[
https://issues.apache.org/jira/browse/FLINK-31903?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17721753#comment-17721753
]
Zhipeng Zhang commented on FLINK-31903:
---------------------------------------
The root cause is that using memory as statebackend has limited state size. So
this is not a bug.
" java.io.IOException: Size of the state is larger than the maximum permitted
memory-backed state. Size=5365814, maxSize=5242880. Consider using a different
checkpoint storage, like the FileSystemCheckpointStorage.".
> Caching records fails in BroadcastUtils#withBroadcastStream
> -----------------------------------------------------------
>
> Key: FLINK-31903
> URL: https://issues.apache.org/jira/browse/FLINK-31903
> Project: Flink
> Issue Type: Bug
> Components: Library / Machine Learning
> Affects Versions: ml-2.3.0
> Reporter: Zhipeng Zhang
> Priority: Major
>
> When caching more than 1,000,000 records using BroadcastUtils#withBroadcast,
> it throws exception as follows:
> {code:java}
> Caused by: org.apache.flink.util.FlinkRuntimeException: Exceeded checkpoint
> tolerable failure threshold.
> at
> org.apache.flink.runtime.checkpoint.CheckpointFailureManager.checkFailureAgainstCounter(CheckpointFailureManager.java:206)
> at
> org.apache.flink.runtime.checkpoint.CheckpointFailureManager.handleTaskLevelCheckpointException(CheckpointFailureManager.java:191)
> at
> org.apache.flink.runtime.checkpoint.CheckpointFailureManager.handleCheckpointException(CheckpointFailureManager.java:124)
> at
> org.apache.flink.runtime.checkpoint.CheckpointCoordinator.abortPendingCheckpoint(CheckpointCoordinator.java:2078)
> at
> org.apache.flink.runtime.checkpoint.CheckpointCoordinator.receiveDeclineMessage(CheckpointCoordinator.java:1038)
> at
> org.apache.flink.runtime.scheduler.ExecutionGraphHandler.lambda$declineCheckpoint$2(ExecutionGraphHandler.java:103)
> at
> org.apache.flink.runtime.scheduler.ExecutionGraphHandler.lambda$processCheckpointCoordinatorMessage$3(ExecutionGraphHandler.java:119)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
> at java.lang.Thread.run(Thread.java:748)
> {code}
> It seems that the bug comes from caching too many records when calling
> AbstractBroadcastWrapperOperator#snapshot.
>
> The failed case could be found here:
> [https://github.com/zhipeng93/flink-ml/tree/FLINK-31903-fail-case]
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
