![Use Spark in Local Mode
In your build.sbt:
• src/test/scala => “provided”
• src/main/scala => “compiled”
The Scala Build Tool is your friend!
Simply:
• Import Spark libraries
• Invoke a Context and/or Session
• Set master to local[*] or local[n]](https://image.slidesharecdn.com/sparksummit2019talkstreamingapps1-190425191322/85/Spark-AI-Summit-2019-Headaches-and-Breakthroughs-in-Building-Continuous-Applications-Spark-Streaming-13-320.jpg)
Spark + AI Summit 2019: Headaches and Breakthroughs in Building Continuous Applications (Spark Streaming)
- 1. WIFI SSID:SparkAISummit | Password: UnifiedAnalytics
- 2. Landon Robinson & Jack Chapa Spark Streaming Headaches and Breakthroughs in Building Continuous Applications #UnifiedAnalytics #SparkAISummit
- 3. Who We Are 3#UnifiedAnalytics #SparkAISummit Landon Robinson Data Engineer Jack Chapa Data Engineer Big Data Team @ SpotX
- 4. Because Spark Streaming... • is very powerful • can supercharge your infrastructure • … and can be very complex! Lots of headaches and breakthroughs! But first… why are we here? 4#UnifiedAnalytics #SparkAISummit
- 5. ● Streaming Basics ● Testing ● Monitoring & Alerts ● Batch Intervals & Resources Takeaways Leave with a few actionable items that we wish we knew when we started with Spark Streaming. Focus Areas ● Helpful Configurations ● Backpressure ● Data Enrichment & Transformations
- 6. Our Company 6#UnifiedAnalytics #SparkAISummit The Trusted Platform For Premium Publishers and Broadcasters
- 7. We Process a Lot of Data Data: - 220 MM+ Total Files/Blocks - 8 PB+ HDFS Space - 20 TB+ new data daily - 100MM+ records/minute - 300+ Data Nodes Apps: - Thousands of daily Spark apps - Hundreds of daily user queries - Multiple 24/7 Streaming apps
- 8. Our uses include: - Rapid ingestion of data into warehouse for querying - Machine learning on near-live data streams - Ability to react to and impact live situations - Accelerated processing / updating of metadata - Real-time visualization of data streams and processing Spark Streaming is Key for Us
- 9. Spark Streaming Basics a brief overview
- 10. Spark Streaming Basics Spark Streaming is an extension of Spark that enables scalable, high-throughput, fault-tolerant processing of live data streams. • Stream == live data stream – Topic == Kafka’s name for a stream • DStream == sequence of RDDs formed from reading a data stream • Batch == a self-contained job within your Streaming app that processes a segment of the stream.
- 11. Testing Rapid development and testing of Spark apps
- 12. Use Spark in Local Mode You can start building Spark Streaming apps in minutes, using Spark locally! On your local machine • No cluster needed! • Great for rough testing We Recommend: IntelliJ Community Edition • with SBT: For dependency management
- 13. Use Spark in Local Mode In your build.sbt: • src/test/scala => “provided” • src/main/scala => “compiled” The Scala Build Tool is your friend! Simply: • Import Spark libraries • Invoke a Context and/or Session • Set master to local[*] or local[n]
- 14. Example Unit Test using just a SparkContext Invoke a local session: • In your unit test classes • Test logic on small datasets Add to your deployment pipeline for a nice pre-release gut check!
- 15. Unit Testing Spark Streaming Apps can easily be unit tested - Using .queueStream() - Using a spark testing library Libraries - spark-testing-base - sscheck - spark-tests Use Cases - DStream actions - Business Logic - Integration
- 16. Example Library: spark-testing-base - Easy to Use - Helpful wrappers - Integrates w/ scalatest - Minimal code required - Clock management - Runs alongside other tests GitHub: https://github.com/holdenk/spark-testing-base
- 17. Monitoring Tracking and visualizing performance of your app
- 18. Monitoring is Awesome It can reveal: • How your app is performing • Problems + Bugs! And provide opportunities to: • See and address issues • Observe behavior visually But monitoring can be tough to implement!
- 19. Monitoring (a less than ideal approach) You could do it all in the app... Example: Looping over RDDs to: • Count records • Track Kafka offsets • Processing time / delays But it’s less than ideal... • Calculating performance significantly impacts performance… not great. • All these metrics are calculated by Spark!
- 20. Monitoring and Visualization (using Listeners) Use Spark Listeners to access metrics in the background! Let Spark do the hard work: • Batch duration, delays • Record throughput • Stream position recovery Come to our talk: Spark Listeners: A Crash Course in Fast, Easy Monitoring! • Room 3016 | Today @ 5:30 PM
- 21. Kafka Offset Recovery Saving your place elsewhere
- 22. Inside the Spark Listener class, after a batch completes... You can access an object generated by Spark containing your offsets processed. Take those offsets and back them up to a DB... Writing Offsets to MySQL
- 23. Your offsets are now stored in a DB after each batch completes. Whenever your app restarts, it reads those offsets from the DB... And starts processing where it last left off! Reading Offsets from MySQL
- 24. Getting Offsets from the Database
- 25. Example: Reading Offsets from MySQL
- 26. Example: Reading Offsets from MySQL
- 27. - Record timing info for fast troubleshooting - Escalate alarms to the appropriate team - Quickly resolve while app continues running Timing Logging (around actions)
- 28. React How do I react to this monitoring? ● Heartbeats ● Scheduled Monitor Jobs ○ Version Updates ○ Ensure Running ○ Act on failure/problem ● Monitoring Alarms ● Look at them!
- 29. Batch Intervals Optimizing for speed and resource efficiency
- 30. You want batches that process faster than the interval, but not so fast that resources are idling and therefore wasted! Setting Appropriate Batch Intervals An appropriate batch interval is key to an app that is quick and efficient. Effectiveness of interval is affected by: • Resources alloc (cpu + ram) • Quantity of work • Quantity of data Batch interval
- 31. Setting Appropriate Batch Intervals Consider these questions: How quickly do I need to process data? • Can I slow it down to save resources? What is my resource budget / allocation? • Can I increase? Can I cut back? • Bigger interval = more time to process • … but also more data to process • Smaller interval = the opposite
- 32. Tips for finding an optimal combination: Start small! a. Short batch interval (seconds) b. Modest resources Whichever you have in more flexible supply (a or b), increasing accordingly. Again: processing time < interval = good Comfortably less, not significantly less. Setting Appropriate Batch Intervals
- 33. Additional Resource Notes - Scale down when possible - Free up resources or save on cloud utilization spend - Avoid preemption - Use resource pools with prioritization - With preemption disabled if you can - Set appropriate # of partitions for Kafka topics - Higher volume == higher partition count - Higher partition count == greater parallelization
- 34. Helpful Configuration Settings Configuring your app to be performant and efficient
- 35. Helpful Configuration Settings Spark • spark.memory.useLegacyMode = true – spark.storage.memoryFraction=0.03 • spark.submit.deployMode = cluster • spark.serializer = org.apache.spark.serializer.KryoSerializer • spark.rdd.compress = true – spark.io.compression.codec=org.apache.spark.io.SnappyCompressionCodec • spark.shuffle.service.enabled = true • spark.streaming.blockInterval = 300 Kafka • enable.auto.commit = ‘false’
- 36. Backpressure Use Case: You have irregular spikes in message throughput from Kafka topics • Backpressure dynamically alters rate data is received per batch from Kafka. • Prevents overwhelming of app at startup and peak load. Settings: • spark.streaming.backpressure.enabled = true • spark.streaming.kafka.maxRatePerPartition = 20000 – max rate (messages/second) at which each Kafka partition will be read • PID Rate Estimator: can be used to tweak the rate based on batch performance – spark.streaming.backpressure.pid.* Source: https://www.linkedin.com/pulse/enable-back-pressure-make-your-spark-streaming-production-lan-jiang/
- 37. Transformations Bringing streaming and static data together
- 38. Transformations (Streaming + Static) transform() ● Allows RDD-level access to data. ● Use case: joining with another RDD updateStateByKey() / mapWithState() ● Apply function to each key - useful for keeping track of state ● Use case: maintaining state between batches (e.g. rolling join w/ two streams) reduceByKey() ● Reduce a keyed RDD with appropriate function. ● Use case: deduping, aggregations
- 39. Using the transform() method on DStream: Apply an RDD-to-RDD function to every RDD of the DStream. Used for arbitrary RDD operations on the DStream. • Useful for applying arbitrary RDD operations on a DStream. • Great for enriching streaming data with supplemental static data Joining Streaming and Static Data Source: https://hadoopsters.net/2017/11/26/how-to-join-static-data-with-streaming-data-dstream-in-spark/ transactions = … // streaming dataset (dstream) transaction_details = … // static dataset (rdd) val complete_transaction_data = transactions.transform(live_transaction => live_transaction.join(transaction_details))
- 40. Effective Static Joining How do we handle static and persistent data? Driver: ● Broadcast if small enough ● Read on driver every batch, then join Worker: ● Connect on worker - lazy val connection object ● Useful for persisting data
- 41. Streaming isn’t always easy… but here are some great takeaways! • Testing: Use Spark Locally w/ Unit Tests • Monitoring: Use Listeners & React • Batch Intervals & Resources: Be thoughtful! • Configuration: Lots of awesome ones! • Transformations: Do more with your streaming data! • Offset Recovery: Stop worrying and love the offset management! Review
- 42. Landon Robinson • lrobinson@spotx.tv Jack Chapa • jchapa@spotx.tv hadoopsters.dev https://gist.github.com/hadoopsters Contact Us
- 43. Q & A #UnifiedAnalytics #SparkAISummit
- 44. DON’T FORGET TO RATE AND REVIEW THE SESSIONS SEARCH SPARK + AI SUMMIT