[FLINK-23003] Resource leak in RocksIncrementalSnapshotStrategy

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.14.0, 1.13.1, 1.12.4
• Fix Version/s: 1.14.0, 1.13.2
• Component/s: Runtime / Checkpointing, Runtime / State Backends
• Labels: pull-request-available

Environment

Flink: 1.14-SNAPSHOT

• Assignee: Yanfei Lei
• Reporter: Yanfei Lei
• Created: 16/Jun/21 02:16
• Updated: 28/Aug/21 12:18
• Resolved: 17/Jun/21 15:33

Description

We found that `RocksDBStateUploader` in `RocksIncrementalSnapshotStrategy` is not closed correctly after being used. It would lead to a resource leak.

`RocksDBStateUploader` inherits `RocksDBStateDataTransfer`, and `RocksDBStateDataTransfer` holds an `ExecutorService`. `RocksDBStateUploader` uses the `ExecutorService` to upload files to DFS asynchronously.

When `RocksDBKeyedStateBackend` is cleaned up, all resources held by the backend should be closed, but now `RocksIncrementalSnapshotStrategy` lacks a close() function.

And we encountered an example caused by this problem. When we benchmarked the performance of incremental rescaling, we observed that the forked VM of JMH can't exit normally.

[INFO]
[INFO] --- exec-maven-plugin:1.6.0:exec (default-cli) @ benchmark ---
# JMH version: 1.19
# VM version: JDK 1.8.0_281, VM 25.281-b09
# VM invoker: /home/leiyanfei.lyf/jdk1.8.0_281/jre/bin/java
# VM options: -Djava.rmi.server.hostname=127.0.0.1 -Dcom.sun.management.jmxremote.authenticate=false -Dcom.sun.management.jmxremote.ssl=false -Dcom.sun.management.jmxremote.ssl
# Warmup: 10 iterations, 1 s each
# Measurement: 10 iterations, 1 s each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.apache.flink.state.RocksIncrementalCheckpointScaleUpBenchmark.ScalingUp
# Parameters: (numberElements = 100, parallelism1 = 2, parallelism2 = 3)# Run progress: 0.00% complete, ETA 00:02:00
# Fork: 1 of 3
# Warmup Iteration   1: 244.717 ms/op
# Warmup Iteration   2: 104.749 ms/op
# Warmup Iteration   3: 104.182 ms/op
...
Iteration   1: 96.600 ms/op
Iteration   2: 108.463 ms/op
Iteration   3: 93.657 ms/op
...<JMH had finished, but forked VM did not exit, are there stray running threads? Waiting 24 seconds more...>Non-finished threads:
...
Thread[pool-15-thread-4,5,main]
  at sun.misc.Unsafe.park(Native Method)
  at java.util.concurrent.locks.LockSupport.park(LockSupport.java:175)
  at java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.await(AbstractQueuedSynchronizer.java:2039)
  at java.util.concurrent.LinkedBlockingQueue.take(LinkedBlockingQueue.java:442)
  at java.util.concurrent.ThreadPoolExecutor.getTask(ThreadPoolExecutor.java:1074)
  at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1134)
  at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
  at java.lang.Thread.run(Thread.java:748)
<shutdown timeout of 30 seconds expired, forcing forked VM to exit>

 
The root cause of this example is that the `RocksDBStateUploader` in `RocksIncrementalSnapshotStrategy` is not closed normally when `RocksDBKeyedStateBackend` is disposed.

 

The solution to this problem is quite straightforward, `RocksDBStateUploader` in `RocksIncrementalSnapshotStrategy` can be closed when cleaning up `RocksDBKeyedStateBackend`.

Issue Links

• is related to: FLINK-22886 Thread leak in RocksDBStateUploader
• links to: GitHub Pull Request #16168
• links to: GitHub Pull Request #16176