[
https://issues.apache.org/jira/browse/FLINK-4193?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15624754#comment-15624754
]
Gyula Fora commented on FLINK-4193:
-----------------------------------
These issues usually happened inside the RocksDB.open(...) method during
initialization of the state backend. If you think that the refactoring can
affect this then we might get lucky :)
We are running this in production applications and haven't ported them to 1.2
but in a week or two I will start working on that.
> Task manager JVM crashes while deploying cancelling jobs
> --------------------------------------------------------
>
> Key: FLINK-4193
> URL: https://issues.apache.org/jira/browse/FLINK-4193
> Project: Flink
> Issue Type: Bug
> Components: Streaming, TaskManager
> Reporter: Gyula Fora
> Priority: Critical
>
> We have observed several TM crashes while deploying larger stateful streaming
> jobs that use the RocksDB state backend.
> As the JVMs crash the logs don't show anything but I have uploaded all the
> info I have got from the standard output.
> This indicates some GC and possibly some RocksDB issues underneath but we
> could not really figure out much more.
> GC segfault
> https://gist.github.com/gyfora/9e56d4a0d4fc285a8d838e1b281ae125
> Other crashes (maybe rocks related)
> https://gist.github.com/gyfora/525c67c747873f0ff2ff2ed1682efefa
> https://gist.github.com/gyfora/b93611fde87b1f2516eeaf6bfbe8d818
> The third link shows 2 issues that happened in parallel...
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
