Re: CommandStatus from insert returning when using a portal.

"David G. Johnston" <david.g.johnston@gmail.com> writes:
> I agree that the documented contract of the insert command tag says it
> reports the size of the entire tuple store maintained by the server during
> the transaction instead of just the most recent count on subsequent fetches.

Where do you see that documented, exactly?  I looked in the protocol
chapter and didn't find anything definite either way.

I'm quite prepared to believe there are bugs here, since this whole
set of behaviors is unreachable via libpq: you can't get it to send
an Execute with a count other than zero (ie, fetch all), nor is it
prepared to deal with the PortalSuspended messages it'd get if it did.

I think that the behavior is arising from this bit in PortalRun:

        switch (portal->strategy)
        {
            ...
            case PORTAL_ONE_RETURNING:
            ...

                /*
                 * If we have not yet run the command, do so, storing its
                 * results in the portal's tuplestore.  But we don't do that
                 * for the PORTAL_ONE_SELECT case.
                 */
                if (portal->strategy != PORTAL_ONE_SELECT && !portal->holdStore)
                    FillPortalStore(portal, isTopLevel);

                /*
                 * Now fetch desired portion of results.
                 */
                nprocessed = PortalRunSelect(portal, true, count, dest);

                /*
                 * If the portal result contains a command tag and the caller
                 * gave us a pointer to store it, copy it and update the
                 * rowcount.
                 */
                if (qc && portal->qc.commandTag != CMDTAG_UNKNOWN)
                {
                    CopyQueryCompletion(qc, &portal->qc);
------>>>           qc->nprocessed = nprocessed;
                }

                /* Mark portal not active */
                portal->status = PORTAL_READY;

                /*
                 * Since it's a forward fetch, say DONE iff atEnd is now true.
                 */
                result = portal->atEnd;

The marked line is, seemingly intentionally, overriding the portal's
rowcount (which ought to count the whole query result) with the number
of rows processed in the current fetch.  Chap's theory that that's
always zero when this is being driven by JDBC seems plausible,
since the query status won't be returned to the client unless we
detect end-of-portal (otherwise we just send PortalSuspended).

It seems plausible to me that we should just remove that marked line.
Not sure about the compatibility implications, though.

            regards, tom lane