Re: CI failure: postgres_fdw_get_connections

On Fri, Jun 12, 2026 at 8:12 AM Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
>
> Hi,
>
> I just tried the new GitHub Actions CI for the first time. Thanks to
> everybody who worked on making that happen. However, I got a failure,
> on Linux-Meson 32 bit only:
>
> # --- /__w/postgresql/postgresql/contrib/postgres_fdw/expected/postgres_fdw.out
> 2026-06-11 19:31:44.347832846 +0000
> # +++ /__w/postgresql/postgresql/build/testrun/postgres_fdw-running/regress/results/postgres_fdw.out
> 2026-06-11 19:47:06.867590217 +0000
> # @@ -12983,8 +12983,7 @@
> # FROM postgres_fdw_get_connections(true);
> # server_name | closed | remote_backend_pid
> # -------------+--------+--------------------
> # - loopback | t | t
> # -(1 row)
> # +(0 rows)
>
> Run is here: https://github.com/robertmhaas/postgresql/actions/runs/27372091232/job/80887169394
>
> Apparently, this comment isn't always correct:
>
> -- is not available. Despite the termination, remote_backend_pid should
> -- still show the non-zero PID of the terminated remote backend.
>
> The issue seems to be that for the entry to appear in the output of
> postgres_fdw_get_connections, it must not yet have been removed from
> ConnectionHash. However, pgfdw_inval_callback can blow away
> connections freely if they haven't yet been used in the current
> transaction, so invalidation processing at the beginning of any
> statement subsequent to the "SELECT 1 FROM ft1 LIMIT 1;" at
> postgres_fdw.sql line 4607 can close the connection if a relevant
> invalidation message has been received. So I guess maybe there's
> enough DDL happening elsewhere concurrently with this test that a
> sinval reset is possible?

I think your analysis is correct.

postgres_fdw_get_connections(true) can report the connection only if
the entry is still present in ConnectionHash. However, for an idle
cached connection, pgfdw_inval_callback() may discard it before the
next statement runs. In that case, returning 0 rows seems like a
legitimate result, so I think the current comment and expected output
are too strict.

What seems better is to split the test expectations into two parts:

1. For the current idle-connection case, accept either outcome:
if the cache entry is still present, verify that it reports closed
status and a nonzero remote_backend_pid; otherwise accept 0 rows.

2. To preserve coverage of the terminated-backend reporting path, add
a separate check inside an explicit transaction. In that case, the
connection is already in use, so invalidation may mark it invalid
but cannot discard it before transaction end, and
postgres_fdw_get_connections(true) should still report it.

The attached patch implements this approach.

Regards,

--
Fujii Masao