Grokking TechTalk #29: Building Realtime Metrics Platform at LinkedIn
- 1. Building Realtime Metrics Platform at LinkedIn Khai Tran Staff Software Engineer
- 2. Agenda ● Self introduction ● Data organization at LinkedIn ● Overview of LinkedIn metrics platform ● Moving from offline to nearline ● Under the hood of the nearline architecture ● Nearline production use cases ● Conclusion
- 3. About me ● Currently teach lead of LinkedIn metrics team ● BS from DHBK HN, K46 ● PhD from University of Wisconsin- Madison, on databases. Thesis: “Realizing parallelism in OLTP workloads” ● Interns at Oracle, Microsoft Research, and Google ● First job - Oracle Labs, mostly working on Oracle query optimizer ● Second job - AWS, in two teams: ○ DynamoDB: on Request Routers, DynamoDB query language ○ Redshift: on query processing, statistic collection
- 4. Data organization at LinkedIn ● Data infrastructure - online world ○ Operational databases: Espresso, Venice/Voldermort, graph database… ○ Streaming: Samza, Kafka, Brooklin/Databus ○ Search infrastructure... ● Analytic Platform and Application - offline world ○ Infrastructure: Hadoop eco, Spark, Presto, Goblin (for ETL), Pinot (OLAP databases)... ○ Platform: metrics platform, data warehouses ○ Applications: XLNT (A/B testing), Raptor (visualization), ThirdEye (anomaly detection)... ● Relevance - machine learning related ○ Machine learning infrastructure and platform ○ Feed ranking, search ranking ○ PYMK, JYMI...
- 5. Overview of LinkedIn metrics platform
- 6. Metrics @ LinkedIn ● Metrics = Measurements over tracking data ● Tracking data: any logged events (web or mobile) ● Crucial for decision making: ○ Experimentation - test everything ○ Reporting - monitor and alert ○ In production, site-facing applications
- 7. Example
- 8. We provide: ● A trusted repository of metrics ● A self-serve platform for sustainable lifecycle of metrics In production Experimentation Reporting Primary Data Unified Metrics Platform LinkedIn unified metrics platform (UMP)
- 9. Growth of UMP Metrics 2016 20172015 6800 4680 1100 Current: 10K+ metrics
- 10. # code LOAD … # data # transformation # code STORE … # config Metrics: - A = SUM(A’) - B = Unique(id) Downstream: - XLNT - Raptor UMP User Code Platform Generated Code To App To App DefineDeclare Onboard Data Metadata Onboarding process User
- 11. Moving from offline to nearline
- 12. Offline computation flows Hourly job latency: 3-6 hours -> want realtime/nearline ...... Metric union User code User code Cubing, Rollup Dimension augmentation HDFS tables Dali views Pinot, Presto Azkaban execution Espresso, Oracle, MySQL
- 13. ... What we want for nearline flows Metric unionUser code User code Samza job Dimension augmentation Pinot
- 14. Latency is not the only requirement Easy to onboard ● Minimum effort to convert existing offline into nearline ● Easy to write user code for new nearline flows Easy to maintain ● Just one version of user code - single source of truth ● Run as a service Latency ● ~5 - 30 mins
- 15. Samza jobs Putting things together Pinot Batch jobs UMP realtime platform UMP offline platform HDFS Raptor Lambda architecture with a single codebase code configMetrics definition
- 16. Current support User code in Pig ● LOAD, STORE ● FILTER, SAMPLE, SPLIT, UNION ● Simple FOREACH ● JOIN - all semantics ● GROUP/COGROUP, DISTINCT ● Record/Array FLATTEN ● Java UDFs, Python UDFs ● Pig Nested FOREACH and sort/limit (in Windows) ● Hive Not yet
- 17. Under the hood of the nearline architecture
- 18. Pig to Samza through SQL processing Open source framework for building dynamic data management systems. Including: ➢ SQL Parser ➢ Relational algebra APIs ➢ Query planning engine We built UMP nearline with: ➢ Pig’s Grunt parser ➢ Calcite relational algebra ➢ Calcite query planning engine
- 19. Architecture ... Metric union User code User code Dimension augmentation Calcite relational algebra as an IR convert generate Samza code optimize Samza physical plan Samza configuration Pig to Calcite Calcite to Samza
- 20. Pig to Calcite # code LOAD … LOAD ... COGROUP ... STORE … GruntParser CO- GROUP LOAD LOAD PigRelConverter FULL OUTERJ OIN AGGRE GATE AGGRE GATE TABLE SCAN TABLE SCAN PRO- JECT User scripts Pig Logical Plan Calcite relational algebra
- 21. Example
- 22. Example
- 23. Example INNER JOIN FILTER FILTER PROJECT PROJECT PROJECT TABLE SCAN TABLE SCAN Calcite logical plan
- 24. Planning/Optimization ➢ Calcite logical plans: ○ Relational algebra: What to do ➢ Samza physical plans: ○ Samza physical node: How to do it ➢ Calcite Samza planner: ○ Calcite logical plan -> optimized Samza physical plan
- 25. Example Stream- Stream Self Join Samza Project Samza Project Samza Filter Samza Filter Samza Project Input Stream INNER JOIN FILTER FILTER PROJECT PROJECT PROJECT TABLE SCAN TABLE SCAN Calcite Samza planner Calcite logical plan Samza physical plan
- 26. Code-gen From Samza physical plans: ➢ Generate Samza code for constructing the stream graph using Samza Fluent APIs . Mapping: ➢ Samza physical nodes -> corresponding stream APIs: ○ Samza project -> stream.map() ○ Samza filter -> stream.filter() ○ ... ➢ Relational expressions -> lambda functions: ○ Filter expressions -> filter() functions ○ Project expressions -> map() functions ○ ...
- 27. Example Stream- Stream Self Join Samza Project 1 Samza Project 2 Samza Filter 1 Samza Filter 2 Samza Project Input Stream
- 28. Example Stream- Stream Self Join Samza Project 1 Samza Project 2 Samza Filter 1 Samza Filter 2 Samza Project Input Stream
- 29. Example Stream- Stream Self Join Samza Project 1 Samza Project 2 Samza Filter 1 Samza Filter 2 Samza Project Input Stream
- 30. Example Stream- Stream Self Join Samza Project 1 Samza Project 2 Samza Filter 1 Samza Filter 2 Samza Project Input Stream
- 31. Example Stream- Stream Self Join Samza Project 1 Samza Project 2 Samza Filter 1 Samza Filter 2 Samza Project Input Stream
- 32. Example Stream- Stream Self Join Samza Project 1 Samza Project 2 Samza Filter 1 Samza Filter 2 Samza Project Input Stream
- 34. Nearline production use cases
- 35. Top stories picked up by editors
- 36. Usage case 1 - Feedback to editor
- 37. Usage case 2 - Recruiter usage statistics
- 38. Conclusion
- 39. Samza jobs From improved Lambda architecture... Pinot Batch jobs UMP realtime platform UMP offline platform HDFS Raptor Lambda architecture with a single codebase code configMetrics definition
- 40. … to our bigger picture Pig Latin Calcite relational algebra HiveQL SparkSQL/ RDD Presto SQL Portable UDFs AORA (Author Once, Run Anywhere) architecture