Mailing-List: contact user-help@flink.apache.org; run by ezmlm
Precedence: bulk
Reply-To: user@flink.apache.org
MIME-Version: 1.0
In-Reply-To: <CAKiyyaF1Ox68NCs95ptKSAp59Y0TOhERvv7A48efripXbjcHpw@mail.gmail.com>
References: <57343E9D.2050107@touk.pl> <CAKiyyaF1Ox68NCs95ptKSAp59Y0TOhERvv7A48efripXbjcHpw@mail.gmail.com>
From: Ufuk Celebi <uce@apache.org>
Date: Thu, 12 May 2016 18:39:12 +0200
Message-ID: <CAKiyyaEwYPbbXkDmXSd1Q34m8kLw+xWAD9UVCGVEYKEj1p+3kw@mail.gmail.com>
Subject: Re: checkpoints not being removed from HDFS
To: Ufuk Celebi <uce@apache.org>
Cc: user@flink.apache.org
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: quoted-printable
archived-at: Thu, 12 May 2016 16:39:54 -0000

The issue is here: https://issues.apache.org/jira/browse/FLINK-3902

(My "explanation" before dosn't make sense actually and I don't see a
reason why this should be related to having many state handles.)

On Thu, May 12, 2016 at 3:54 PM, Ufuk Celebi <uce@apache.org> wrote:
> Hey Maciek,
>
> thanks for reporting this. Having files linger around looks like a bug to=
 me.
>
> The idea behind having the recursive flag set to false in the
> AbstractFileStateHandle.discardState() call is that the
> FileStateHandle is actually just a single file and not a directory.
> The second call trying to delete the parent directory only succeeds
> when all other files in that directory have been deleted as well. I
> think this is what sometimes fails with many state handles. For
> RocksDB there is only a single state handle, which works well.
>
> I will open an issue for this and try to reproduce it reliably and then f=
ix it.
>
> =E2=80=93 Ufuk
>
>
> On Thu, May 12, 2016 at 10:28 AM, Maciek Pr=C3=B3chniak <mpr@touk.pl> wro=
te:
>> Hi,
>>
>> we have stream job with quite large state (few GB), we're using
>> FSStateBackend and we're storing checkpoints in hdfs.
>> What we observe is that v. often old checkpoints are not discarded prope=
rly.
>> In hadoop logs I can see:
>>
>> 2016-05-10 12:21:06,559 INFO BlockStateChange: BLOCK* addToInvalidates:
>> blk_1084791727_11053122 10.10.113.10:50010
>> 2016-05-10 12:21:06,559 INFO org.apache.hadoop.ipc.Server: IPC Server
>> handler 9 on 8020, call
>> org.apache.hadoop.hdfs.protocol.ClientProtocol.delete from 10.10.113.9:4=
9233
>> Call#12337 Retry#0
>> org.apache.hadoop.fs.PathIsNotEmptyDirectoryException:
>> `/flink/checkpoints_test/570d6e67d571c109daab468e5678402b/chk-62 is non
>> empty': Directory is not empty
>>         at
>> org.apache.hadoop.hdfs.server.namenode.FSDirDeleteOp.delete(FSDirDeleteO=
p.java:85)
>>         at
>> org.apache.hadoop.hdfs.server.namenode.FSNamesystem.delete(FSNamesystem.=
java:3712)
>>
>> While on flink side (jobmanager log) we don't see any problems:
>> 2016-05-10 12:20:22,636 [Checkpoint Timer] INFO
>> org.apache.flink.runtime.checkpoint.CheckpointCoordinator     - Triggeri=
ng
>> checkpoint 62 @ 1462875622636
>> 2016-05-10 12:20:32,507 [flink-akka.actor.default-dispatcher-240088] INF=
O
>> org.apache.flink.runtime.checkpoint.CheckpointCoordinator - Completed
>> checkpoint 62 (in 9843 ms)
>> 2016-05-10 12:20:52,637 [Checkpoint Timer] INFO
>> org.apache.flink.runtime.checkpoint.CheckpointCoordinator     - Triggeri=
ng
>> checkpoint 63 @ 1462875652637
>> 2016-05-10 12:21:06,563 [flink-akka.actor.default-dispatcher-240028] INF=
O
>> org.apache.flink.runtime.checkpoint.CheckpointCoordinator - Completed
>> checkpoint 63 (in 13909 ms)
>> 2016-05-10 12:21:22,636 [Checkpoint Timer] INFO
>> org.apache.flink.runtime.checkpoint.CheckpointCoordinator     - Triggeri=
ng
>> checkpoint 64 @ 1462875682636
>>
>> I see in the code that delete operations in flink are done with recursiv=
e
>> flag set to false - but I'm not sure why the contents are not being dele=
ted
>> before?
>> When we were using RocksDB backed we didn't encounter such situation.
>> we're using flink 1.0.1 and hdfs 2.7.2.
>>
>> Do anybody has any idea why this could be happening?
>>
>> thanks,
>> maciek
>>
>>
>>