[
https://issues.apache.org/jira/browse/FLINK-14110?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16932672#comment-16932672
]
Aaron Levin commented on FLINK-14110:
-------------------------------------
[~carp84] [~klion26] I have not made a patch. I would be happy to look into it,
but am not quite sure where to start (familiarity with RocksDB and Flink's
state backend is low). I worry a little that this issue could be upstream in
RocksDB itself (or a mishandling of exceptions at the JNI layer), but haven't
checked.
if your intuition is that this shouldn't be too hard to patch, then I can take
a look!
> Deleting state.backend.rocksdb.localdir causes silent failure
> -------------------------------------------------------------
>
> Key: FLINK-14110
> URL: https://issues.apache.org/jira/browse/FLINK-14110
> Project: Flink
> Issue Type: Bug
> Components: Runtime / State Backends
> Affects Versions: 1.8.1, 1.9.0
> Environment: Flink {{1.8.1}} and {{1.9.0}}.
> JVM 8
> Reporter: Aaron Levin
> Priority: Minor
>
> Suppose {{state.backend.rocksdb.localdir}} is configured as:
> {noformat}
> state.backend.rocksdb.localdir: /flink/tmp
> {noformat}
> If I then run \{{rm -rf /flink/tmp/job_*}} on a host while a Flink
> application is running, I will observe the following:
> * throughput of my operators running on that host will drop to zero
> * the application will not fail or restart
> * the task manager will not fail or restart
> * in most cases there is nothing in the logs to indicate a failure (I've run
> this several times and only once seen an exception - I believe I was lucky
> and deleted those directories during a checkpoint or something)
> The desired behaviour here would be to throw an exception and crash, instead
> of silently dropping throughput to zero. Restarting the Task Manager will
> resolve the issues.
> I only tried this on Flink {{1.8.1}} and {{1.9.0}}.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
