Re: hyrax vs. RelationBuildPartitionDesc

Robert Haas <robertmhaas@gmail.com> writes:
> On Tue, Jun 11, 2019 at 1:57 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:

> I think the change is responsive to your previous complaint that the
> timing of stuff getting freed is not very well pinned down.  With this
> change, it's much more tightly pinned down: it happens when the
> refcount goes to 0.  That is definitely not perfect, but I think that
> it is a lot easier to come up with scenarios where the leak
> accumulates because no cache flush happens while the relfcount is 0
> than it is to come up with scenarios where the refcount never reaches
> 0.  I agree that the latter type of scenario probably exists, but I
> don't think we've come up with one yet.

I don't know why you think that's improbable, given that the changes
around PartitionDirectory-s cause relcache entries to be held open much
longer than before (something I've also objected to on this thread).

>> As I said upthread, my current inclination is to do nothing in this
>> area for v12 and then try to replace the whole thing with proper
>> reference counting in v13.  I think the cases where we have a major
>> leak are corner-case-ish enough that we can leave it as-is for one
>> release.

> Is this something you're planning to work on yourself?

Well, I'd rather farm it out to somebody else, but ...

> Do you have a
> design in mind?  Is the idea to reference-count the PartitionDesc?

What we discussed upthread was refcounting each of the various
large sub-objects of relcache entries, not just the partdesc.
I think if we're going to go this way we should bite the bullet and fix
them all.  I really want to burn down RememberToFreeTupleDescAtEOX() in
particular ... it seems likely to me that that's also a source of
unpleasant memory leaks.

			regards, tom lane