Lambda architecture @ Indix
7 likes2,715 views
The document discusses the Lambda architecture for building big data systems, focusing on integrating batch and real-time processing to handle large-scale data. It highlights challenges such as mutable state, operational complexity, and high latency while proposing solutions for human fault tolerance and accuracy. The architecture is described as domain agnostic and emphasizes the need for higher-level abstractions and tools that are still maturing.
Lambda architecture @ Indix
- 1. Lambda Architecture Analyzing large scale, unstructured, dynamic data Rajesh Muppalla (@codingnirvana) rajesh@indix.com
- 2. Indix - Quick Overview Am I priced higher or lower w.r.t my competitor on Nikon D700? Which product has the UPC - 8745354434? What are all the variants of Apple Macbook Air 13â? What is the average price change of all Nike Shoes in Walmart in the last 3 months?
- 3. Data Pipeline @ Indix C Crawling Parsing ML Model ML Model Classification C1 C1 C1 C1 C2 C2 C2 C2 C2 Matching Product & Price Catalog
- 4. Data Pipeline @ Indix Analytics (Precomputes, Insights) Search Index Product & Price Catalog Experiences We released the v1.0 of our API today - developer.indix.com
- 5. Data is Dynamic C C1 C1 C1 C1 C2 C2 C2 C2 C2 ML Model ML Model (new) Crawling Parsing Classification Matching
- 6. Data Scale 400 M Product URLs 4 TB HTML Data Crawled Daily 100 TB Data Processed Daily 3000 Categories 10 B Price Points 2000 Sites
- 8. Batch using HBase & MapReduce
- 9. Problem 1 Mutable State Data Systems should be Human Fault Tolerant
- 10. Problem 2 Compactions Random Write databases are hard to manage at large scale
- 11. Problem 3 16 hours 16 hours latency is a lot. We wanted it to be couple of hours
- 12. Three Problems â No Human Fault Tolerance â Mutable State â Operational Complexity â Random Writes (Compactions) â High Latency â Batch system architectural tradeoff
- 13. Rethink our data systems
- 15. Lambda Architecture â An approach to build big data systems â Architectural Components & Principles â Ties Batch & Real Time Systems â General Purpose - Domain Agnostic â Coined by Nathan Marz â Ex-Twitter Engineer â Creator of Storm
- 16. Data System - Traditional Approach HBase Application Source of Truth
- 17. Data System - New Approach Immutable Raw Data Application Processed View(s) Source of Truth
- 18. Letâs take an example Find the count of unique products in any given category for the entire time range
- 19. Two Requirements â Recomputations â Large Scale
- 20. Batch Layer Implementation HDFS (Vertical Partitioning) HBase C1 5 C2 7 C3 4 C4 7 C5 1 Products Master Data 9 am 10 am 11 am 12 pm 1 pm 2 pm Query Intermediate view C1 C2 C3 C4 C5 MR Job 1 Batch View New Data MR Job 2
- 21. Handling Recomputations HDFS (Vertical Partitioning) HBase C1 5 C2 7 C3 4 C4 7 C5 1 Products Master Data 9 am 10 am 11 am 12 pm 1 pm 2 pm Query Intermediate view C1 C2 C3 C4 C5 MR Job 1 Batch View New Data MR Job 2
- 22. Handling Scale â Hadoop HDFS, MapReduce, HBase â Proven Linear Scalability
- 23. Three Problems (Recap) â No Human Fault Tolerance â Mutable State â Operational Complexity â Random Writes (Compactions) â High Latency â Batch system architectural tradeoff
- 24. Human Fault Tolerance â Bugs in the batch jobs â Discard views & Recompute â Bugs in the master data jobs â Re-process the master data to hide the old data â Bugs in the query â Re-deploy the query layer â Traceability as a side effect
- 25. Operational Complexity â No random writes in the batch layer â Bulk Updates to build the batch view
- 26. Great⊠What about Latency?
- 27. Speed Layer Queue (Kafka) Recent Data Real Time Processing (Storm) HHyyppeHerylroplogeglrolloogg gS lSeoetgst s Query Random Writes (Updates) Read-Write Data Store (Riak, HBase, Cassandra)
- 28. Speed Layer has mutation... But â Speed layer deals with much smaller data â Batch Layer - Months/years of data â Speed Layer - Few hours or 1 day of data â Easy to manage operationally Complexity Isolation
- 29. Final Step - Merging Results Batch Layer Speed Layer Data Query Merged Results C1 - 50000 C1 - 499 (Approximate with error 0.02%) C1 - 50499
- 30. What about Accuracy? Batch Layer Speed Layer Data Query Merged Results C1 - 499 (Approximate with error 0.02%) C1â - 50500 Batch Layer CC11â -- 5500050000 Eventually Accurate
- 32. Lambda Architecture @ INDIX
- 33. Lambda Architecture @ Indix
- 34. Batch Layer @ Indix â Pail â Vertical partitioning â Consolidation of small files â Scalding â Thrift for enforcing schemas â HBase/Solr for views â Bulk updates to create views
- 35. Speed Layer @ Indix â Still WIP â To reduce latency â Micro batches for Speed layer â Use the last batch run + bulk update views
- 36. Open Challenges â Managing both Batch & Real Time still painful â Two broad directions â Abstractions â SummingBird (Twitter) â Unified Stack â Spark â Kafka + Samza/Storm (LinkedIn) â Cloud Data Flow (Google)
- 37. In Conclusion... â Lambda Architecture â A different approach to build data systems â Solid principles â Domain Agnostic â Tools not yet mature
- 38. Resources â Indix Engineering Blog - http://engineering.indix.com â Runaway Complexity in Big Data Systems â Lambda Architecture â Big Data Book - Manning â Scalding â Spark â Pail â Summingbird
- 39. Key Takeaways - Human Fault Tolerance - Complexity Isolation - Higher Level Abstractions
- 40. Thank You
- 41. Batch vs Real Time Choices
- 42. Tying it all together - Go-CD
- 43. Extras â Monoids â LA is not new â Search Engines (fast, slow crawl) â Event Sourcing (immutable events to maintain state) â Patch, Audit, Bootstrap
- 44. Problem Statement - Optimization