Hi,
On Fri, Jun 19, 2026 at 01:28:21PM +0000, Bertrand Drouvot wrote:
> Hi,
>
> On Thu, Jun 18, 2026 at 04:21:33PM -0700, Jeff Davis wrote:
>
> > We could keep the transaction ID in the tracking record, and ignore
> > entries from an aborted subxact. But it's getting fairly complex and
> > delicate.
As discussed off-list, let's move forward with this proposal (as the save a copy
of the object's ACL + ROLE membership mentioned in [1] would be more complex and
fragile).
So PFA the new version doing so. It's the same patch series as the previous one
except that it adds:
0004: Discard ACL tracking entries from aborted subtransactions
It stores SubTransactionId in each AclCheckEntry at recording time.
On subtransaction abort, AtEOSubXact_AclTrack() discards entries whose subxid
is >= the aborting subtransaction's ID.
The code is pretty simple but I suspect that adding a test would probably need to
create a brand new module (say test_acltrack) with C code forcing
recordDependencyOn() and BeginInternalSubTransaction()) and I'm not sure it's
worth it. Indeed, I think the test would need to do things similar to what you
shared in [2].
The patch is split to ease the review but should probably be merged as a single
patch.
[1]: https://postgr.es/m/ajVF7YWf%2Bpxs4cOf%40bdtpg
[2]: https://postgr.es/m/5c629d2455946ad2fde3c184f64ea2c323ef2133.camel@j-davis.com
Regards,
--
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com
