Re: PATCH: Batch/pipelining support for libpq

On Tue, 16 Feb 2021 at 09:19, Craig Ringer <craig(dot)ringer(at)enterprisedb(dot)com>
wrote:
>
> So long as there is a way to "send A", "send B", "send C", "read results
> from A", "send D", and there's a way for the application to associate some
> kind of state (an application specific id or index, a pointer to an
> application query-queue struct, whatever) so it can match queries to
> results ... then I'm happy.
>

FWIW I'm also thinking of revising the docs to mostly use the term
"pipeline" instead of "batch". Use "pipelining and batching" in the chapter
topic, and mention "batch" in the index, and add a para that explains how
to run batches on top of pipelining, but otherwise use the term "pipeline"
not "batch".

That's because pipelining isn't actually batching, and using it as a naïve
batch interface will get you in trouble. If you PQsendQuery(...) a long
list of queries without consuming results, you'll block unless you're in
libpq's nonblocking-send mode. You'll then be deadlocked because the server
can't send results until you consume some (tx buffer full) and can't
consume queries until it can send some results, but you can't consume
results because you're blocked on a send that'll never end.

An actual batch interface where you can bind and send a long list of
queries might be worthwhile, but should be addressed separately, and it'd
be confusing if this pipelining interface claimed the term "batch". I'm not
convinced enough application developers actually code directly against
libpq for it to be worth creating a pretty batch interface where you can
submit (say) an array of struct PQbatchEntry { const char * query, int
nparams, ... } to a PQsendBatch(...) and let libpq handle the socket I/O
for you. But if someone wants to add one later it'll be easier if we don't
use the term "batch" for the pipelining feature.

I'll submit a reworded patch in a bit.