Open Source Lambda Architecture with Hadoop, Kafka, Samza and Druid
44 likesā¢6,895 views
This document discusses using an open source Lambda architecture with Kafka, Hadoop, Samza, and Druid to handle event data streams. It describes the problem of interactively exploring large volumes of time series data. It outlines how Druid was developed as a fast query layer for Hadoop to enable low-latency queries over aggregated data. The architecture ingests raw data streams in real-time via Kafka and Samza, aggregates the data in Druid, and enables reprocessing via Hadoop for reliability.
Open Source Lambda Architecture with Hadoop, Kafka, Samza and Druid
- 1. OPEN SOURCE LAMBDA ARCHITECTURE KAFKA Ā· HADOOP Ā· SAMZA Ā· DRUID FANGJIN YANG Ā· GIAN MERLINO Ā· DRUID COMMITTERS
- 2. PROBLEM DEALING WITH EVENT DATA MOTIVATION EVOLUTION OF A āREAL-TIMEā STACK ARCHITECTURE THE āRADā-STACK NEXT STEPS TRY IT OUT FOR YOURSELF OVERVIEW
- 3. THE PROBLEM
- 4. 2013 THE PROBLEM ā£ Arbitrary and interactive exploration of time series data ā¢ Ad-tech, system/app metrics, network/website trafļ¬c analysis ā£ Multi-tenancy: lots of concurrent users ā£ Scalability: 10+ TB/day, ad-hoc queries on trillions of events ā£ Recency matters! Real-time analysis
- 5. 2013 FINDING A SOLUTION ā£ Load all your data into Hadoop. Query it. Done! ā£ Good job guys, letās go home
- 7. 2013 PROBLEMS WITH THE NAIVE SOLUTION ā£ MapReduce can handle almost every distributed computing problem ā£ MapReduce over your raw data is ļ¬exible but slow ā£ Hadoop is not optimized for query latency ā£ To optimize queries, we need a query layer
- 8. 2013 FINDING A SOLUTION Hadoop (pre-processing and storage) Query Layer Hadoop EventStreams Insight
- 9. A FASTER QUERY LAYER
- 10. 2013 MAKE QUERIES FASTER ā£ What types of queries to optimize for? ā¢ Revenue over time broken down by demographic ā¢ Top publishers by clicks over the last month ā¢ Number of unique visitors broken down by any dimension ā¢ Not dumping the entire dataset ā¢ Not examining individual events
- 11. 2013 FINDING A SOLUTION Hadoop (pre-processing and storage) RDBMS Hadoop EventStreams Insight
- 12. 2013 FINDING A SOLUTION Hadoop (pre-processing and storage) NoSQL K/V Stores Hadoop EventStreams Insight
- 13. 2013 FINDING A SOLUTION Hadoop (pre-processing and storage) Commercial Databases Hadoop EventStreams Insight
- 14. DRUID AS A QUERY LAYER
- 15. 2013 DRUID ā£ Druid project started in 2011, went open source in 2012 ā£ Designed for low latency ingestion and ad-hoc aggregations ā£ Designed for keeping around a lot of history (years are ok) ā£ Growing Community ā¢ ~100 contributors ā¢ Used in production at numerous large and small organizations
- 16. 2014 REALTIME INGESTION >500K EVENTS / SECOND AVERAGE >1M EVENTS / SECOND PEAK 10 ā 100K EVENTS / SECOND / CORE DRUID IN PRODUCTION
- 17. 2014 0.0 0.5 1.0 1.5 0 1 2 3 4 0 5 10 15 20 90%ile95%ile99%ile Feb 03 Feb 10 Feb 17 Feb 24 time querytime(seconds) datasource a b c d e f g h Query latency percentiles QUERY LATENCY (500MS AVERAGE) 90% < 1S 95% < 5S 99% < 10S DRUID IN PRODUCTION
- 18. 2013 RAW DATA timestamp publisher advertiser gender country click price 2011-01-01T01:01:35Z bieberfever.com google.com Male USA 0 0.65 2011-01-01T01:03:63Z bieberfever.com google.com Male USA 0 0.62 2011-01-01T01:04:51Z bieberfever.com google.com Male USA 1 0.45 ... 2011-01-01T01:00:00Z ultratrimfast.com google.com Female UK 0 0.87 2011-01-01T02:00:00Z ultratrimfast.com google.com Female UK 0 0.99 2011-01-01T02:00:00Z ultratrimfast.com google.com Female UK 1 1.53
- 19. 2013 ROLLUP DATA timestamp publisher advertiser gender country impressions clicks revenue 2011-01-01T01:00:00Z ultratrimfast.com google.com Male USA 1800 25 15.70 2011-01-01T01:00:00Z bieberfever.com google.com Male USA 2912 42 29.18 2011-01-01T02:00:00Z ultratrimfast.com google.com Male UK 1953 17 17.31 2011-01-01T02:00:00Z bieberfever.com google.com Male UK 3194 170 34.01 ā£ Truncate timestamps ā£ GroupBy over string columns (dimensions) ā£ Aggregate numeric columns (metrics)
- 20. 2013 PARTITION DATA timestamp publisher advertiser gender country impressions clicks revenue 2011-01-01T01:00:00Z ultratrimfast.com google.com Male USA 1800 25 15.70 2011-01-01T01:00:00Z bieberfever.com google.com Male USA 2912 42 29.18 2011-01-01T02:00:00Z ultratrimfast.com google.com Male UK 1953 17 17.31 2011-01-01T02:00:00Z bieberfever.com google.com Male UK 3194 170 34.01 ā£ Shard data by time ā£ Immutable chunks of data called āsegmentsā Segment 2011-01-01T02/2011-01-01T03 Segment 2011-01-01T01/2011-01-01T02
- 21. 2013 IMMUTABLE SEGMENTS ā£ Fundamental storage unit in Druid ā£ Read consistency ā£ One thread scans one segment ā£ Multiple threads can access same underlying data ā£ Segment sizes -> computation completes in ms ā£ Simpliļ¬es distribution & replication
- 22. 2013 COLUMN ORIENTATION timestamp publisher advertiser gender country impressions clicks revenue 2011-01-01T01:00:00Z ultratrimfast.com google.com Male USA 1800 25 15.70 2011-01-01T01:00:00Z bieberfever.com google.com Male USA 2912 42 29.18 ā£ Scan/load only what you need ā£ Compression! ā£ Indexes!
- 23. DRUID INGESTION ā£ Must have denormalized, ļ¬at data ā£ Druid cannot do stateful processing at ingestion time ā£ ā¦like stream-stream joins ā£ ā¦or user session reconstruction ā£ ā¦or a bunch of other useful things! ā£ Many Druid users need an ETL pipeline
- 24. 2013 DRUID REAL-TIME INGESTION Druid Realtime Workers Immediate Druid Historical Nodes Periodic Druid Broker Nodes Data Source User queries
- 26. 2013 DRUID REAL-TIME INGESTION Druid Realtime Workers Immediate Druid Historical Nodes Periodic Druid Broker Nodes Data Source Stream Processor User queries
- 27. 2013 DRUID REAL-TIME INGESTION Druid Realtime Workers Immediate Druid Historical Nodes Periodic Druid Broker Nodes User queries
- 29. AN EXAMPLE: ONLINE ADS ā£ Input data: impressions, clicks, ID-to-name mappings ā£ Output: enhanced impressions ā£ Steps ā£ Join impressions with clicks ->āclicksā ā£ Look up IDs to names -> āadvertiserā, āpublisherā, ā¦ ā£ Geocode -> ācountryā,Ā ā¦ ā£ Lots of other additions
- 31. PIPELINE Impressions Partition 0 {key: 186bd591-9442-48f0, publisher: foo, ā¦} {key: 9b5e2cd2-a8ac-4232, publisher: qux, ā¦} ā¦ Partition 1 {key: 1079026c-7151-4871, publisher: baz, ā¦} ā¦ Clicks Partition 0 ā¦ Partition 1 {key: 186bd591-9442-48f0} ā¦
- 34. PIPELINE Shufļ¬ed Partition 0 {type: impression, key: 186bd591-9442-48f0, publisher: foo, ā¦} {type: impression, key: 1079026c-7151-4871, publisher: baz, ā¦} {type: click, key: 186bd591-9442-48f0} ā¦ Partition 1 {type: impression, key: 9b5e2cd2-a8ac-4232, publisher: qux, ā¦} ā¦
- 37. PIPELINE Joined Partition 0 {key: 186bd591-9442-48f0, is_clicked: true, publisher: foo, ā¦} {key: 1079026c-7151-4871, is_clicked: false, publisher: baz, ā¦} ā¦ Partition 1 {key: 9b5e2cd2-a8ac-4232, is_clicked: false, publisher: qux, ā¦} ā¦
- 41. REPROCESSING
- 42. WHY REPROCESS DATA? ā£ Bugs in processing code ā£ Imprecise streaming operations ā£ ā¦like using short join windows ā£ Limitations of current software ā£ ā¦Kafka 0.8.x, Samza 0.9.x can generate duplicate messages ā£ ā¦Druid 0.7.x streaming ingestion is best-effort
- 43. LAMBDA ARCHITECTURES ā£ Hybrid batch/streaming data pipeline ā£ Batch technologies ā¢ Hadoop MapReduce ā¢ Spark ā£ Streaming technologies ā¢ Samza ā¢ Storm ā¢ Spark Streaming
- 44. LAMBDA ARCHITECTURES ā£ Advantages? ā¢ Works as advertised ā¢ Works with a huge variety of open software ā¢ Druid supports batch-replace-by-time-range through Hadoop
- 45. LAMBDA ARCHITECTURES ā£ Disadvantages? ā£ Need code to run on two very different systems ā£ Maintaining two codebases is perilous ā£ ā¦productivity loss ā£ ā¦code drift ā£ ā¦difļ¬culty training new developers
- 49. KAPPA ARCHITECTURE ā£ Pure streaming ā£ Reprocess data by replaying the input stream ā£ Doesnāt require operating two systems ā£ Doesnāt overcome software limitations ā£ I donāt have much experience with this ā£ http://radar.oreilly.com/2014/07/questioning-the-lambda- architecture.html
- 50. OPERATIONS
- 51. NICE THINGS ABOUT KAFKA ā£ Scalable, replicated pub/sub ā£ Replayable message logs ā£ New consumers can read all old messages ā£ Existing consumers can reprocess all old messages
- 52. NICE THINGS ABOUT SAMZA ā£ Multi-tenancy: one main thread per container ā£ Robustness: isolated containers limit slowness and failure ā£ Visibility ā£ Multistage jobs, lots of metrics per stage ā£ Can inspect the message queue in Kafka ā£ State is simple ā£ Logging and restoring handled for you ā£ Single-threaded programming
- 53. NICE THINGS ABOUT DRUID ā£ Fast ingestion, fast queries ā£ Seamlessly merge stream-ingested and batch-ingested data ā£ Batch loads can āreplaceā stream loads for the same time range
- 54. NICE THINGS ABOUT HADOOP ā£ Solid batch processing system ā£ Easy to partition and reprocess data by time range ā£ Jobs can process all data, or a pre-partitioned slice
- 55. MONITORING ā£ Kafka partition availability ā£ Kafka log cleaner ā£ Samza consumer offsets ā£ Druid ingestion process rate ā£ Druid ingestion drop rate ā£ Druid query latency ā£ System metrics: CPU, network, disk ā£ Event counts at various stages
- 56. STREAM METRICS
- 57. STREAM METRICS
- 58. DO TRY THIS AT HOME
- 59. 2013 CORNERSTONES ā£ Druid - druid.io - @druidio ā£ Samza - samza.apache.org - @samzastream ā£ Kafka - kafka.apache.org - @apachekafka ā£ Hadoop - hadoop.apache.org
- 60. GLUE Tranquility Camus / Secor Druid Hadoop indexer
- 61. GLUE Camus / Secor Druid Hadoop indexer druid-kaka-eight
- 62. TAKE AWAYS ā£ Consider Kafka for making your streams available ā£ Consider Samza for streaming data integration ā£ Consider Druid for interactive exploration of streams ā£ Metrics, metrics, metrics ā£ Have a reprocessing strategy if youāre interested in historical data
- 63. THANK YOU