Re: PoC: Simplify recovery after dropping a table by LOGGING the restore LSN

Hi Jim,

Thanks for the review.

On Thu, Oct 29, 2026 at 6:24 PM Jim Jones <jim(dot)jones(at)uni-muenster(dot)de> wrote:
> == callback functions ==
>
> The comments of RegisterXactCallback and RegisterSubXactCallback suggest
> they're meant for extensions: "These functions are intended for use by
> dynamically loaded modules.". Are they really the best fit here?
>
I removed the RegisterXactCallback and RegisterSubXactCallback calls
and replaced them with AtEOSubXact_DropOrTruncateTable and
AtEOXact_DropOrTruncateTable. They are now called directly from
xact.c, next to AtEOXact_on_commit_actions and
AtEOSubXact_on_commit_actions, which already handle ON COMMIT actions
this way.

> == RegisterDropOrTruncateTable ==
>
> The function already expects reloid, so it could look for rel_name and
> rel_namespace itself if needed -- currently the caller needs to do it.
> Have you considered simply (relid, is_truncate)?
>
Done. It now takes (relid, is_truncate) and looks up the names itself.

> == doDeletion ==
>
> Is the relname != NULL guard and the "unknown" schemaname really
> necessary? I imagined that at this point the relation is locked and must
> exist. Also, IMHO silently skipping a log entry or just logging a
> placeholder is arguably not a good idea for this feature.
>
Agreed. I removed the relname check and the "unknown" schema name.

> == prepared transactions ==
>
> A simplistic approach would be just to make it explicit in the docs and
> not log anything.
>
Agreed. Nothing is logged for prepared transactions now, and the docs say so.

While testing I also found that TRUNCATE of a partitioned table with
no partitions hit an Assert when wal_level is minimal, because no
commit record is written. Fixed in this version.

v12 attached.

Best,
Salma Elsayed