liam saffioti <liam.saffiotti@gmail.com> writes:
> Do you think timescaledb is causing this problem?
Seems possible.
> 2021-09-10 01:00:32.691 156668 FATAL: cannot read pg_class without
> having selected a database
> 2021-09-10 01:00:33.356 110257 LOG: autovacuum launcher process (PID
> 156668) exited with exit code 1
This is pretty clear evidence that some code running inside the
autovacuum launcher process tried to read a non-shared catalog, which it
cannot do because the launcher isn't connected to any specific database
within the cluster. It is highly unlikely that anything in core Postgres
is doing that; we'd have noticed. So I conclude that some extension code
is carelessly trying to do catalog accesses without checking to see if
it's running in a background process that doesn't support that.
Of the extensions you listed, timescaledb is perhaps the most likely to
contain such a bug; but I don't know much about hypopg, pg_partman,
pg_stat_kcache, or powa, so I can't completely rule those out. It would
have to be an extension that sometimes acts of its own accord without
being called by a SQL command, so the list of suspects isn't all that
long.
Anyway, your first step should be to make sure you have up-to-date
versions of all those extensions. If you do, then try filing a bug
with timescaledb.
regards, tom lane
