[
https://issues.apache.org/jira/browse/FLINK-6284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16008047#comment-16008047
]
ASF GitHub Bot commented on FLINK-6284:
---------------------------------------
GitHub user tillrohrmann opened a pull request:
https://github.com/apache/flink/pull/3884
[FLINK-6284] Correct sorting of completed checkpoints in
ZooKeeperStateHandleStore
In order to store completed checkpoints in an increasing order in ZooKeeper,
the paths for the completed checkpoint is no generated by
`String.format("/%019d", checkpointId)` instead of `String.format("/%s",
checkpointId)`.
This makes sure that the converted long will always have the same length
with
leading 0s.
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/tillrohrmann/flink fixZooKeeperSorting
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/flink/pull/3884.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #3884
----
commit 5bd499329d68c6f3236b4e89ba25fdb9acb7e422
Author: Till Rohrmann <trohrm...@apache.org>
Date: 2017-05-12T12:23:37Z
[FLINK-6284] Correct sorting of completed checkpoints in
ZooKeeperStateHandleStore
In order to store completed checkpoints in an increasing order in ZooKeeper,
the paths for the completed checkpoint is no generated by
String.format("/%019d", checkpointId) instead of String.format("/%s",
checkpointId).
This makes sure that the converted long will always have the same length
with
leading 0s.
----
> Incorrect sorting of completed checkpoints in
> ZooKeeperCompletedCheckpointStore
> -------------------------------------------------------------------------------
>
> Key: FLINK-6284
> URL: https://issues.apache.org/jira/browse/FLINK-6284
> Project: Flink
> Issue Type: Bug
> Components: State Backends, Checkpointing
> Reporter: Xiaogang Shi
> Priority: Blocker
> Fix For: 1.3.0
>
>
> Now all completed checkpoints are sorted in their paths when they are
> recovered in {{ZooKeeperCompletedCheckpointStore}} . In the cases where the
> latest checkpoint's id is not the largest in lexical order (e.g., "100" is
> smaller than "99" in lexical order), Flink will not recover from the latest
> completed checkpoint.
> The problem can be easily observed by setting the checkpoint ids in
> {{ZooKeeperCompletedCheckpointStoreITCase#testRecover()}} to be 99, 100 and
> 101.
> To fix the problem, we should explicitly sort found checkpoints in their
> checkpoint ids, without the usage of
> {{ZooKeeperStateHandleStore#getAllSortedByName()}}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
