[jira] [Commented] (FLINK-3431) Add retrying logic for RocksDB snapshots

["Shuyi Chen (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-3431?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16142536#comment-16142536 ] 

Shuyi Chen commented on FLINK-3431:
-----------------------------------

Anyone looking at this issue? I can take a look otherwise.

> Add retrying logic for RocksDB snapshots
> ----------------------------------------
>
>                 Key: FLINK-3431
>                 URL: https://issues.apache.org/jira/browse/FLINK-3431
>             Project: Flink
>          Issue Type: Improvement
>          Components: DataStream API, State Backends, Checkpointing
>            Reporter: Gyula Fora
>            Priority: Critical
>
> Currently the RocksDB snapshots rely on hdfs copy not failing while taking the snapshots.
> In some cases when the state size is big enough the HDFS nodes might get so overloaded that the copy operation fails on errors like this:
> AsynchronousException{java.io.IOException: All datanodes 172.26.86.90:50010 are bad. Aborting...}
> at org.apache.flink.streaming.runtime.tasks.StreamTask$1.run(StreamTask.java:545)
> Caused by: java.io.IOException: All datanodes 172.26.86.90:50010 are bad. Aborting...
> at org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer.setupPipelineForAppendOrRecovery(DFSOutputStream.java:1023)
> at org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer.processDatanodeError(DFSOutputStream.java:838)
> at org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer.run(DFSOutputStream.java:483)
> I think it would be important that we don't immediately fail the job in these cases but retry the copy operation after some random sleep time. It might be also good to do a random sleep before the copy depending on the state size to smoothen out IO a little bit.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)