[
https://issues.apache.org/jira/browse/FLINK-6485?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17757764#comment-17757764
]
Hangxiang Yu commented on FLINK-6485:
-------------------------------------
IIUC, ChangelogStateBackend ([Generalized incremental
checkpoints|https://cwiki.apache.org/confluence/display/FLINK/FLIP-158%3A+Generalized+incremental+checkpoints])
could help to resolve this basically.
> Use buffering to avoid frequent memtable flushes for short intervals in
> RockdDB incremental checkpoints
> -------------------------------------------------------------------------------------------------------
>
> Key: FLINK-6485
> URL: https://issues.apache.org/jira/browse/FLINK-6485
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Checkpointing
> Reporter: Stefan Richter
> Priority: Not a Priority
> Labels: auto-deprioritized-major, auto-deprioritized-minor
>
> The current implementation of incremental checkpoitns in RocksDB needs to
> flush the memtable to disk prior to a checkpoint and this will generate a SST
> file.
> What is required for fast checkpoint intervals is an alternative mechanism to
> quickly determine a delta from the previous incremental checkpoint to avoid
> this frequent flushing. This could be implemented through custom buffering
> inside the backend, e.g. a changelog buffer that is maintain up to a certain
> size.
> The buffer's content becomes part of the private state in the incremental
> snapshot and the buffer is dropped i) after each checkpoint or ii) after
> exceeding a certain size that justifies flushing and writing a new SST file.
> This mechanism should not be blocking, which we can achieve in the following
> way:
> 1) We have a clear upper limit to the buffer size (e.g. 64MB), once the limit
> of diffs is reached, we can drop the buffer because we can assume enough work
> was done to justify a new SST file
> 2) We write the buffer to a local FS, so we can expect this to be reasonable
> fast and that it will not suffer from the kind of blocking that we have in
> DFS. I mean technically, also flushing the SST file can block. Then, in the
> async part, we can transfer the locally written buffer file to DFS.
> There might be other mechanisms in RocksDB that we could exploit for this,
> such as the write ahead log, but this could be already be a good solution.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
