osdir.com

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Setting an allowable number of checkpoint failures


Hi vino,

Yes, I believe we are on the same page. I created
https://issues.apache.org/jira/browse/FLINK-10074 to track it.

Thanks,
Thomas

On Mon, Aug 6, 2018 at 8:42 AM vino yang <yanghua1127@xxxxxxxxx> wrote:

> Hi Thomas,
>
> What I am saying is what you mean, maybe I am not very accurate.
>
> Thanks, vino.
>
> 2018-08-06 21:22 GMT+08:00 Thomas Weise <thw@xxxxxxxxxx>:
>
> > Hi,
> >
> > What we are looking for is that the job does *not* restart on transient
> > checkpoint failures and we would like to cap the number of allowable
> > subsequent failures until a restart occurs.
> >
> > The reason is that every restart is a service interruption that is
> > potentially very expensive.
> >
> > Thanks,
> > Thomas
> >
> >
> >
> >
> >
> >
> >
> > On Mon, Aug 6, 2018 at 5:09 AM vino yang <yanghua1127@xxxxxxxxx> wrote:
> >
> > > Hi Till,
> > >
> > > I think the way you proposed is a solution. But I think we also can
> > provide
> > > a solution to prevent Checkpoint from failing indefinitely, in case the
> > Job
> > > does not fail.
> > >
> > > Instead, a threshold is given to allow the checkpoint to fail a few
> > times.
> > > When this threshold is reached, we decide to let the job fail.
> > >
> > > Thanks, vino.
> > >
> > > 2018-08-06 15:14 GMT+08:00 Till Rohrmann <trohrmann@xxxxxxxxxx>:
> > >
> > > > Hi Lakshmi,
> > > >
> > > > you could somewhat achieve the described behaviour by setting
> > > > setFailOnCheckpointintErrors(true) and using the
> > > > FailureRateRestartStrategy
> > > > as the restart strategy. That way checkpoint failures will trigger a
> > job
> > > > restart (this is the downside) which is handled by the restart
> > strategy.
> > > > The FailureRateRestartStrategy allows for x failures to happen within
> > in
> > > a
> > > > given time interval. If this number is exceeded, then the job will
> > > > terminally fail.
> > > >
> > > > Cheers,
> > > > Till
> > > >
> > > > On Sat, Aug 4, 2018 at 4:58 AM vino yang <yanghua1127@xxxxxxxxx>
> > wrote:
> > > >
> > > > > Hi Lakshmi,
> > > > >
> > > > > Your understanding of "
> > > > > *CheckpointConfig#setFailOnCheckpointingErrors(false)*" is correct,
> > If
> > > > this
> > > > > is set to false, the task will only decline a the checkpoint and
> > > continue
> > > > > running.
> > > > >
> > > > > I think it is also a good choice to allow a number of failures to
> be
> > > set.
> > > > > Flink currently only supports whether the Task fails if the
> > checkpoint
> > > > > fails. It is not supported to configure a threshold.
> > > > >
> > > > > You can create an issue in JIRA to feedback this requirement.
> > > > >
> > > > > Thanks, vino.
> > > > >
> > > > > 2018-08-04 4:28 GMT+08:00 Lakshmi Gururaja Rao <lrao@xxxxxxxx>:
> > > > >
> > > > > > Hi,
> > > > > >
> > > > > > We are running into intermittent checkpoint failures while
> > > > checkpointing
> > > > > to
> > > > > > S3.
> > > > > >
> > > > > > As described in this thread -
> > > > > >  http://apache-flink-user-mailing-list-archive.2336050.
> > > > > > n4.nabble.com/1-5-some-thing-weird-td21309.html
> > > > > > <http://apache-flink-user-mailing-list-archive.2336050.
> > > > > > n4.nabble.com/1-5-some-thing-weird-td21309.html>,
> > > > > > we see that the job restarts when it encounters such a failure.
> > > > > >
> > > > > > As mentioned in the thread, I see that there is an option to not
> > fail
> > > > > tasks
> > > > > > on checkpoint errors -
> > > > > > *CheckpointConfig#setFailOnCheckpointingErrors(false)**.
> *However,
> > > > this
> > > > > > would mean that the job would continue running even in the case
> of
> > > > > > persistent checkpoint failures. Is my understanding here correct?
> > > > > >
> > > > > > If above is true, then is there a way to configure an allowable
> > > number
> > > > of
> > > > > > checkpoint failures? i.e. something along the lines of "Don't
> fail
> > > the
> > > > > job
> > > > > > if there are <=X number of checkpoint failures", so that *only
> > > > *transient
> > > > > > failures can be ignored.
> > > > > >
> > > > > > Thanks,
> > > > > > Lakshmi
> > > > > >
> > > > >
> > > >
> > >
> >
>