Kinesis and Spark Streaming - Advanced AWS Meetup - August 2014
Download as PPTX, PDF70 likes17,686 views
Spark Streaming allows for processing of real-time data streams using Spark. The document discusses using Spark Streaming with Amazon Kinesis for streaming data ingestion. It covers the Spark Streaming and Kinesis integration architecture, how the Spark Kinesis receiver works, scaling considerations, and fault tolerance mechanisms through checkpointing. Examples of monitoring and tuning Spark Streaming jobs on Kinesis data are also provided.
Ad
More Related Content
Viewers also liked (13)
Similar to Kinesis and Spark Streaming - Advanced AWS Meetup - August 2014 (20)
Ad
More from Chris Fregly (20)
Ad
Recently uploaded (20)
![Pixologic ZBrush Crack Plus Activation Key [Latest 2025] New Version](https://cdn.slidesharecdn.com/ss_thumbnails/fashionevolution2-250322112409-f76abaa7-250428124909-b51264ff-250504160528-fc2bb1c5-thumbnail.jpg?width=560&fit=bounds)
Kinesis and Spark Streaming - Advanced AWS Meetup - August 2014
- 1. Spark Streaming + Kinesis Chris Fregly AdvancedAWS Meetup 8/12/2014 Kinesis Streaming
- 2. Who am I? Former Netflix’er (netflix.github.io) Spark Contributor (github.com/apache/spark) Consultant (fluxcapacitor.com) Author (sparkinaction.com)
- 3. Quick Poll • Spark, Spark Streaming? • Hadoop, Hive, Pig? • EMR, Redshift? • Flume, Kafka, Kinesis, Storm? • Lambda Architecture?
- 5. Spark Overview • Part of Berkeley Data Analytics Stack (“badass”) • ~2009, Berkeley AMPLab • Written in Scala • Supports Java, Python, SQL, and R • In-memory whenever possible • Improved efficiency over MapReduce – 100x in-memory, 2-10x on-disk • Compatible with Hadoop – File formats, SerDes, and UDFs – Hive (Shark) and Pig (Spork)
- 6. Spark API • Richer, more expressive API than just map() and reduce() – filter(), join(), distinct(), groupByKey() • Supports Java, Python, SQL, and R • Resilient Distributed Dataset (RDD) – Core Spark abstraction – Partition across cluster – Parallel, fault tolerant, immutable, recomputable • Unified API across all libraries
- 8. Spark Libraries • Spark SQL (Data Processing) • Spark Streaming (Streaming) • MLlib (Machine Learning) • GraphX (Graph Processing) • BlinkDB (Approximate Queries)
- 9. Similar Projects • Spark – Microsoft Dryad – Apache Tez – Impala – Google BigQuery – Google Cloud Dataflow
- 10. Spark + Hadoop Cluster View
- 11. Master High Availability • Multiple Master Nodes • ZooKeeper maintains current Master • Existing applications and workers will be notified of new Master election • New applications and workers need to explicitly specify current Master • Alternatives (Not recommended) – Local filesystem – NFS Mount
- 12. Spark Execution Engine • General cluster computing engine • Parallel, distributed, DAG-based • Lazy evaluation • Allows optimizations • Data locality and rack awareness • Fine-grained fault tolerance using RDD lineage graphs
- 15. Spark Streaming Overview • Low latency, high throughput, fault-tolerant • DStream: Micro-batches of RDDs – Operations are similar to RDD – Lineage for fault-tolerance • Supports Flume, Kafka, Twitter, Kinesis, etc. • Built on Spark Core Execution Engine and API • Long-running Spark Application
- 16. Spark Streaming API • Rich, expressive API based on core Spark API – filter(), join(), distinct(), groupByKey() • Maintain State – updateStateByKey() • Window Operations – Window size & slide interval • Checkpointing • Register DStream as a SQL table
- 18. Similar Projects • Spark Streaming – Twitter Storm – Yahoo! S4 – LinkedIn Samsa – Google Millwheel
- 19. Spark Streaming Use Cases • Operational dashboards • ETL on streaming data ingestion • Anomaly, malware, and fraud detection • Predictive maintenance – Sensors • NLP analysis – Twitter Firehose • Lambda architecture – Unified batch and streaming – ie. Different machine learning models for different time frames
- 20. Spark Streaming Cluster View
- 21. Spark Streaming + Kinesis Spark Kinesis Producer Spark Kinesis Producer Spark St reaming Kinesis Archit ect ure Kinesis Spark St reaming, Kinesis Client Library Applicat ion Kinesis St ream Shard 1 Shard 2 Shard 3 Spark Kinesis Producer Spark Kinesis Receiver Spark Kinesis Receiver Spark Kinesis Receiver
- 22. Spark Kinesis Receiver Internals
- 23. Scaling • Horizontally scale by adding more Kinesis Receivers • Kinesis Client Library within each Kinesis Receiver will negotiate and rebalance shard processing • Never need more Kinesis Receivers than the number of stream shards • Each Kinesis Receiver can support multiple shards • Supports Kinesis shard splitting/merging • Recommendation: over provision shards and avoid splitting/merging
- 25. Streaming Receiver Failure • Upon failure, backup receiver takes over • Checkpoint sources like Kafka and kinesis allow multiple receivers to pull from the same stream (ie. during a failover) – De-duping is handled by Spark • Supports graceful shutdown to allow in-flight message draining • Recommendation: choose buffered sources like Flume, Kafka and Kinesis
- 26. Streaming Driver Failure • Streaming Driver app is long-running – Monitor driver, receiver worker nodes, and streams – Alert upon failure or unusually high latency • Driver failure interrupts stream processing • Enable checkpointing for backup Driver to take over • Use StreamingContext.getOrCreate(…) in Driver app
- 27. Types of Checkpoints Spark 1. Spark checkpointing of StreamingContext DStreams and metadata 2. Lineage of state and window DStream operations Kinesis 3. Kinesis Client Library (KCL) checkpoints current position within shard – Checkpoint info is stored in DynamoDB per Kinesis Application keyed by shard
- 28. Monitoring • Streaming tab in Spark Web UI • CloudWatch • StreamingListener callback
- 30. Tuning • Batch interval – High: reduce overhead of submitting new tasks for each batch – Low: keeps latencies low – Recommendation: test & find the sweet spot • Checkpoint interval – High: reduce load on checkpoint overhead – Low: reduce amount of data loss on failure – Recommendation: 5-10x sliding window interval • Explicitly uncache DStreams when no longer needed • Use CMS GC for consistent processing times • Use Kryo serialization • DStreams are already serialized as byte arrays (versus Java objects) to minimize GC
- 32. Spark + AWS + Lambda
- 33. Spark + AWS + Lambda + ML
- 34. Summary • Spark • Spark Streaming + Kinesis • Scaling • Fault Tolerance • Monitoring • Tuning • Lambda Architecture • Spark in Action – Oct 2014 MEAP – Early access: http://sparkinaction.com