Re: POC: Lock updated tuples in tuple_update() and tuple_delete()

On Sun, Apr 2, 2023 at 3:47 AM Andres Freund <andres(at)anarazel(dot)de> wrote:
> On 2023-04-02 03:37:19 +0300, Alexander Korotkov wrote:
> > On Sat, Apr 1, 2023 at 8:21 AM Andres Freund <andres(at)anarazel(dot)de> wrote:
> > > Given that the in-tree state has been broken for a week, I think it probably
> > > is time to revert the commits that already went in.
> >
> > The revised patch is attached. The most notable change is getting rid
> > of LazyTupleTableSlot. Also get rid of complex computations to detect
> > how to initialize LazyTupleTableSlot. Instead just pass the oldSlot
> > as an argument of ExecUpdate() and ExecDelete(). The price for this
> > is just preallocation of ri_oldTupleSlot before calling ExecDelete().
> > The slot allocation is quite cheap. After all wrappers it's
> > table_slot_callbacks(), which is very cheap, single palloc() and few
> > fields initialization. It doesn't seem reasonable to introduce an
> > infrastructure to evade this.
> >
> > I think patch resolves all the major issues you've highlighted. Even
> > if there are some minor things missed, I'd prefer to push this rather
> > than reverting the whole work.
>
> Shrug. You're designing new APIs, days before the feature freeze. This just
> doesn't seem ready in time for 16. I certainly won't have time to look at it
> sufficiently in the next 5 days.

OK. Reverted.

------
Regards,
Alexander Korotkov