Hi Alvaro,
On Sat, Apr 3, 2021 at 12:01 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:
> On 2021-Apr-02, Amit Langote wrote:
>
> > On Sat, Mar 20, 2021 at 10:21 PM Amit Langote <amitlangote09@gmail.com> wrote:
> > > Updated patches attached. Sorry about the delay.
> >
> > Rebased over the recent DETACH PARTITION CONCURRENTLY work.
> > Apparently, ri_ReferencedKeyExists() was using the wrong snapshot.
>
> Hmm, I wonder if that stuff should be using a PartitionDirectory? (I
> didn't actually understand what your code is doing, so please forgive if
> this is a silly question.)
No problem, I wondered about that too when rebasing.
My instinct *was* that maybe there's no need for it, because
find_leaf_pk_rel()'s use of a PartitionDesc is pretty limited in
duration and scope of the kind of things it calls that there's no need
to worry about it getting invalidated while in use. But I may be
wrong about that, because get_partition_for_tuple() can call arbitrary
user-defined functions, which may result in invalidation messages
being processed and an unguarded PartitionDesc getting wiped out under
us.
So, I've added PartitionDirectory protection in find_leaf_pk_rel() in
the attached updated version.
--
Amit Langote
EDB: http://www.enterprisedb.com
