Flink Forward SF 2017: Stephan Ewen - Experiences running Flink at Very Large Scale
7 likes1,558 views
The document discusses lessons learned from running Apache Flink at large scale, including challenges with task deployment, RPC volume, checkpointing, and state management. It outlines the importance of understanding and optimizing checkpoints, network performance, and file system interactions to mitigate issues related to scalability and performance. Key recommendations are provided to improve recovery time, CPU efficiency, and overall application robustness.
Flink Forward SF 2017: Stephan Ewen - Experiences running Flink at Very Large Scale
- 1. Experiences in running Apache Flink® at large scale Stephan Ewen (@StephanEwen)
- 2. 2 Lessons learned from running Flink at large scale Including various things we never expected to become a problem and evidently still did… Also a preview to various fixes coming in Flink…
- 3. What is large scale? 3 Large Data Volume ( events / sec ) Large Application State (GBs / TBs) High Parallelism (1000s of subtasks) Complex Dataflow Graphs (many operators)
- 5. Deploying Tasks 5 Happens during initial deployment and recovery JobManager TaskManager Akka / RPC Akka / RPC Blob Server Blob Server Deployment RPC Call Contains - Job Configuration - Task Code and Objects - Recover State Handle - Correlation IDs
- 6. Deploying Tasks 6 Happens during initial deployment and recovery JobManager TaskManager Akka / RPC Akka / RPC Blob Server Blob Server Deployment RPC Call Contains - Job Configuration - Task Code and Objects - Recover State Handle - Correlation IDs KBs up to MBs KBs few bytes
- 7. RPC volume during deployment 7 (back of the napkin calculation) number of tasks 2 MB parallelism size of task objects 100010 x x x x = = RPC volume 20 GB ~20 seconds on full 10 GBits/s net > 1 min with avg. of 3 GBits/s net > 3 min with avg. of 1GBs net
- 8. Timeouts and Failure detection 8 ~20 seconds on full 10 GBits/s net > 1 min with avg. of 3 GBits/s net > 3 min with avg. of 1GBs net Default RPC timeout: 10 secs default settings lead to failed deployments with RPC timeouts Solution: Increase RPC timeout Caveat: Increasing the timeout makes failure detection slower Future: Reduce RPC load (next slides)
- 9. Dissecting the RPC messages 9 Message part Size Variance across subtasks and redeploys Job Configuration KBs constant Task Code and Objects up to MBs constant Recover State Handle KBs variable Correlation IDs few bytes variable
- 10. Upcoming: Deploying Tasks 10 Out-of-band transfer and caching of large and constant message parts JobManager TaskManager Akka / RPC Akka / RPC Blob Server Blob Cache (1) Deployment RPC Call (Recover State Handle, Correlation IDs, BLOB pointers) (2) Download and cache BLOBs (Job Config, Task Objects) MBs KBs
- 12. 12 …is the most important part of running a large Flink program Robustly checkpointing…
- 13. Review: Checkpoints 13 Trigger checkpoint Inject checkpoint barrier stateful operation source / transform
- 14. Review: Checkpoints 14 Take state snapshot Trigger state snapshot stateful operation source / transform
- 15. Review: Checkpoint Alignment 15 1 2 3 4 5 6 a b c d e f begin aligning checkpoint barrier n xy operator 7 8 9 c d e f g h aligning ab operator 23 1 4 5 6 input buffer y
- 16. Review: Checkpoint Alignment 16 7 8 9 d e f g h bc operator 23 1 5 6 emit barrier n 7 8 9 d e f g h c operator 23 1 5 6 input buffer continuecheckpoint 4 4 i a
- 18. Understanding Checkpoints 18 How well behaves the alignment? (lower is better) How long do snapshots take? delay = end_to_end – sync – async
- 19. Understanding Checkpoints 19 How well behaves the alignment? (lower is better) How long do snapshots take? delay = end_to_end – sync – async long delay = under backpressure under constant backpressure means the application is under provisioned too long means ! too much state per node ! snapshot store cannot keep up with load (low bandwidth) changes with incremental checkpoints most important metric
- 20. Alignments: Limit in-flight data ▪ In-flight data is data "between" operators • On the wire or in the network buffers • Amount depends mainly on network buffer memory ▪ Need some to buffer out network fluctuations / transient backpressure ▪ Max amount of in-flight data is max amount buffered during alignment 20 1 2 3 4 5 6 a b c d e f checkpoint barrier xyoperator
- 21. Alignments: Limit in-flight data ▪ Flink 1.2: Global pool that distributes across all tasks • Rule-of-thumb: set to 4 * num_shuffles * parallelism * num_slots ▪ Flink 1.3: Limits the max in-flight data automatically • Heuristic based on of channels and connections involved in a transfer step 21 1 2 3 4 5 6 a b c d e f checkpoint barrier xyoperator
- 22. Heavy alignments ▪ A heavy alignment typically happens at some point ! Different load on different paths ▪ Big window emission concurrent to a checkpoint ▪ Stall of one operator on the path 22
- 23. Heavy alignments ▪ A heavy alignment typically happens at some point ! Different load on different paths ▪ Big window emission concurrent to a checkpoint ▪ Stall of one operator on the path 23
- 24. Heavy alignments ▪ A heavy alignment typically happens at some point ! Different load on different paths ▪ Big window emission concurrent to a checkpoint ▪ Stall of one operator on the path 24 GC stall
- 25. Catching up from heavy alignments ▪ Operators that did heavy alignment need to catch up again ▪ Otherwise, next checkpoint will have a heavy alignment as well 25 7 8 9 d e f g h operator 5 6 7 8 9 d e f g h bc operator 23 1 5 6 4 a consumed first after checkpoint completed bc a
- 26. Catching up from heavy alignments ▪ Giving the computation time to catch up before starting the next checkpoint • Useful: Set the min-time-between-checkpoints ▪ Asynchronous checkpoints help a lot! • Shorter stalls in the pipelines means less build-up of in-flight data • Catch up already happens concurrently to state materialization 26
- 27. Asynchronous Checkpoints 27 Durably persist snapshots asynchronously Processing pipeline continues stateful operation source / transform
- 28. Asynchrony of different state types 28 State Flink 1.2 Flink 1.3 Flink 1.3 + Keyed state RocksDB ✔ ✔ ✔ Keyed State on heap ✘ (✔) (hidden in 1.2.1) ✔ ✔ Timers ✘ ✔/✘ ✔ Operator State ✘ ✔ ✔
- 29. When to use which state backend? 29 Async. Heap RocksDB State ≥ Memory ? Complex Objects? (expensive serialization) high data rate? no yes yes no yes no a bit simplified
- 30. File Systems, Object Stores, and Checkpointed State 30
- 31. Exceeding FS request capacity ▪ Job size: 4 operators ▪ Parallelism: 100s to 1000 ▪ State Backend: FsStateBackend ▪ State size: few KBs per operator, 100s to 1000 of files ▪ Checkpoint interval: few secs ▪ Symptom: S3 blocked off connections after exceeding 1000s HEAD requests / sec 31
- 32. Exceeding FS request capacity What happened? ▪ Operators prepare state writes, ensure parent directory exists ▪ Via the S3 FS (from Hadoop), each mkdirs causes 2 HEAD requests ▪ Flink 1.2: Lazily initialize checkpoint preconditions (dirs.) ▪ Flink 1.3: Core state backends reduce assumption of directories (PUT/GET/DEL), rich file systems support them as fast paths 32
- 33. Reducing FS stress for small state 33 JobManager TaskManager Checkpoint Coordinator Task TaskManager Task TaskTask Root Checkpoint File (metadata) checkpoint data files Fs/RocksDB state backend for most states
- 34. Reducing FS stress for small state 34 JobManager TaskManager Checkpoint Coordinator Task TaskManager Task TaskTask checkpoint data directly in metadata file Fs/RocksDB state backend for small states ack+data Increasing small state threshold reduces number of files (default: 1KB)
- 35. Lagging state cleanup 35 JobManager TM TMTM TMTM TM TMTMTM Symptom: Checkpoints get cleaned up too slow State accumulates over time many TaskManagers create files One JobManager deleting files
- 36. Lagging state cleanup ▪ Problem: FileSystems and Object Stores offer only synchronous requests to delete state object Time to delete a checkpoint may accumulates to minutes. ▪ Flink 1.2: Concurrent checkpoint deletes on the JobManager ▪ Flink 1.3: For FileSystems with actual directory structure, use recursive directory deletes (one request per directory) 36
- 37. Orphaned Checkpoint State 37 JobManager Checkpoint Coordinator TaskManager TaskTask (1) TaskManager writes state (2) Ack and transfer ownership of state (3) JobManager records state reference Who owns state objects at what time?
- 38. Orphaned Checkpoint State 38 fs:///checkpoints/job-61776516/ chk-113 / chk-129 / chk-221 / chk-271 / chk-272 / It gets more complicated with incremental checkpoints… Upcoming: Searching for orphaned state latest retained periodically sweep checkpoint directory for leftover dirs
- 40. 40 The closer you application is to saturating either network, CPU, memory, FS throughput, etc. the sooner an extraordinary situation causes a regression Enough headroom in provisioned capacity means fast catchup after temporary regressions Be aware that certain operations are spiky (like aligned windows) Production test always with checkpoints ;-)
- 41. Recommendations (part 1) Be aware of the inherent scalability of primitives ▪ Broadcasting state is useful, for example for updating rules / configs, dynamic code loading, etc. ▪ Broadcasting does not scale, i.e., adding more nodes does not. Don't use it for high volume joins ▪ Putting very large objects into a ValueState may mean big serialization effort on access / checkpoint ▪ If the state can be mappified, use MapState – it performs much better 41
- 42. Recommendations (part 2) If you are about recovery time ▪ Having spare TaskManagers helps bridge the time until backup TaskManagers come online ▪ Having a spare JobManager can be useful • Future: JobManager failures are non disruptive 42
- 43. Recommendations (part 3) If you care about CPU efficiency, watch your serializers ▪ JSON is a flexible, but awfully inefficient data type ▪ Kryo does okay - make sure you register the types ▪ Flink's directly supported types have good performance basic types, arrays, tuples, … ▪ Nothing ever beats a custom serializer ;-) 43