Re: Slow catchup of 2PC (twophase) transactions on replica in LR

On Tue, Apr 16, 2024 at 4:25 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:

>
> >
>
> Can you please once consider the idea shared by me at [1] (One naive
> idea is that on the publisher .....) to solve this problem?
>
> [1] -
> https://www.postgresql.org/message-id/CAA4eK1K1fSkeK%3Dkc26G5cq87vQG4%3D1qs_b%2Bno4%2Bep654SeBy1w%40mail.gmail.com
>
>
>
Expanding on Amit's idea, we found out that there is already a mechanism in
code to fully decode prepared transactions prior to a defined LSN where
two_phase is enabled using the "two_phase_at" LSN in the slot. Look at
ReorderBufferFinishPrepared() on how this is done. This code was not
working as expected in our patch because
we were setting two_phase on the slot to true as soon as the alter command
was received. This was not the correct way, initially when two_phase is
enabled, the two_phase changes to pending state and two_phase option on the
slot should only be set to true when two_phase moves from pending to
enabled. This will happen once the replication is restarted with two_phase
option. Look at code in CreateDecodingContext() on how "two_phase_at" is
set in the slot when done this way. So we changed the code to not remotely
alter two_phase when toggling from false to true. With this change, now
even if there are pending transactions on the publisher when toggling
two_phase from false to true, these pending transactions will be fully
decoded and sent once the commit prepared is decoded as the pending
prepared transactions are prior to the "two_phase_at" LSN. With this patch,
now we are able to handle both pending prepared transactions when altering
two_phase from true to false as well as false to true.

Attaching the patch for your review and comments. Big thanks to Kuroda-san
for also working on the patch.

regards,
Ajin Cherian
Fujitsu Australia.