatra/inject/inject-logback/src/main/scala",
"loglens/loglens-logback/src/main/scala/com/twitter/loglens/logback",
"representation-scorer/server/src/main/resources",
"representation-scorer/server/src/main/scala/com/twitter/representationscorer",
"twitter-server/logback-classic/src/main/scala",
],
)
# Aurora Workflows build phase convention requires a jvm_app named with ${project-name}-app
jvm_app(
name = "representation-scorer-app",
archive = "zip",
binary = ":bin",
tags = ["bazel-compatible"],
)

9
representation-scorer/server/src/main/resources/BUILD Normal file
View File

@ -0,0 +1,9 @@
resources(
sources = [
"*.xml",
"*.yml",
"com/twitter/slo/slo.json",
"config/*.yml",
],
tags = ["bazel-compatible"],
)

55
representation-scorer/server/src/main/resources/com/twitter/slo/slo.json Normal file
View File

@ -0,0 +1,55 @@
{
"servers": [
{
"name": "strato",
"indicators": [
{
"id": "success_rate_3m",
"indicator_type": "SuccessRateIndicator",
"duration": 3,
"duration_unit": "MINUTES"
}, {
"id": "latency_3m_p99",
"indicator_type": "LatencyIndicator",
"duration": 3,
"duration_unit": "MINUTES",
"percentile": 0.99
}
],
"objectives": [
{
"indicator": "success_rate_3m",
"objective_type": "SuccessRateObjective",
"operator": ">=",
"threshold": 0.995
},
{
"indicator": "latency_3m_p99",
"objective_type": "LatencyObjective",
"operator": "<=",
"threshold": 50
}
],
"long_term_objectives": [
{
"id": "success_rate_28_days",
"objective_type": "SuccessRateObjective",
"operator": ">=",
"threshold": 0.993,
"duration": 28,
"duration_unit": "DAYS"
},
{
"id": "latency_p99_28_days",
"objective_type": "LatencyObjective",
"operator": "<=",
"threshold": 60,
"duration": 28,
"duration_unit": "DAYS",
"percentile": 0.99
}
]
}
],
"@version": 1
}

155
representation-scorer/server/src/main/resources/config/decider.yml Normal file
View File

@ -0,0 +1,155 @@
enableLogFavBasedApeEntity20M145KUpdatedEmbeddingCachedStore:
comment: "Enable to use the non-empty store for logFavBasedApeEntity20M145KUpdatedEmbeddingCachedStore (from 0% to 100%). 0 means use EMPTY readable store for all requests."
default_availability: 0
enableLogFavBasedApeEntity20M145K2020EmbeddingCachedStore:
comment: "Enable to use the non-empty store for logFavBasedApeEntity20M145K2020EmbeddingCachedStore (from 0% to 100%). 0 means use EMPTY readable store for all requests."
default_availability: 0
representation-scorer_forward_dark_traffic:
comment: "Defines the percentage of traffic to forward to diffy-proxy. Set to 0 to disable dark traffic forwarding"
default_availability: 0
"representation-scorer_load_shed_non_prod_callers":
comment: "Discard traffic from all non-prod callers"
default_availability: 0
enable_log_fav_based_tweet_embedding_20m145k2020_timeouts:
comment: "If enabled, set a timeout on calls to the logFavBased20M145K2020TweetEmbeddingStore"
default_availability: 0
log_fav_based_tweet_embedding_20m145k2020_timeout_value_millis:
comment: "The value of this decider defines the timeout (in milliseconds) to use on calls to the logFavBased20M145K2020TweetEmbeddingStore, i.e. 1.50% is 150ms. Only applied if enable_log_fav_based_tweet_embedding_20m145k2020_timeouts is true"
default_availability: 2000
enable_log_fav_based_tweet_embedding_20m145kUpdated_timeouts:
comment: "If enabled, set a timeout on calls to the logFavBased20M145KUpdatedTweetEmbeddingStore"
default_availability: 0
log_fav_based_tweet_embedding_20m145kUpdated_timeout_value_millis:
comment: "The value of this decider defines the timeout (in milliseconds) to use on calls to the logFavBased20M145KUpdatedTweetEmbeddingStore, i.e. 1.50% is 150ms. Only applied if enable_log_fav_based_tweet_embedding_20m145kUpdated_timeouts is true"
default_availability: 2000
enable_cluster_tweet_index_store_timeouts:
comment: "If enabled, set a timeout on calls to the ClusterTweetIndexStore"
default_availability: 0
cluster_tweet_index_store_timeout_value_millis:
comment: "The value of this decider defines the timeout (in milliseconds) to use on calls to the ClusterTweetIndexStore, i.e. 1.50% is 150ms. Only applied if enable_cluster_tweet_index_store_timeouts is true"
default_availability: 2000
representation_scorer_fetch_signal_share:
comment: "If enabled, fetches share signals from USS"
default_availability: 0
representation_scorer_fetch_signal_reply:
comment: "If enabled, fetches reply signals from USS"
default_availability: 0
representation_scorer_fetch_signal_original_tweet:
comment: "If enabled, fetches original tweet signals from USS"
default_availability: 0
representation_scorer_fetch_signal_video_playback:
comment: "If enabled, fetches video playback signals from USS"
default_availability: 0
representation_scorer_fetch_signal_block:
comment: "If enabled, fetches account block signals from USS"
default_availability: 0
representation_scorer_fetch_signal_mute:
comment: "If enabled, fetches account mute signals from USS"
default_availability: 0
representation_scorer_fetch_signal_report:
comment: "If enabled, fetches tweet report signals from USS"
default_availability: 0
representation_scorer_fetch_signal_dont_like:
comment: "If enabled, fetches tweet don't like signals from USS"
default_availability: 0
representation_scorer_fetch_signal_see_fewer:
comment: "If enabled, fetches tweet see fewer signals from USS"
default_availability: 0
# To create a new decider, add here with the same format and caller's details : "representation-scorer_load_shed_by_caller_id_twtr:{{role}}:{{name}}:{{environment}}:{{cluster}}"
# All the deciders below are generated by this script - ./strato/bin/fed deciders ./ --service-role=representation-scorer --service-name=representation-scorer
# If you need to run the script and paste the output, add only the prod deciders here. Non-prod ones are being taken care of by representation-scorer_load_shed_non_prod_callers
"representation-scorer_load_shed_by_caller_id_all":
comment: "Reject all traffic from caller id: all"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:frigate:frigate-pushservice-canary:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:frigate:frigate-pushservice-canary:prod:atla"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:frigate:frigate-pushservice-canary:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:frigate:frigate-pushservice-canary:prod:pdxa"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:frigate:frigate-pushservice-send:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:frigate:frigate-pushservice-send:prod:atla"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:frigate:frigate-pushservice:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:frigate:frigate-pushservice:prod:atla"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:frigate:frigate-pushservice:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:frigate:frigate-pushservice:prod:pdxa"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:frigate:frigate-pushservice:staging:atla":
comment: "Reject all traffic from caller id: twtr:svc:frigate:frigate-pushservice:staging:atla"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:frigate:frigate-pushservice:staging:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:frigate:frigate-pushservice:staging:pdxa"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:home-scorer:home-scorer:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:home-scorer:home-scorer:prod:atla"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:home-scorer:home-scorer:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:home-scorer:home-scorer:prod:pdxa"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:stratostore:stratoapi:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:stratostore:stratoapi:prod:atla"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:stratostore:stratoserver:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:stratostore:stratoserver:prod:atla"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:stratostore:stratoserver:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:stratostore:stratoserver:prod:pdxa"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:timelinescorer:timelinescorer:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:timelinescorer:timelinescorer:prod:atla"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:timelinescorer:timelinescorer:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:timelinescorer:timelinescorer:prod:pdxa"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:topic-social-proof:topic-social-proof:prod:atla":
comment: "Reject all traffic from caller id: twtr:svc:topic-social-proof:topic-social-proof:prod:atla"
default_availability: 0
"representation-scorer_load_shed_by_caller_id_twtr:svc:topic-social-proof:topic-social-proof:prod:pdxa":
comment: "Reject all traffic from caller id: twtr:svc:topic-social-proof:topic-social-proof:prod:pdxa"
default_availability: 0
"enable_sim_clusters_embedding_store_timeouts":
comment: "If enabled, set a timeout on calls to the SimClustersEmbeddingStore"
default_availability: 10000
sim_clusters_embedding_store_timeout_value_millis:
comment: "The value of this decider defines the timeout (in milliseconds) to use on calls to the SimClustersEmbeddingStore, i.e. 1.50% is 150ms. Only applied if enable_sim_clusters_embedding_store_timeouts is true"
default_availability: 2000

165
representation-scorer/server/src/main/resources/logback.xml Normal file
View File

@ -0,0 +1,165 @@
<configuration>
<shutdownHook class="ch.qos.logback.core.hook.DelayingShutdownHook"/>
<!-- ===================================================== -->
<!-- Service Config -->
<!-- ===================================================== -->
<property name="DEFAULT_SERVICE_PATTERN"
value="%-16X{traceId} %-12X{clientId:--} %-16X{method} %-25logger{0} %msg"/>
<property name="DEFAULT_ACCESS_PATTERN"
value="%msg"/>
<!-- ===================================================== -->
<!-- Common Config -->
<!-- ===================================================== -->
<!-- JUL/JDK14 to Logback bridge -->
<contextListener class="ch.qos.logback.classic.jul.LevelChangePropagator">
<resetJUL>true</resetJUL>
</contextListener>
<!-- ====================================================================================== -->
<!-- NOTE: The following appenders use a simple TimeBasedRollingPolicy configuration. -->
<!-- You may want to consider using a more advanced SizeAndTimeBasedRollingPolicy. -->
<!-- See: https://logback.qos.ch/manual/appenders.html#SizeAndTimeBasedRollingPolicy -->
<!-- ====================================================================================== -->
<!-- Service Log (rollover daily, keep maximum of 21 days of gzip compressed logs) -->
<appender name="SERVICE" class="ch.qos.logback.core.rolling.RollingFileAppender">
<file>${log.service.output}</file>
<rollingPolicy class="ch.qos.logback.core.rolling.TimeBasedRollingPolicy">
<!-- daily rollover -->
<fileNamePattern>${log.service.output}.%d.gz</fileNamePattern>
<!-- the maximum total size of all the log files -->
<totalSizeCap>3GB</totalSizeCap>
<!-- keep maximum 21 days' worth of history -->
<maxHistory>21</maxHistory>
<cleanHistoryOnStart>true</cleanHistoryOnStart>
</rollingPolicy>
<encoder>
<pattern>%date %.-3level ${DEFAULT_SERVICE_PATTERN}%n</pattern>
</encoder>
</appender>
<!-- Access Log (rollover daily, keep maximum of 21 days of gzip compressed logs) -->
<appender name="ACCESS" class="ch.qos.logback.core.rolling.RollingFileAppender">
<file>${log.access.output}</file>
<rollingPolicy class="ch.qos.logback.core.rolling.TimeBasedRollingPolicy">
<!-- daily rollover -->
<fileNamePattern>${log.access.output}.%d.gz</fileNamePattern>
<!-- the maximum total size of all the log files -->
<totalSizeCap>100MB</totalSizeCap>
<!-- keep maximum 7 days' worth of history -->
<maxHistory>7</maxHistory>
<cleanHistoryOnStart>true</cleanHistoryOnStart>
</rollingPolicy>
<encoder>
<pattern>${DEFAULT_ACCESS_PATTERN}%n</pattern>
</encoder>
</appender>
<!--LogLens -->
<appender name="LOGLENS" class="com.twitter.loglens.logback.LoglensAppender">
<mdcAdditionalContext>true</mdcAdditionalContext>
<category>${log.lens.category}</category>
<index>${log.lens.index}</index>
<tag>${log.lens.tag}/service</tag>
<encoder>
<pattern>%msg</pattern>
</encoder>
</appender>
<!-- LogLens Access -->
<appender name="LOGLENS-ACCESS" class="com.twitter.loglens.logback.LoglensAppender">
<mdcAdditionalContext>true</mdcAdditionalContext>
<category>${log.lens.category}</category>
<index>${log.lens.index}</index>
<tag>${log.lens.tag}/access</tag>
<encoder>
<pattern>%msg</pattern>
</encoder>
</appender>
<!-- Pipeline Execution Logs -->
<appender name="ALLOW-LISTED-PIPELINE-EXECUTIONS" class="ch.qos.logback.core.rolling.RollingFileAppender">
<file>allow_listed_pipeline_executions.log</file>
<rollingPolicy class="ch.qos.logback.core.rolling.TimeBasedRollingPolicy">
<!-- daily rollover -->
<fileNamePattern>allow_listed_pipeline_executions.log.%d.gz</fileNamePattern>
<!-- the maximum total size of all the log files -->
<totalSizeCap>100MB</totalSizeCap>
<!-- keep maximum 7 days' worth of history -->
<maxHistory>7</maxHistory>
<cleanHistoryOnStart>true</cleanHistoryOnStart>
</rollingPolicy>
<encoder>
<pattern>%date %.-3level ${DEFAULT_SERVICE_PATTERN}%n</pattern>
</encoder>
</appender>
<!-- ===================================================== -->
<!-- Primary Async Appenders -->
<!-- ===================================================== -->
<property name="async_queue_size" value="${queue.size:-50000}"/>
<property name="async_max_flush_time" value="${max.flush.time:-0}"/>
<appender name="ASYNC-SERVICE" class="com.twitter.inject.logback.AsyncAppender">
<queueSize>${async_queue_size}</queueSize>
<maxFlushTime>${async_max_flush_time}</maxFlushTime>
<appender-ref ref="SERVICE"/>
</appender>
<appender name="ASYNC-ACCESS" class="com.twitter.inject.logback.AsyncAppender">
<queueSize>${async_queue_size}</queueSize>
<maxFlushTime>${async_max_flush_time}</maxFlushTime>
<appender-ref ref="ACCESS"/>
</appender>
<appender name="ASYNC-ALLOW-LISTED-PIPELINE-EXECUTIONS" class="com.twitter.inject.logback.AsyncAppender">
<queueSize>${async_queue_size}</queueSize>
<maxFlushTime>${async_max_flush_time}</maxFlushTime>
<appender-ref ref="ALLOW-LISTED-PIPELINE-EXECUTIONS"/>
</appender>
<appender name="ASYNC-LOGLENS" class="com.twitter.inject.logback.AsyncAppender">
<queueSize>${async_queue_size}</queueSize>
<maxFlushTime>${async_max_flush_time}</maxFlushTime>
<appender-ref ref="LOGLENS"/>
</appender>
<appender name="ASYNC-LOGLENS-ACCESS" class="com.twitter.inject.logback.AsyncAppender">
<queueSize>${async_queue_size}</queueSize>
<maxFlushTime>${async_max_flush_time}</maxFlushTime>
<appender-ref ref="LOGLENS-ACCESS"/>
</appender>
<!-- ===================================================== -->
<!-- Package Config -->
<!-- ===================================================== -->
<!-- Per-Package Config -->
<logger name="com.twitter" level="INHERITED"/>
<logger name="com.twitter.wilyns" level="INHERITED"/>
<logger name="com.twitter.configbus.client.file" level="INHERITED"/>
<logger name="com.twitter.finagle.mux" level="INHERITED"/>
<logger name="com.twitter.finagle.serverset2" level="INHERITED"/>
<logger name="com.twitter.logging.ScribeHandler" level="INHERITED"/>
<logger name="com.twitter.zookeeper.client.internal" level="INHERITED"/>
<!-- Root Config -->
<!-- For all logs except access logs, disable logging below log_level level by default. This can be overriden in the per-package loggers, and dynamically in the admin panel of individual instances. -->
<root level="${log_level:-INFO}">
<appender-ref ref="ASYNC-SERVICE"/>
<appender-ref ref="ASYNC-LOGLENS"/>
</root>
<!-- Access Logging -->
<!-- Access logs are turned off by default -->
<logger name="com.twitter.finatra.thrift.filters.AccessLoggingFilter" level="OFF" additivity="false">
<appender-ref ref="ASYNC-ACCESS"/>
<appender-ref ref="ASYNC-LOGLENS-ACCESS"/>
</logger>
</configuration>

13
representation-scorer/server/src/main/scala/com/twitter/representationscorer/BUILD Normal file
View File

@ -0,0 +1,13 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"finagle-internal/slo/src/main/scala/com/twitter/finagle/slo",
"finatra/inject/inject-thrift-client",
"representation-scorer/server/src/main/scala/com/twitter/representationscorer/columns",
"strato/src/main/scala/com/twitter/strato/fed",
"strato/src/main/scala/com/twitter/strato/fed/server",
"twitter-server-internal/src/main/scala",
],
)

38
representation-scorer/server/src/main/scala/com/twitter/representationscorer/RepresentationScorerFedServer.scala Normal file
View File

@ -0,0 +1,38 @@
package com.twitter.representationscorer
import com.google.inject.Module
import com.twitter.inject.thrift.modules.ThriftClientIdModule
import com.twitter.representationscorer.columns.ListScoreColumn
import com.twitter.representationscorer.columns.ScoreColumn
import com.twitter.representationscorer.columns.SimClustersRecentEngagementSimilarityColumn
import com.twitter.representationscorer.columns.SimClustersRecentEngagementSimilarityUserTweetEdgeColumn
import com.twitter.representationscorer.modules.CacheModule
import com.twitter.representationscorer.modules.EmbeddingStoreModule
import com.twitter.representationscorer.modules.RMSConfigModule
import com.twitter.representationscorer.modules.TimerModule
import com.twitter.representationscorer.twistlyfeatures.UserSignalServiceRecentEngagementsClientModule
import com.twitter.strato.fed._
import com.twitter.strato.fed.server._
object RepresentationScorerFedServerMain extends RepresentationScorerFedServer
trait RepresentationScorerFedServer extends StratoFedServer {
override def dest: String = "/s/representation-scorer/representation-scorer"
override val modules: Seq[Module] =
Seq(
CacheModule,
ThriftClientIdModule,
UserSignalServiceRecentEngagementsClientModule,
TimerModule,
RMSConfigModule,
EmbeddingStoreModule
)
override def columns: Seq[Class[_ <: StratoFed.Column]] =
Seq(
classOf[ListScoreColumn],
classOf[ScoreColumn],
classOf[SimClustersRecentEngagementSimilarityUserTweetEdgeColumn],
classOf[SimClustersRecentEngagementSimilarityColumn]
)
}

16
representation-scorer/server/src/main/scala/com/twitter/representationscorer/columns/BUILD Normal file
View File

@ -0,0 +1,16 @@
scala_library(
compiler_option_sets = ["fatal_warnings"],
platform = "java8",
tags = ["bazel-compatible"],
dependencies = [
"content-recommender/thrift/src/main/thrift:thrift-scala",
"finatra/inject/inject-core/src/main/scala",
"representation-scorer/server/src/main/scala/com/twitter/representationscorer/common",
"representation-scorer/server/src/main/scala/com/twitter/representationscorer/modules",
"representation-scorer/server/src/main/scala/com/twitter/representationscorer/scorestore",
"representation-scorer/server/src/main/scala/com/twitter/representationscorer/twistlyfeatures",
"representation-scorer/server/src/main/thrift:thrift-scala",
"strato/src/main/scala/com/twitter/strato/fed",
"strato/src/main/scala/com/twitter/strato/fed/server",
],
)

13
representation-scorer/server/src/main/scala/com/twitter/representationscorer/columns/Info.scala Normal file
View File

@ -0,0 +1,13 @@
package com.twitter.representationscorer.columns
import com.twitter.strato.config.{ContactInfo => StratoContactInfo}
object Info {
val contactInfo: StratoContactInfo = StratoContactInfo(
description = "Please contact Relevance Platform team for more details",
contactEmail = "no-reply@twitter.com",
ldapGroup = "representation-scorer-admins",
jiraProject = "JIRA",
links = Seq("http://go.twitter.biz/rsx-runbook")
)
}

116
representation-scorer/server/src/main/scala/com/twitter/representationscorer/columns/ListScoreColumn.scala Normal file
View File

@ -0,0 +1,116 @@
package com.twitter.representationscorer.columns
import com.twitter.representationscorer.thriftscala.ListScoreId
import com.twitter.representationscorer.thriftscala.ListScoreResponse
import com.twitter.representationscorer.scorestore.ScoreStore
import com.twitter.representationscorer.thriftscala.ScoreResult
import com.twitter.simclusters_v2.common.SimClustersEmbeddingId.LongInternalId
import com.twitter.simclusters_v2.common.SimClustersEmbeddingId.LongSimClustersEmbeddingId
import com.twitter.simclusters_v2.thriftscala.Score
import com.twitter.simclusters_v2.thriftscala.ScoreId
import com.twitter.simclusters_v2.thriftscala.ScoreInternalId
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingId
import com.twitter.simclusters_v2.thriftscala.SimClustersEmbeddingPairScoreId
import com.twitter.stitch
import com.twitter.stitch.Stitch
import com.twitter.strato.catalog.OpMetadata
import com.twitter.strato.config.ContactInfo
import com.twitter.strato.config.Policy
import com.twitter.strato.data.Conv
import com.twitter.strato.data.Description.PlainText
import com.twitter.strato.data.Lifecycle
import com.twitter.strato.fed._
import com.twitter.strato.thrift.ScroogeConv
import com.twitter.util.Future
import com.twitter.util.Return
import com.twitter.util.Throw
import javax.inject.Inject
class ListScoreColumn @Inject() (scoreStore: ScoreStore)
extends StratoFed.Column("recommendations/representation_scorer/listScore")
with StratoFed.Fetch.Stitch {
override val policy: Policy = Common.rsxReadPolicy
override type Key = ListScoreId
override type View = Unit
override type Value = ListScoreResponse
override val keyConv: Conv[Key] = ScroogeConv.fromStruct[ListScoreId]
override val viewConv: Conv[View] = Conv.ofType
override val valueConv: Conv[Value] = ScroogeConv.fromStruct[ListScoreResponse]
override val contactInfo: ContactInfo = Info.contactInfo
override val metadata: OpMetadata = OpMetadata(
lifecycle = Some(Lifecycle.Production),
description = Some(
PlainText(
"Scoring for multiple candidate entities against a single target entity"
))
)
override def fetch(key: Key, view: View): Stitch[Result[Value]] = {
val target = SimClustersEmbeddingId(
embeddingType = key.targetEmbeddingType,
modelVersion = key.modelVersion,
internalId = key.targetId
)
val scoreIds = key.candidateIds.map { candidateId =>
val candidate = SimClustersEmbeddingId(
embeddingType = key.candidateEmbeddingType,
modelVersion = key.modelVersion,
internalId = candidateId
)
ScoreId(
algorithm = key.algorithm,
internalId = ScoreInternalId.SimClustersEmbeddingPairScoreId(
SimClustersEmbeddingPairScoreId(target, candidate)
)
)
}
Stitch
.callFuture {
val (keys: Iterable[ScoreId], vals: Iterable[Future[Option[Score]]]) =
scoreStore.uniformScoringStore.multiGet(scoreIds.toSet).unzip
val results: Future[Iterable[Option[Score]]] = Future.collectToTry(vals.toSeq) map {
tryOptVals =>
tryOptVals map {
case Return(Some(v)) => Some(v)
case Return(None) => None
case Throw(_) => None
}
}
val scoreMap: Future[Map[Long, Double]] = results.map { scores =>
keys
.zip(scores).collect {
case (
ScoreId(
_,
ScoreInternalId.SimClustersEmbeddingPairScoreId(
SimClustersEmbeddingPairScoreId(
_,
LongSimClustersEmbeddingId(candidateId)))),
Some(score)) =>
(candidateId, score.score)
}.toMap
}
scoreMap
}
.map { (scores: Map[Long, Double]) =>
val orderedScores = key.candidateIds.collect {
case LongInternalId(id) => ScoreResult(scores.get(id))
case _ =>
// This will return None scores for candidates which don't have Long ids, but that's fine:
// at the moment we're only scoring for Tweets
ScoreResult(None)
}
found(ListScoreResponse(orderedScores))
}
.handle {
case stitch.NotFound => missing
}
}
}

Some files were not shown because too many files have changed in this diff Show More