June 2014 HUG: Interactive analytics over hadoop
3 likes1,429 views
This document discusses interactive analytics for human timescales using feature sequences to calculate non-additive metrics like instant overlaps between large user groups. It describes Yahoo's advertising data warehouse that handles petabytes of data daily and provides normalized views and analytics across systems in milliseconds. Custom algorithms like feature sequence encoding enable exact overlap calculations in under a minute for billions of user events, compared to 19 hours for existing approaches.
June 2014 HUG: Interactive analytics over hadoop
- 1. Interactive Analytics in Human Time S u p r e e t h R a o , S u n i l G u p t a ⎪ J u n e 1 8 , 2 0 1 4 4 5 t h B a y A r e a H U G , S u n n yv a l e , C a l i f o r n i a
- 2. Interactive – How we see it? 2 Yahoo Confidential & Proprietary 60B events, 3.5TB of compressed data Response 400ms Serve an ad and get insights < 2s
- 3. Agenda:
- 4. Motivation Approach Problem Deepdive- Instant Overlap Summary Questions
- 5. Motivation:
- 6. Lots of data Analytics Data restatement - batch and real time Human time
- 7. Lots of data ~30B advertising events/day ~10s of TB of compressed data/day Minutes to Year Grain Multi-quarter data retention Data Aging
- 8. Analytics Reporting Metrics Attribution Multi-level hierarchical computation Bidding/Targeting optimization Non-additive computation
- 9. Data Restatement Real time Batch Producer Consumer quick path, lower amount of checks or reconciliation, typically no lookups high latency path, checks and reconciliations, can have lookbacks and lookups
- 10. Human Time <1s ( 99 percentile) Default time grain ( < 300 ms) Instant overlap ( < 60s) Data ingested, insights available ( < 2s)
- 11. Lots of data Analytics Data restatement - batch and real time Human time
- 12. Approach:
- 13. Data Ingestion or Collection Transformations Data Ingestion Persistence Runtime Compute Caching API Optional Middleware Business API UI Data Collection Data Pipelines Data Warehouse/ Analytics and Optimizations Reporting Application/UI Logical View - Scope Transformations/Aggs Data Ingestion Persistence Runtime Compute Caching API Optional Middleware Business API UI Data Collection Impacts Out of scope
- 14. Transformations/Aggs Data Ingestion Persistence Runtime Compute Caching API Optional Middleware Business API UI Data Collection Batch processing DAG, Real-time topology, SOX, Traffic protection, Late processing, Retention, Completeness Monitoring, PII cleansing/masking Compatible with HDFS, Performance (Indexed, Columnar, Compression, Serialization, Flexibility, Concurrency, Grain of data stored) Distributed/Stand-alone, Caching objects vs caching results Access to data with group by, order by etc..; SQL or SQL like Translate JSON to SQL(optional) Logical View - Characteristics Impacts Out of scope
- 15. Transformations/Aggs Data Ingestion Persistence Runtime Compute Caching API Optional Middleware Business API UI Data Collection Hadoop MR/PIG /Oozie(Lotus)/Storm(Trident) Druid, Shark, Hive, Oracle RAC, Mysql, Hbase, Impala memcached_y, Redis JSON-REST API ; JDBC; ODBC Data Ingestion Persistence Runtime Compute Caching API Optional Middleware Business API UI Data Collection Logical View - Choices Impacts Out of scope
- 16. How we do what we do? Components of Advertising Data Warehouse Druid JDBC/ODBC Data Warehouse-Persistence Hive Metrics Store JSON-API Persistence and run-time compute Computation and Ingestion Quick cache ( using a database for now) Upstream: API layer, MSTR, Adhoc access, Identity Service, Ad-Serving manifests Data Producers; Serving, Scoring, Booking, 3rd/1st Party Data Real time and batch compute engine (Hadoop/Storm ) Data filtering/transformations: Transformations, format conversions Custom Algorithms : computing recursive uniques, indexing
- 17. Human time, How? Druid for interactive queries Storm-Druid for quick ingestion and index Specialized computation and processing for quicker response › Sketches › Feature sequence based overlaps › Custom indexing
- 18. Problem Deepdive: Instant Overlap
- 21. Overlap Non-additive › Require access to raw (user level data) to compute non-additive • Billions of events a day • TBs of data a day 1-1 vs 1-n vs few-n › Between car commuter and vegan what is the overlap › For Car commuter which are the top overlap groups › For Vegan, Car commuters what are the top overlap groups
- 22. Re-stating motivation Given two sets having identifiers, how can we do exact overlaps in close to real time? ( < 1 min). Overlap is like a AND operation or a set
- 23. Existing Approaches ● Use exact compute paradigms o Do joins for intersections which will lead to exact results Hive, PIG, MR can all support efficient joins Exact but not real time ● Use sketches o Approximate algorithms HLL, KMV, accuracy vs size, performance Approx, needs high perf tuning close to real time but not exact
- 24. Using Feature Sequences – 1/4 Feature sequence encoding o Encode the sequence {Ram} - { car commuter, soccer fan,...} {Tom} - { soccer fan, vegan...} {Sam} - { car commuter, soccer fan, vegan...} ….
- 25. Using Feature Sequences – 2/4 Eliminate the user on encoded bitmaps {car commuter, soccer fan, vegan...}- count -c1- # {soccer fan, vegan...} - count - c2 - # {car commuter, vegan...} - count - c2 - # Counts become additive now
- 26. Using Feature Sequences – 3/4 ● Store row qualifications into a bitmap o Car commuter- Row1, Row3 1010000000 o Vegan - Row1, Row2, Row3 1110000000 ● Load the bitmap into Druid using a custom indexer o in-memory or memory mapped
- 27. Using Feature Sequences – 4/4 Data Structures › {feature_sequence}->count › Feature->row qualification bitmaps AND is now an “AND” on bitmaps › supported within Druid › Very efficient Works alongside topN and groupBys
- 28. Comparison with existing algorithm ● 1-n – Bulk Overlap on grid o 19 hours on grid o Few-n calls for a re-process o 1-1 ( <1s) ● Instant Overlap o < 60s ( pre-processing 3-4 hours) o Supports “exact” AND o Flexible ( few-n, 1-n) o 1-1 ( < 1s)
- 29. Summary ● Yahoo’s Advertising Data Warehouse o Peta Byte Scale o Normalized view across many systems o Analytics and optimizations with specialized algorithms o Data restatement - batch and realtime o Human time
- 30. Thank You @supreeth_ @_skgupta We are hiring! Reach out to us at bigdata@yahoo-inc.com.
- 31. Data Ingestion or Collection Transformations Data Ingestion Persistence Runtime Compute Caching API Optional Middleware Business API UI Data Collection Data Pipelines Data Warehouse/ Analytics and Optimizations Reporting Application/UI Logical View - Scope Transformations/Aggs Data Ingestion Persistence Runtime Compute Caching API Optional Middleware Business API UI Data Collection Impacts Out of scope
- 32. Dimension Flexibility Many dimensions Adding new dimensions Time zones Time grain
- 33. Normalized view across systems PaidSearch Display Native Programmatic buying and selling Ad-targeting
- 34. Hardware Configs ●High-memory boxes ●SSD preferred ●Savings due to better compression
Editor's Notes
- #14: Logical view of a typical data driven application architecture Sox compliance
- #17: -quick cache store for querying all metrics in a single fetch, to support one-page load UI architecture - Hive for scheduled job and for adhoc long range generic queries which are not supported on the interactive interface
- #18: Bitmap indexed, columnar, can operate on compressed bitmaps, distributed -