Nargarjun, thanks a lot for the reply, which makes sense to me. Yes, we are running with AT_LEAST_ONCE mode.
On Wed, Nov 21, 2018 at 3:19 PM Nagarjun Guraja <nagar...@gmail.com> wrote: > Hi Steven, > > The metric 'Buffered During Alignment' you are talking about will always > be zero when the job is run in ATLEAST_ONCE mode. Is that the case with > your job? My understanding is, backpressure can only be monitored by > sampling thread stacktraces and interpreting the situation based on the > contention for network buffers on demand. > > Regards, > Nagarjun > > *Success is not final, failure is not fatal: it is the courage to continue > that counts. * > *- Winston Churchill - * > > > On Wed, Nov 21, 2018 at 1:50 PM Steven Wu <stevenz...@gmail.com> wrote: > >> >> Flink has two backpressure related metrics: “ >> lastCheckpointAlignmentBuffered” and “checkpointAlignmentTime”. But they >> seems to always report zero. Similar thing in web UI, “Buffered During >> Alignment” always shows zero, even backpressure testing shows high >> backpressure for some operators. Has anyone else seen similar problem? >> >> We are running flink 1.4.0 with some cherry-picked fixes. there was a bug >> and fix for 1.5 and above, which shouldn't affect us >> https://issues.apache.org/jira/browse/FLINK-10135 >> >> Thanks, >> Steven >> >
