On 11/12/2023 09:31, Richard Guo wrote:
> On Fri, Dec 8, 2023 at 3:13 PM Alexander Pyhalov
> <a.pyhalov@postgrespro.ru <mailto:a.pyhalov@postgrespro.ru>> wrote:
> Andrei Lepikhov писал(а) 2023-12-08 07:37:
> > I'd already clashed with Tom on copying the required_relids field
> and
> > voluntarily made unnecessary copies in the project [1].
> > And ... stuck into huge memory consumption. The reason was in
> > Bitmapsets:
> > When we have 1E3-1E4 partitions and try to reparameterize a join,
> one
> > bitmapset field can have a size of about 1kB. Having bitmapset
> > referencing Relation with a large index value, we had a lot of (for
> > example, 1E4 * 1kB) copies on each reparametrization of such a
> field.
> > Alexander Pyhalov should remember that case.
> Yes. If it matters, this happened during reparametrization when 2
> partitioned tables with 1000 partitions each were joined. Then
> asymmetric pw join managed to eat lots of memory for bitmapsets (by
> lots of memory I mean all available on the test VM).
> By reparametrization did you mean the work done in
> reparameterize_path_by_child()? If so maybe you'd be interested in the
> patch [1] which postpones reparameterization of paths until createplan.c
> and thus can help avoid unnecessary reparametrization work.
Yeah, I have discovered it already. It is a promising solution and only
needs a bit more review. But here, I embraced some corner cases with the
idea that we may not see other cases right now. And also, sometimes the
Bitmapset field is significant - it is not a corner case.
--
regards,
Andrei Lepikhov
Postgres Professional