Hi Shihao,

I re-ran the reproduction from my earlier mail against v6 (REL_18_STABLE
4d6ff8a051 + v6-0001..0003): the standby runs the unpatched summarizer,
then the promoted server is restarted on 18 + v6.

The timeline 1 summary is the same as before (0/02000028-0/030051F0, ending
at the switch point 0/30051F0). The incremental backup and pg_combinebackup
succeed, and the restored table now matches the promoted server (1|0 2|0
3|0 4|1), where v5 gave 4|0.

With v6-0002's check changed back to `ws->end_lsn > tli_end_lsn`, the
restored table is stale again (4|0), so the `>=` is what fixes it. The
control case, with v6 binaries on the standby from the start, also restores
correctly.

I didn't run the new
https://www.google.com/url?q=http://015_summary_past_switch.pl&source=gmail&ust=1795093490688000&sa=E
in this build.

Thanks,
Sai Pranav

On Tue, Nov 17, 2026 07:07 AM, shihao zhong <[email protected]> wrote:

> Hi Zsolt and Raja,
>
> Thanks for the review! I appreciate it!
>
> In v6, 0002 also checks a summary that ends at the switch point. It
> no longer raises an error. Instead it backs up any relation fork that
> such a summary truncates in full. That only costs space in an
> incremental backup whose prior backup is older than the promotion.
>
> Thanks,
> Shihao
>

Reply via email to