[
https://issues.apache.org/jira/browse/FLINK-17468?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17096317#comment-17096317
]
Yun Tang commented on FLINK-17468:
----------------------------------
[~qqibrow] I think the async phase of checkpoint would only do one thing:
writing data to distributed file system. No matter it's uploading sst files in
RocksDB incremental checkpoints or just writing to output stream in RocksDB
savepoint or FsStateBackend's checkpoint/savepoint. From my experience, if the
async phase took too long, it means the performance of writing data to
distributed storage behaves bad (indicates the poor performance of DFS). I
cannot actually see the real difference if we track the uploading part in async
uploading. Would you please share you successful story here?
> Provide more detailed metrics why asynchronous part of checkpoint is taking
> long time
> -------------------------------------------------------------------------------------
>
> Key: FLINK-17468
> URL: https://issues.apache.org/jira/browse/FLINK-17468
> Project: Flink
> Issue Type: New Feature
> Components: Runtime / Checkpointing, Runtime / Metrics, Runtime /
> State Backends
> Affects Versions: 1.10.0
> Reporter: Piotr Nowojski
> Priority: Major
>
> As [reported by
> users|https://lists.apache.org/thread.html/r0833452796ca7d1c9d5e35c110089c95cfdadee9d81884a13613a4ce%40%3Cuser.flink.apache.org%3E]
> it's not obvious why asynchronous part of checkpoint is taking so long time.
> Maybe we could provide some more detailed metrics/UI/logs about uploading
> files, materializing meta data, or other things that are happening during the
> asynchronous checkpoint process?
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
