the-algorithm/unified_user_actions/service/deploy/uua-client-event.aurora
twitter-team 617c8c787d Open-sourcing Unified User Actions
Unified User Action (UUA) is a centralized, real-time stream of user actions on Twitter, consumed by various product, ML, and marketing teams. UUA makes sure all internal teams consume the uniformed user actions data in an accurate and fast way.
2023-04-14 16:45:37 -05:00

175 lines
5.3 KiB
Plaintext

import os
import itertools
import subprocess
import math
SERVICE_NAME = 'uua-client-event'
CPU_NUM = 3
HEAP_SIZE = 3 * GB
RAM_SIZE = HEAP_SIZE + 1 * GB
# We make disk size larger than HEAP so that if we ever need to do a heap dump, it will fit on disk.
DISK_SIZE = HEAP_SIZE + 2 * GB
class Profile(Struct):
package = Default(String, SERVICE_NAME)
cmdline_flags = Default(String, '')
log_level = Default(String, 'INFO')
instances = Default(Integer, 1000)
kafka_bootstrap_servers = Default(String, '/s/kafka/client-events:kafka-tls')
kafka_bootstrap_servers_remote_dest = Default(String, '/s/kafka/bluebird-1:kafka-tls')
source_topic = Default(String, 'client_event')
sink_topics = Default(String, 'unified_user_actions,unified_user_actions_engagements')
decider_overlay = Default(String, '')
resources = Resources(
cpu = CPU_NUM,
ram = RAM_SIZE,
disk = DISK_SIZE
)
install = Packer.install(
name = '{{profile.package}}',
version = Workflows.package_version()
)
async_profiler_install = Packer.install(
name = 'async-profiler',
role = 'csl-perf',
version = 'latest'
)
setup_jaas_config = Process(
name = 'setup_jaas_config',
cmdline = '''
mkdir -p jaas_config
echo "KafkaClient {
com.sun.security.auth.module.Krb5LoginModule required
principal=\\"discode@TWITTER.BIZ\\"
useKeyTab=true
storeKey=true
keyTab=\\"/var/lib/tss/keys/fluffy/keytabs/client/discode.keytab\\"
doNotPrompt=true;
};" >> jaas_config/jaas.conf
'''
)
main = JVMProcess(
name = SERVICE_NAME,
jvm = Java11(
heap = HEAP_SIZE,
extra_jvm_flags =
'-Djava.net.preferIPv4Stack=true'
' -XX:MaxMetaspaceSize=536870912'
' -XX:+UseNUMA'
' -XX:+AggressiveOpts'
' -XX:+PerfDisableSharedMem' # http://www.evanjones.ca/jvm-mmap-pause.html
' -Dlog_level={{profile.log_level}}'
' -Dlog.access.output=access.log'
' -Dlog.service.output={{name}}.log'
' -Djava.security.auth.login.config=jaas_config/jaas.conf'
),
arguments =
'-jar {{name}}-bin.jar'
' -admin.port=:{{thermos.ports[health]}}'
' -kafka.bootstrap.servers={{profile.kafka_bootstrap_servers}}'
' -kafka.bootstrap.servers.remote.dest={{profile.kafka_bootstrap_servers_remote_dest}}'
' -kafka.group.id={{name}}-{{environment}}'
' -kafka.producer.client.id={{name}}-{{environment}}'
' -kafka.max.pending.requests=10000'
# CE events is about 0.4-0.6kb per message on the consumer side. A fetch size of 6~18 MB get us
# about 10k ~ 20k of messages per batch. This fits the size of our pending requests queue and
# within the limit of the max poll records.
' -kafka.consumer.fetch.max=9.megabytes'
' -kafka.consumer.fetch.min=3.megabytes'
' -kafka.max.poll.records=40000'
' -kafka.commit.interval=20.seconds'
' -kafka.producer.batch.size=4.megabytes'
' -kafka.producer.buffer.mem=64.megabytes'
' -kafka.producer.linger=100.millisecond'
' -kafka.producer.request.timeout=30.seconds'
' -kafka.producer.compression.type=lz4'
' -kafka.worker.threads=4'
' -kafka.source.topic={{profile.source_topic}}'
' -kafka.sink.topics={{profile.sink_topics}}'
' -decider.base=decider.yml'
' -decider.overlay={{profile.decider_overlay}}'
' -cluster={{cluster}}'
' {{profile.cmdline_flags}}',
resources = resources
)
stats = Stats(
library = 'metrics',
port = 'admin'
)
job_template = Service(
name = SERVICE_NAME,
role = 'discode',
instances = '{{profile.instances}}',
contact = 'disco-data-eng@twitter.com',
constraints = {'rack': 'limit:1', 'host': 'limit:1'},
announce = Announcer(
primary_port = 'health',
portmap = {'aurora': 'health', 'admin': 'health'}
),
task = Task(
resources = resources,
name = SERVICE_NAME,
processes = [async_profiler_install, install, setup_jaas_config, main, stats],
constraints = order(async_profiler_install, install, setup_jaas_config, main)
),
health_check_config = HealthCheckConfig(
initial_interval_secs = 100,
interval_secs = 60,
timeout_secs = 60,
max_consecutive_failures = 4
),
update_config = UpdateConfig(
batch_size = 1000,
watch_secs = 90,
max_per_shard_failures = 3,
max_total_failures = 0,
rollback_on_failure = False
)
)
PRODUCTION = Profile(
# go/uua-decider
decider_overlay = '/usr/local/config/overlays/discode-default/UnifiedUserActions/prod/{{cluster}}/decider_overlay.yml'
)
STAGING = Profile(
package = SERVICE_NAME+'-staging',
cmdline_flags = '',
kafka_bootstrap_servers_remote_dest = '/s/kafka/custdevel:kafka-tls',
decider_overlay = '/usr/local/config/overlays/discode-default/UnifiedUserActions/staging/{{cluster}}/decider_overlay.yml' # go/uua-decider
)
DEVEL = STAGING(
log_level = 'INFO',
)
prod_job = job_template(
tier = 'preferred',
environment = 'prod',
).bind(profile = PRODUCTION)
staging_job = job_template(
environment = 'staging'
).bind(profile = STAGING)
devel_job = job_template(
environment = 'devel'
).bind(profile = DEVEL)
jobs = []
for cluster in ['atla', 'pdxa']:
jobs.append(prod_job(cluster = cluster))
jobs.append(staging_job(cluster = cluster))
jobs.append(devel_job(cluster = cluster))