> On 30 Jul 2026, at 19:44, Kirill Reshke <[email protected]> wrote:
>
> <v8-0003-Guard-against-configuration-hazards-in-archive_mo.patch><v8-0002-Add-primary_last_archived-column-to-pg_stat_wal_r.patch>
Hi Kirill,
Thanks for v8!
The fallback currently treats the absence of reports for three
archive_status_report_interval periods as a failure, but the walsender
sends nothing when last_archived_wal has not changed. Thus the first
segment after an idle period can be archived locally immediately,
racing with the primary.
I suggest using two states:
* before the first report, use archive_status_report_interval to detect
that the upstream cannot provide reports;
* after the first report, start a longer archive_timeout-based timer
only when an unconfirmed .ready segment appears.
The latter timer must start when there is work awaiting confirmation,
not at the last report. We also need to decide whether it uses the
standby's archive_timeout or a value reported by the upstream, and
allow time for archive_command and report delivery.
Report scheduling also needs adjustment. WalSndArchivalReport() is
called after early returns in WalSndKeepaliveIfNecessary(), so
wal_sender_timeout=0 disables reports. The report deadline is not
considered by WalSndComputeSleeptime() either.
ProcessArchivalReport() also marks the segment containing an ancestor
timeline's switchpoint as done. A descendant report proves that its
common prefix exists under the descendant filename, but I don't see
what guarantees that the ancestor filename is present in the archive.
This needs either an explanation or an archive-only recovery test with
recovery_target_timeline set to the ancestor.
Some smaller items remain: the new GUC and view column are missing from
config.sgml and monitoring.sgml; changing
archive_status_report_interval by SIGHUP does not renegotiate the
existing connection; and the Windows archive paths in 055-057 are still
not escaped as in enable_archiving() [0]. There is also a duplicated
condition in ProcessArchivalReport(), a misspelled SGML ID, and trailing
whitespace.
Best regards, Andrey Borodin.
[0] https://github.com/x4m/postgres_g/actions/runs/30611537052/job/91095599886
