Re: postgres fe/be protocol

Chris Bitmead <chrisb(at)nimrod(dot)itg(dot)telstra(dot)com(dot)au> writes:
> Tom Lane wrote:
>> I'm pretty sure the backend sends only one 'Z' per query cycle.

> I put in a printf in parseInput in fe-exec.c in libpq, straight after
> it reads the id.

Oh. What you're missing is that the input message isn't necessarily
*consumed* when first seen. The sequence of events is

* absorb bufferload of data into libpq input buffer;

* parse messages in parseInput until 'C' is seen;

* when 'C' message is completely read, set asyncStatus = PGASYNC_READY
(fe-exec.c line 756 in current sources), then absorb that message by
advancing inCursor (line 878) and loop back around. Now it sees the
final 'Z' message, but decides at line 713 not to process it just yet.
So it returns to PQgetResult, which finishes off and stashes away the
PGresult object, and then sets asyncStatus = PGASYNC_BUSY again.

* PQexec will now loop around and call PQgetResult and thence parseInput
again, which will now absorb the 'Z' and set asyncStatus = PGASYNC_IDLE,
which finally allows the PQexec loop to return. In the meantime your
printf printed the 'Z' a second time.

This is kind of baroque, I agree, but it seemed to be necessary to
support asynchronous data reading without doing too much damage to
backward compatibility of PQexec() semantics. In particular, the
critical thing here is that we need to be able to deliver a sequence
of PGresult objects when the query string contains multiple commands
and the application is using PQgetResult --- but the old behavior of
PQexec was that you only got back the *last* PGresult if the query
string produced more than one, so I had to preserve that...

regards, tom lane