On Tue, Jul 11, 2023 at 4:30 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> Jeevan Chalke <jeevan.chalke@enterprisedb.com> writes:
> > Attached patch.
>
> I would be astonished if this fixes anything. The code still doesn't
> know which paths are referenced by which other ones, and so the place
> where we free a previously-added path can't know what to do.
>
> I've speculated about adding some form of reference counting to paths
> (maybe just a "pin" flag rather than a full refcount) so that we could
> be smarter about this. The existing kluge for "don't free IndexPaths"
> could be replaced by setting the pin mark on any IndexPath that we
> make a bitmap path from. Up to now it hasn't seemed necessary to
> generalize that hack, but maybe it's time. Can you show a concrete
> case where we are freeing a still-referenced path?
Set of patches in [1] add infrastructure to reference, link and unlink
paths.The patches are raw and have some TODOs there. But I think that
infrastructure will solve this problem as a side effect. Please take a
look and let me know if this is as per your speculation. It's more
than just pinning though.
The patch set uses references to free memory consumed by paths which
remain unused. The memory consumed is substantial when partitionwise
join is used and there are thousands of partitions.
[1] https://www.postgresql.org/message-id/CAExHW5tUcVsBkq9qT%3DL5vYz4e-cwQNw%3DKAGJrtSyzOp3F%3DXacA%40mail.gmail.com
--
Best Wishes,
Ashutosh Bapat
