[
https://issues.apache.org/jira/browse/FLINK-5670?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15842584#comment-15842584
]
Aljoscha Krettek commented on FLINK-5670:
-----------------------------------------
The problem is that we create to many subdirectories and
{{RocksDBKeyedStateBackend.dispose}} will only clean up the last part of this.
I recently fixed that for 1.1 here:
https://github.com/apache/flink/commit/e1861db7f71d8583c40b358ee474c77210c761f9.
I think we should also get this into Flink 1.2. What do you think [~rmetzger]?
I know of at least one big user of Flink for which this was a potential problem.
> Local RocksDB directories not cleaned up
> ----------------------------------------
>
> Key: FLINK-5670
> URL: https://issues.apache.org/jira/browse/FLINK-5670
> Project: Flink
> Issue Type: Bug
> Components: State Backends, Checkpointing
> Reporter: Ufuk Celebi
> Priority: Minor
>
> After cancelling a job with a RocksDB backend all files are properly cleaned
> up, but the parent directories still exist and are empty:
> {code}
> 859546fec3dac36bb9fcc8cbdd4e291e
> +- StreamFlatMap_3_0
> +- StreamFlatMap_3_3
> +- StreamFlatMap_3_4
> +- StreamFlatMap_3_5
> +- StreamFlatMap_3_6
> {code}
> The number of empty folders varies between runs.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
