Ben Snaidero <bensnaidero@geotab.com> writes:
> I am running into a strange issue with Postgres 10 when using pg_dump with
> the directory format and jobs parameter set it intermittently hangs. Seems
> to occur less frequently the lower I set the jobs parameter but does happen
> eventually even when set to 2 (could not reproduce when jobs=1). I've
> tested with Postgres 11 and no matter how high I set the jobs parameter it
> never hangs (tested all the way up to jobs=8). Also with Postgres 10 and
> other formats I don't run into the issue. Here are my server specs and
> software versions (OS and Postgres)
FWIW, it looks like the hang is on the client side (i.e., pg_dump itself).
The one active server process is just trying to send COPY data and is
waiting for the client to take it. So I'm wondering about deadlocks
inside parallel pg_dump.
This doesn't present any obvious answer, because there is not much
difference between the parallel pg_dump logic in 10 and 11. But
maybe you could get stack trace(s) out of the stuck pg_dump job on 10?
(Keep in mind that unlike the backend, the pg_dump process will have
multiple threads in this situation.)
Another thing you could do to triangulate is to verify that v11 pg_dump
is okay working against the v10 server. Unfortunately the other
combination can't be tested, but this'd be enough to exonerate the v10
server side.
regards, tom lane