On Tue, Aug 10, 2021 at 8:15 PM Ha Ka <klasahubert@gmail.com> wrote:
>
> sob., 19 cze 2021 o 12:14 Amit Kapila <amit.kapila16@gmail.com> napisał(a):
>
> We increased logical_decoding_work_mem for our production database
> from 64 to 192 MB and it looks like the issue still persists. The
> frequency with which replication hangs has remained the same.
>
Sounds strange. I think one thing to identify at the time slowdown has
happened is whether there are a very large number of in-progress
transactions at the time slowdown happened. Because the profile shared
last time seems to be spending more time in hash_seq_search than in
actually serializing the exact. Another possibility to try out for
your case is to just always serialize the current xact and see what
happens, this might not be an actual solution but can help in
diagnosing the problem.
> Do you
> need any additional perf reports after our change?
>
It might be good if you can share the WALSender portion of perf as
shared in one of the emails above?
--
With Regards,
Amit Kapila.
