At Wed, 15 Jun 2022 14:56:42 -0400, Tom Lane <tgl@sss.pgh.pa.us> wrote in
> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:
> > So, git archaeology led me to this thread
> > https://postgr.es/m/202106072107.d4i55hdscxqj@alvherre.pgsql
> > which is why we added that message in the first place.
>
> Um. Good thing you looked. I doubt we want to revert that change now.
>
> > Alternatives:
> > - Have the client not complain if it gets CloseComplete in idle state.
> > (After all, it's a pretty useless message, since we already do nothing
> > with it if we get it in BUSY state.)
>
> ISTM the actual problem here is that we're reverting to IDLE state too
> soon. I didn't try to trace down exactly where that's happening, but
Yes. I once visited that fact but also I thought that in the
comparison with non-pipelined PQsendQuery, the three messages look
extra. Thus I concluded (at the time) that removing Close is enough
here.
> I notice that in the non-pipeline case we don't go to IDLE till we've
> seen 'Z' (Sync). Something in the pipeline logic must be jumping the
> gun on that state transition.
PQgetResult() resets the state to IDLE when not in pipeline mode.
fe-exec.c:2171
> if (conn->pipelineStatus != PQ_PIPELINE_OFF)
> {
> /*
> * We're about to send the results of the current query. Set
> * us idle now, and ...
> */
> conn->asyncStatus = PGASYNC_IDLE;
And actually that code let the connection state enter to IDLE before
CloseComplete. In the test case I posted, the following happens.
PQsendQuery(conn, "SELECT 1;");
PQsendFlushRequest(conn);
PQgetResult(conn); // state enters IDLE, reads down to <CommandComplete>
PQgetResult(conn); // reads <CloseComplete comes>
PQpipelineSync(conn); // sync too late
Pipeline feature seems intending to allow PQgetResult called before
PQpipelineSync. And also seems allowing to call QPpipelineSync() after
PQgetResult().
I haven't come up with a valid *fix* of this flow..
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center