Upload
yi-pan
View
798
Download
1
Embed Size (px)
Citation preview
Stream Processing @Scale in LinkedIn
Yi PanData Infrastructure
Samza Team @LinkedIn
Databus
• What is Stream Processing?• What is Samza?• Stream Processing @LinkedIn• Upcoming features
Overview
• What’s stream processing– Input: an unbounded sequence of events
• E.g. web server logs, user activity tracking events, database changelogs, etc.
– Latency: near real-time• From milliseconds to minutes, instead of hours to
days– Output: an unbounded sequence of changes to
the derived dataset• The derived dataset is usually the final or partial
analytic results that can either be in another stream, or a serving data store
Stream Processing
Response latency
Stream
Processing
Milliseconds to minutes
RPC
Synchronous Later. Possibly much later.
0 ms
Stream Processing
• What are the application requirements?– Scalable, fast, stateful stream processing– What scale should we operate at?
• Traffic Volume: 1.4 Trillion events/day• Intermediate State Size: multi TB / colo (*)
– Why is it expensive to run stream processing at scale?
• Intermediate data set needs to be stored to allow low latency processing
• Large volume of data needs to be pulled and pushed via network
Stream Processing
• What is Stream Processing?• What is Samza?• Stream Processing @LinkedIn• Upcoming features
Overview
• Samza is a distributed Turing machine– Single Task Samza Job is a stateful
Turing machine
What’s Samza
Samza TaskInput stream Output stream
Statechangelogch
eckp
oint
– Scaling a Samza job: partition the streams
What’s SamzaIn
put s
trea
m A
partition 0
partition 1
partition 2
partition 3
partition n
– Scaling a Samza job: partition the streams
What’s SamzaIn
put s
trea
m B
partition 0
partition 1
partition 2
partition 3
partition n
– Scaling a Samza job: replicating the state machine
What’s Samza
shared checkpoint
Job
• Samza Execution in Yarn
What’s Samza
Host 1 Host 2 Host 3
Application Master
Samza container Samza container
Samza container
Deploy Samza job
• States in Samza– Checkpoints
• Offsets per input stream partitions– State Stores
• In-memory or on-disk (RocksDB) derived data set
What’s Samza
Samza TaskOutput stream partitions
State changelog partitionsch
eckp
oint
Host 1
• States in Samza– Checkpoints and local state stores are backed
by distributed logs
What’s Samza
Samza TaskOutput stream partitions
State changelog partitionsch
eckp
oint
Host 1
• What is Stream Processing?• What is Samza?• Stream Processing @LinkedIn• Upcoming features
Overview
Stream Processing @ LinkedIn
WebServersWebServers
WebServersWebServers
WebServersWebServers
WebServersMonitorServers
Oracle
Espresso
Kafka Databus
Trackingevents
Metrics
changelog
changelog
Samza JobsSamza
JobsSamza JobsSamza
Jobs
bootstrap
bootstrap
VoldemortDerivedData
DerivedData
Stream Processing @ LinkedIn
• Tracking aggregate/analysis (ACG)
Stream Processing @ LinkedIn
• Content standardization w/ adjunct data setMember
Profile DBBootstrap
JobDatabus
Kafka
Content Standardization
Kafka
Kafka
Stream Processing @ LinkedIn
• Kafka Deployment– 1.1 Trillion messages / day
• Databus Deployment– 300 Billion messages / day
• Samza Deployment– multiple colos– 10+ Yarn clusters– 200+ nodes– 100+ Jobs in production
• What is Stream Processing?• What’s Samza• Stream Processing @LinkedIn• Upcoming features
Overview
• New features– Local state store improvements
• RocksDB TTL support• Fast recovery
– Dynamic configuration– Easier deployment w/ standalone jobs– High-level query language for faster
development
Upcoming Features
Contact Us / Get Involved• Open Source
–Documentation: samza.apache.org–Mailing list:
[email protected]– JIRA: https
://issues.apache.org/jira/browse/SAMZA