Re: logical decoding : exceeded maxAllocatedDescs for .spill files

On Wed, 20 Nov 2019 at 19:24, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:
>
> On 2019-Nov-20, Juan José Santamaría Flecha wrote:
>
> > I was not able to reproduce the Permission denied error with current HEAD,
> > until I opened another CMD inside the "pg_replslot/regression_slot" folder.
> > This will be problematic, is the deletion of the folder actually needed?
>
> Yes :-(  The code assumes that if the directory is there, then it's
> valid.  Trying to remove that assumption is probably a more invasive
> fix.
>
> I think ReplicationSlotDropAcquired is too pessimistic (no recourse if
> the rename fails) and too optimistic (this will almost never happen).
> We could change it so that the rename is retried a few times, and avoid
> the failure.  (Naturally, the rmtree should also be retried.)  The code
> seems written with the POSIX semantics in mind, but it seems easy to
> improve.

Just to be clear, there are two issues being discussed here :

1. Issue with the patch, where pg_replslot/slotname/xid-*.spill files
can't be removed because the same backend process has left these files
opened because of an abort. This is happening despite the file being
opened using FILE_SHARE_DELETE flag. I am going to investigate
(possibly the flag is not applicable in case a single process is
involved)

2. This existing issue where pg_replslot/slotname directory removal
will fail if someone else is accessing this directory. This has
nothing to do with the patch.


--
Thanks,
-Amit Khandekar
EnterpriseDB Corporation
The Postgres Database Company