Robert Haas wrote:
It seems to me that you need to start by thinking about what kinds of
queries could be usefully parallelized. What I think you're proposing
here, modulo large amounts of hand-waving, is that we should basically
find a branch of the query tree, cut it off, and make that branch the
responsibility of a subprocess. What kinds of things would be
sensible to hand off in this way? Well, you'd want to find nodes that
are not likely to be repeatedly re-executed with different parameters,
like subplans or inner-indexscans, because otherwise you'll get
pipeline stalls handing the new parameters back and forth. And you
want to find nodes that are expensive for the same reason. So maybe
this would work for something like a merge join on top of two sorts -
one backend could perform each sort, and then whichever one was the
child would stream the tuples to the parent for the final merge. Of
course, this assumes the I/O subsystem can keep up, which is not a
given - if both tables are fed by the same, single spindle, it might
be worse than if you just did the sorts consecutively.
This approach might also benefit queries that are very CPU-intensive,
on a multi-core system with spare cycles. Suppose you have a big tall
stack of hash joins, each with a small inner rel. The child process
does about half the joins and then pipelines the results into the
parent, which does the other half and returns the results.
But there's at least one other totally different way of thinking about
this problem, which is that you might want two processes to cooperate
in executing the SAME query node - imagine, for example, a big
sequential scan with an expensive but highly selective filter
condition, or an enormous sort. You have all the same problems of
figuring out when it's actually going to help, of course, but the
details will likely be quite different.
I'm not really sure which one of these would be more useful in
practice - or maybe there are even other strategies. What does
$COMPETITOR do?
I'm also ignoring the difficulties of getting hold of a second backend
in the right state - same database, same snapshot, etc. It seems to
me unlikely that there are a substantial number of real-world
applications for which this will not work very well if we have to
actually start a new backend every time we want to parallelize a
query. IOW, we're going to need, well, a connection pool in core.
*ducks, runs for cover*
One thing that might work quite well is slicing up by partition
(properly implemented partitioning would go along with this nicely too...)
regards
Mark
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers