Re: Proposal: Conflict log history table for Logical Replication

On Tue, Oct 13, 2026 at 9:39 AM shveta malik <shveta(dot)malik(at)gmail(dot)com> wrote:
>
> On Mon, Oct 12, 2026 at 11:31 AM shveta malik <shveta(dot)malik(at)gmail(dot)com> wrote:
> >
> > On Mon, Oct 12, 2026 at 6:04 AM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > >
> > > On Fri, Oct 9, 2026 at 1:43 PM shveta malik <shveta(dot)malik(at)gmail(dot)com> wrote:
> > > >
> > > > Another concern on v79 is:
> > > >
> > > > I noticed that one CLT row contains three timestamp representations
> > > > with three different behaviors:
> > > >
> > > > 1) replica_identity is frozen in the apply worker’s DateStyle.
> > > > Session's Datestyle has no impact on it.
> > > > 2) remote_commit_ts is represented in the reader’s (session's) DateStyle, and
> > > > 3) local_conflicts.commit_ts is always ISO.
> > > >
> > > > This may cause some conufsion. Please see the results below. I can
> > > > share the testcase if needed:
> > > >
> > > > Conflicts:
> > > >
> > > > postgres=# select conflict_type, remote_commit_ts, replica_identity,
> > > > local_conflicts from pg_conflict.pg_conflict_log_16390;
> > > > conflict_type | remote_commit_ts |
> > > > replica_identity |
> > > > local_conflicts
> > > > -----------------------+--------------------------------+-----------------------------------+------------------------------------------------------------------------------------------
> > > > delete_origin_differs | 09/10/2026 14:15:29.055423 IST |
> > > > {"k":"2025-12-25 10:00:00+05:30"} |
> > > > {"{\"xid\":\"672\",\"commit_ts\":\"2026-10-09T14:15:22.033391+05:30\",\"origin\":null}"}
> > > > delete_origin_differs | 09/10/2026 14:17:12.929024 IST |
> > > > {"k":"02/01/2025 08:00:00 IST"} |
> > > > {"{\"xid\":\"678\",\"commit_ts\":\"2026-10-09T14:17:09.326423+05:30\",\"origin\":null}"}
> > > > (2 rows)
> > > >
> > > >
> > > > 1) First conflict row was inserted with default Datastyle of DB and
> > > > the apply worker which is:
> > > > postgres=# show DateStyle;
> > > > DateStyle
> > > > -----------
> > > > ISO, DMY
> > > >
> > > > 2) Before inserting second conflict, below was done.
> > > >
> > > > ALTER DATABASE postgres SET DateStyle = 'SQL, DMY';
> > > > ALTER SUBSCRIPTION sub1 DISABLE; ALTER SUBSCRIPTION sub1 ENABLE;
> > > >
> > > > (So Above conflict data is read when session's format was also 'SQL,
> > > > DMY', thus we see remote_commit_ts in SQL format)
> > > > ----
> > > >
> > > > We can see the differnce in 'replica_identity' values in 2 rows, while
> > > > remote_commit_ts and local_conflicts.commit_ts remain same. Second RI
> > > > is 02/01/2025 in DMY format, meaning 2nd january.
> > > >
> > > > 3)
> > > > Now if I do this:
> > > > SET DateStyle = 'ISO, MDY'
> > > >
> > > > postgres=# select conflict_type, remote_commit_ts, replica_identity,
> > > > local_conflicts from pg_conflict.pg_conflict_log_16390;
> > > > conflict_type | remote_commit_ts |
> > > > replica_identity |
> > > > local_conflicts
> > > > -----------------------+----------------------------------+-----------------------------------+------------------------------------------------------------------------------------------
> > > > delete_origin_differs | 2026-10-09 14:15:29.055423+05:30 |
> > > > {"k":"2025-12-25 10:00:00+05:30"} |
> > > > {"{\"xid\":\"672\",\"commit_ts\":\"2026-10-09T14:15:22.033391+05:30\",\"origin\":null}"}
> > > > delete_origin_differs | 2026-10-09 14:17:12.929024+05:30 |
> > > > {"k":"02/01/2025 08:00:00 IST"} |
> > > > {"{\"xid\":\"678\",\"commit_ts\":\"2026-10-09T14:17:09.326423+05:30\",\"origin\":null}"}
> > > > (2 rows)
> > > >
> > > > See remote_commit_ts changed from previous SQL to ISO, while
> > > > replica_identity and local_conflicts.commit_ts are the same.
> > > >
> > > > If I try to read it as documented (casting it back to original
> > > > column's type), I get misleading results:
> > > >
> > > > postgres=# SET DateStyle = 'ISO, MDY';
> > > >
> > > > postgres=#
> > > > SELECT (local_conflicts[1]->>'commit_ts')::timestamptz as local_ts,
> > > > (replica_identity->>'k')::timestamptz as ri_ts
> > > > FROM pg_conflict.pg_conflict_log_16390;
> > > >
> > > > local_ts | ri_ts
> > > > ----------------------------------+---------------------------
> > > > 2026-10-09 14:15:22.033391+05:30 | 2025-12-25 10:00:00+05:30
> > > > 2026-10-09 14:17:09.326423+05:30 | 2025-02-01 08:00:00+05:30
> > > > (2 rows)
> > > >
> > > > See second row for ri_ts, it will now represent '1st Feb' rather than
> > > > '2nd January'.
> > > >
> > > > Shouldn't all three timestamps be in the same format?
> > > >
> > > > IMO, remote_commit_ts being session-dependent is expected.
> > > > The concern is the other two JSON text columns. replica_identity
> > > > stores the worker’s DateStyle, so the ->>'k'::timestamptz can
> > > > silently return a different value (here 2nd Jan read back as 1st Feb);
> > > > while local_conflicts.commit_ts is always ISO. Shouldn’t the recorded
> > > > JSON timestamps use a single, stable format (ISO) so they’re
> > > > unambiguous and safe to cast back?
> > > >
> > >
> > > What are the values if you don't change the Datestyle manually and use
> > > the default or what was used by worker while reading it?
> > >
> >
> >
> > With default, values are:
> >
> > postgres=# SELECT conflict_type,
> > remote_commit_ts,
> > replica_identity->>'k' AS ri_key_timestamp,
> > local_conflicts[1]->>'commit_ts' AS lc_commit_ts
> > FROM pg_conflict.pg_conflict_log_16390;
> >
> > conflict_type | remote_commit_ts |
> > ri_key_timestamp | lc_commit_ts
> > -----------------------+----------------------------------+---------------------------+----------------------------------
> > delete_origin_differs | 2026-10-12 10:47:07.349306+05:30 | 2025-12-25
> > 10:00:00+05:30 | 2026-10-12T10:47:04.618018+05:30
> > delete_origin_differs | 2026-10-12 10:48:22.419669+05:30 | 2025-01-02
> > 08:00:00+05:30 | 2026-10-12T10:48:20.277719+05:30
> > (2 rows)
> >
> > Default DateStyle of Db:
> >
> > postgres=# show DateStyle;
> > DateStyle
> > -----------
> > ISO, DMY
> > (1 row)
> >
> > Same is used by worker while insert and in same we have read the above data.
> >
> > ~~
> >
> > So with the default (ISO), or whenever the reader’s DateStyle matches
> > the one the worker used at write time, the values read back correctly;
> > the misleading result only appears when the reader’s DateStyle differs
> > from the worker’s write-time DateStyle (specifically when the worker
> > wrote under a non-ISO style).
> >
> > ~~
> >
> > Further observations and analysis:
> >
> > 1) Apply worker uses current datestyle of DB for replica_identity, and
> > then materializes it into text (json col).
> >
> > 2) local_conflicts always uses ISO because it’s built via row_to_json.
> > See JsonEncodeDateTime() → EncodeDateTime(), which is called with
> > USE_XSD_DATES; in EncodeDateTime() that falls through the same branch
> > as USE_ISO_DATES.
> >
> > 3) remote_commit_ts is stored as timestamp column rather than text in json.
> >
> > The one stored as timestamptz (3) will have no issues. Because it’s a
> > genuine timestamptz, we always get back the correct instant; only its
> > display changes with the reader’s DateStyle.
> >
> > Both (1) and (2) are frozen as text, but IIUC, only replica_identity
> > is at risk.
> >
> > local_conflicts.commit_ts avoids it by alwyas using ISO (written
> > always as YYYY-MM-DD, where month and day are in fixed positions. Thus
> > no reader’s DMY/MDY setting can swap them.
> >
> > replica_identity is the only one that stores the worker’s DateStyle
> > into stored text (unknown and un-recoverable later). Different rows
> > can be frozen under different styles (config change), so a reader has
> > no reliable style to pick while reading.
> >
> > The fix could be to have the apply worker also use ISO style while
> > inserting; similar to row_to_json, and how logical replication’s
> > libpqwalreceiver standardizes the wire timestamp values using ISO. See
> > libpqrcv_connect using datestyle=ISO.
> >
>
> If we want to address this, I have attached the patch for a potential
> solution. I have also attached the testcase.
>
> In the worker, we can change the config to datestyle=ISO before
> building the JSON values.
>
> --NewGUCNestLevel() opens a new GUC nesting level and returns its
> number. It does not change any setting.
> --SetConfigOption() then changes the GUC (datestyle=ISO).
> --AtEOXact_GUC() restores the GUC stack back to that marked point. In
> case of error, AtEOXact_GUC() will be called by Abort-Txn.

I have reviewed this change, IMHO there are slight error, basically,
SetConfigOption() does not restore the old value at AtEOXact_GUC(true,
save_nestlevel). SetConfigOption calls set_config_option(...,
GUC_ACTION_SET, ...), which records the stack entry with state GUC_SET
rather than GUC_SAVE. When AtEOXact_GUC(true, save_nestlevel) is
called with isCommit = true: It only restores the prior value when
stack->state == GUC_SAVE not GUC_SET. So for fixing this correctly we
need to directly call set_config_option() with GUC_ACTION_SAVE
instead of calling SetConfigOption() as shown below.

save_nestlevel = NewGUCNestLevel();
(void) set_config_option("datestyle", "ISO",
PGC_USERSET, PGC_S_SESSION,
GUC_ACTION_SAVE, true, 0, false);
.....
AtEOXact_GUC(true, save_nestlevel);

--
Regards,
Dilip Kumar
Google