On Fri, 10 Mar 2023 at 21:42, Alexander Korotkov <aekorotkov@gmail.com> wrote:
>
> I wonder why does ExecMergeMatched() determine the lock mode using
> ExecUpdateLockMode(). Why don't we use lock mode set by
> table_tuple_update() like ExecUpdate() does? I skim through the
> MERGE-related threads, but didn't find an answer.
>
> I also noticed that we use ExecUpdateLockMode() even for CMD_DELETE.
> That ends up by usage of LockTupleNoKeyExclusive for CMD_DELETE, which
> seems plain wrong for me.
>
> The proposed change is attached.
>
That won't work if it did a cross-partition update, since it won't
have done a table_tuple_update() in that case, and updateCxt.lockmode
won't have been set. Also, when it loops back and retries, it might
execute a different action next time round. So I think it needs to
unconditionally use LockTupleExclusive, since it doesn't know if it'll
end up executing an update or a delete.
I'm currently working on a patch for bug #17809 that might change that
code though.
Regards,
Dean