Re: Fix premature xmin advancement during fast forward decoding

On Tue, Apr 22, 2025 at 12:36 PM Zhijie Hou (Fujitsu)
<houzj.fnst@fujitsu.com> wrote:
>
> Hi,
>
> To fix this, I think we can allow the base snapshot to be built during fast
> forward decoding, as implemented in the patch 0001 (We already built base
> snapshot in fast-forward mode for logical message in logicalmsg_decode()).

The idea and code looks okay to me and the performance impact is also
not that huge.

IIUC, fast_forward decoding mode is used only in two cases. 1)
pg_replication_slot_advance and 2) in upgrade flow to check if there
are any pending WAL changes which are not yet replicated. See
'binary_upgrade_logical_slot_has_caught_up'-->'LogicalReplicationSlotHasPendingWal'.
It seems like this change will not have any negative impact in the
upgrade flow as well (in terms of performance and anything else).
Thoughts?

>
> Moreover, I conducted a basic test[2] to test the patch's impact, noting that
> advancing the slot incurs roughly a 4% increase in processing time after
> applying the patch, which appears to be acceptable. Additionally, the cost
> associated with building the snapshot via SnapBuildBuildSnapshot() did not show
> up in the profile. Therefore, I think it's reasonable to refrain from
> further optimization at this stage.

I agree on this.

thanks
Shveta