[jira] [Commented] (FLINK-9702) Improvement in (de)serialization of keys and values for RocksDB state

["Stefan Richter (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-9702?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16529733#comment-16529733 ] 

Stefan Richter commented on FLINK-9702:
---------------------------------------

I have a WIP branch that implements many of the optimizations mentioned in the description. It is currently free for takers because I have to finish some more pressing issues first.

https://github.com/StefanRRichter/flink/tree/serialiation-improvements

> Improvement in (de)serialization of keys and values for RocksDB state
> ---------------------------------------------------------------------
>
>                 Key: FLINK-9702
>                 URL: https://issues.apache.org/jira/browse/FLINK-9702
>             Project: Flink
>          Issue Type: Improvement
>          Components: State Backends, Checkpointing
>    Affects Versions: 1.6.0
>            Reporter: Stefan Richter
>            Priority: Major
>
> When Flink interacts with state in RocksDB, object (de)serialization often contributes significantly to performance overhead. I think there are some aspects that we can improve here to reduce the costs in this area. In particular, currently every state has to serialize the backen's current key before each state access. We could reduce this effort by sharing serialized key bytes across all state interactions. Furthermore, we can reduce the amount of  `byte[]` and stream/view that are involved.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)