On Fri, Nov 18, 2022 at 11:08 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Thomas Munro <thomas.munro@gmail.com> writes:
> > I wonder why the walreceiver didn't start in
> > 008_min_recovery_point_node_3.log here:
> > https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mamba&dt=2022-11-16%2023%3A13%3A38
>
> mamba has been showing intermittent failures in various replication
> tests since day one. My guess is that it's slow enough to be
> particularly subject to the signal-handler race conditions that we
> know exist in walreceivers and elsewhere. (Now, it wasn't any faster
> in its previous incarnation as a macOS critter. But maybe modern
> NetBSD has different scheduler behavior than ancient macOS and that
> contributes somehow. Or maybe there's some other NetBSD weirdness
> in here.)
>
> I've tried to reproduce manually, without much success :-(
>
> Like many of its other failures, there's a suggestive postmaster
> log entry at the very end:
>
> 2022-11-16 19:45:53.851 EST [2036:4] LOG: received immediate shutdown request
> 2022-11-16 19:45:58.873 EST [2036:5] LOG: issuing SIGKILL to recalcitrant children
> 2022-11-16 19:45:58.881 EST [2036:6] LOG: database system is shut down
>
> So some postmaster child is stuck somewhere where it's not responding
> to SIGQUIT. While it's not unreasonable to guess that that's a
> walreceiver, there's no hard evidence of it here. I've been wondering
> if it'd be worth patching the postmaster so that it's a bit more verbose
> about which children it had to SIGKILL. I've also wondered about
> changing the SIGKILL to SIGABRT in hopes of reaping a core file that
> could be investigated.
I wonder if it's a runtime variant of the other problem. We do
load_file("libpqwalreceiver", false) before unblocking signals but
maybe don't resolve the symbols until calling them, or something like
that...
