[
https://issues.apache.org/jira/browse/FLINK-21351?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17282968#comment-17282968
]
Till Rohrmann commented on FLINK-21351:
---------------------------------------
Would it work if we don't count savepoints towards the number of maximum
retained checkpoints? That way we would only remove checkpoints if a
non-savepoint is being added. Consequently, we would maintain the assumption on
which the current implementation of incremental checkpoints are built that SST
files are only deleted if there is no other checkpoint relying on it.
Implementation wise this could be done by introducing a separate checkpoint
counter to the {{DefaultCompletedCheckpointStore}}. It adds a bit of logic to
always keep this counter correctly set but it should be a rather small change.
> Incremental checkpoint data would be lost once a non-stop savepoint completed
> -----------------------------------------------------------------------------
>
> Key: FLINK-21351
> URL: https://issues.apache.org/jira/browse/FLINK-21351
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Checkpointing
> Affects Versions: 1.11.3, 1.12.1, 1.13.0
> Reporter: Yun Tang
> Priority: Blocker
> Fix For: 1.11.4, 1.12.2, 1.13.0
>
>
> FLINK-10354 counted savepoint as retained checkpoint so that job could
> failover from latest position. I think this operation is reasonable, however,
> current implementation would let incremental checkpoint data lost immediately
> once a non-stop savepoint completed.
> Current general phase of incremental checkpoints: once a newer checkpoint
> completed, it would be added to checkpoint store. And if the size of
> completed checkpoints larger than max retained limit, it would subsume the
> oldest one. This lead to the reference of incremental data decrease one and
> data would be deleted once reference reached to zero. As we always ensure to
> register newer checkpoint and then unregister older checkpoint, current phase
> works fine as expected.
> However, if a non-stop savepoint (a median manual trigger savepoint) is
> completed, it would be also added into checkpoint store and just subsume
> previous added checkpoint (in default retain one checkpoint case), which
> would unregister older checkpoint without newer checkpoint registered,
> leading to data lost.
> Thanks for [~banmoy] reporting this problem first.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
