flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stephan Ewen <se...@apache.org>
Subject Re: DFS problem with removing checkpoint
Date Mon, 02 Apr 2018 13:58:38 GMT
Can you clarify which one does not get deleted? The file in the
"high-availability.storageDir",
or the  "state.backend.fs.checkpointdir/JobId/check-<INTEGER>", or both?

Could you also tell us which file system you use?

There is a known issue in some versions of Flink that S3 "directories" are
not deleted. This means that Hadoop's S3 marker files (the way that
Hadoop's s3n and s3a  imitate directories in S3) are not deleted. This is
fixed in Flink 1.5. A workaround for Flink 1.4 is to use the
"flink-s3-fs-presto", which does not uses these marker files.


On Thu, Mar 29, 2018 at 8:49 PM, Szymon Szczypiński <simon_t@poczta.fm>
wrote:

> Thank you for your replay.
>
> But my problem is not that Flink doesn't remove files/directories after
> incorrect job cancellation. My problem is different.
> Precisely when everything is ok and job is in RUNNING state, when
> checkpoint for job is done, then there is created file in
> "high-availability.storageDir" with name completedCheckpoint<SomeID> and
> also is created dir in location "state.backend.fs.
> checkpointdir/JobId/check-<INTEGER>". After the next checkpoint is
> completed, previous file and dir are deleted, and this is ok, because i
> always have only one checkpoint.
> But in my case when next checkpoint is completed, the previous is not
> deleted and this happens when job is in running state.
>
> My be you know why those files/dirs are not deleted.
>
> Best regards
>
> Szymon Szczypiński
>
>
> On 29.03.2018 11:23, Stephan Ewen wrote:
>
> Flink removes these files / directories only if you properly cancel the
> job. If you kill the processes (stop-cluster.sh) this looks like a failure,
> and Flink will try to recover. The recovery always starts in ZooKeeper, not
> in the DFS.
>
> Best way to prevent this is to
>   - properly cancel jobs, not just kill processes
>   - use separate cluster IDs for the standalone clusters, so that the new
> cluster knows that it is not supposed to recover the previous jobs and
> checkpoints
>
>
>
> On Wed, Mar 28, 2018 at 11:22 PM, Szymon Szczypiński <simon_t@poczta.fm>
> wrote:
>
>> Hi,
>>
>> i have problem with Flink in version 1.3.1.
>>
>> I have standalone cluster with two JobManagers and four TaskManager, as
>> DFS i use windows high available storage mounted by cifs protocol.
>>
>> And sometimes i'm starting having problem that Flink doesn't remove
>> checkpoint dirs for job and completedCheckpoint files from
>> "high-availability.storageDir".
>>
>> To bring back cluster to normal working i need to remove all dirs from
>> DFS and start everything from beginning.
>>
>>
>> Maybe someone of Flink users had the same problem. For now i doesn't
>> have any idea how to bring back cluster to normal work without deleting
>> dirs from DFS.
>>
>> I don't want to delete dirs from DFS because than  i need to redeploy
>> all jobs.
>>
>>
>> Best regards
>>
>> Szymon Szczypiński
>>
>>
>
>

Mime
View raw message