Re: Revert RI fast-path batching from REL_19_STABLE

Hi,

On Thu, Sep 10, 2026 at 3:20 PM Amit Langote <amitlangote09(at)gmail(dot)com> wrote:
> I have now reverted batching in REL_19_STABLE after pushing the
> snapshot fix for the per-row path to master and REL_19_STABLE. and .
> Batching remains in master for now.

I propose removing batching from master as well. The patch to do so is attached.

I had hoped we could continue fixing the remaining issues on master,
but delaying checks to accumulate a batch has implications that I
haven't fully accounted for. For example, as mentioned upthread, Tomas
reported to me off-list that other AFTER ROW triggers can modify the
referenced table before a buffered check runs, changing its result
compared with checking immediately.

We could address that particular case by disabling batching when such
triggers exist. I have also posted patches for other reported issues,
but I'd prefer to step back and revert all the batching-related code
for now. My ability to work on these fixes is also limited in the near
term due to some personal circumstances.

I'd still like to revisit batching in a separate proposal, or at least
bring back some of the useful pieces removed by this revert, such as
caching of opened relations and TupleTableSlots.

Any objections to removing it from master for now?

--
Thanks, Amit Langote