Re: Fix premature xmin advancement during fast forward decoding - Mailing list pgsql-hackers

On Sat, Apr 26, 2025 at 12:01 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
>
> On Fri, Apr 25, 2025 at 4:42 AM Amit Kapila <amit.kapila16@gmail.com> wrote:
> >
> > On Fri, Apr 25, 2025 at 10:46 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
> > >
> > > What I'm concerned about is the back branches. With this approach all
> > > back branches will have such degradations and it doesn't make sense to
> > > me to optimize SnapBuildCommitTxn() codes in back branches.
> > >
> >
> > One possibility could be that instead of maintaining an entire
> > snapshot in fast_forward mode, we can maintain snapshot's xmin in each
> > ReorderBufferTXN. But then also, how would we get the minimum
> > txns_by_base_snapshot_lsn as we are getting now in
> > ReorderBufferGetOldestXmin? I think we need to traverse the entire
> > list of txns to get it in fast_forward mode but that may not show up
> > because it will not be done for each transaction. We can try such a
> > thing, but it won't be clean to have fast_forward specific code and
> > also it would be better to add such things only for HEAD.
>
> Agreed.
>

We may need to consider handling xmin in SnapBuildCommitTxn(), where
we also set the base snapshot.

> > Can you think of any better ideas?
>
> No idea. Hmm, there seems no reasonable way to fix this issue for back
> branches. I consented to the view that these costs were something that
> we should have paid from the beginning.
>

Right, I feel we should go with the simple change proposed by Hou-San
for now to fix the bug. If, in the future, we encounter any cases
where such optimizations can help for fast-forward mode, then we can
consider it. Does that sound reasonable to you?

--
With Regards,
Amit Kapila.