[
https://issues.apache.org/jira/browse/FLINK-5715?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15940608#comment-15940608
]
ASF GitHub Bot commented on FLINK-5715:
---------------------------------------
Github user StefanRRichter commented on the issue:
https://github.com/apache/flink/pull/3602
Thanks for the review, @StephanEwen ! A agree with undoing
`MemoryStateBackend` and `MathUtil` as proposed, but would argue for keeping
the change in `TimeWindow`. This change fixes skew problems with the old way
the hash code was generated (being timestamps, typically only a small bandwith
of bits in the hash code was used by the other approach). This skew can have
particular impact in the new, flattened hashmap structure of the async backends.
> Asynchronous snapshotting for HeapKeyedStateBackend
> ---------------------------------------------------
>
> Key: FLINK-5715
> URL: https://issues.apache.org/jira/browse/FLINK-5715
> Project: Flink
> Issue Type: New Feature
> Components: State Backends, Checkpointing
> Affects Versions: 1.3.0
> Reporter: Stefan Richter
> Assignee: Stefan Richter
> Fix For: 1.3.0
>
>
> Blocking snapshots render the HeapKeyedStateBackend practically unusable for
> many user in productions. Their jobs can not tolerate stopped processing for
> the time it takes to write gigabytes of data from memory to disk.
> Asynchronous snapshots would be a solution to this problem. The challenge for
> the implementation is coming up with a copy-on-write scheme for the in-memory
> hash maps that build the foundation of this backend. After taking a closer
> look, this problem is twofold. First, providing CoW semantics for the hashmap
> itself, as a mutible structure, thereby avoiding costly locking or blocking
> where possible. Second, CoW for the mutable value objects, e.g. through
> cloning via serializers.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
