Re: Proposal: Conflict log history table for Logical Replication

On Wed, 27 May 2026 at 14:04, vignesh C <vignesh21(at)gmail(dot)com> wrote:
>
>
> I have fixed the rest of the comments. The attached v41 version patch
> has the changes for the same. Additionally the comments from [1] have
> also been fixed.

I was evaluating whether the existing pg_upgrade changes for conflict
log tables can handle the addition of new columns in a future release.
To validate this, I performed the following:
Added two new columns to the conflict log table:
v20_new_col1 TEXT
v20_new_col2 TEXT

These changes are present in patch '0001'.

For adding new columns during binary upgrade, the following
version-specific logic is required in 'pg_dump':
ALTER TABLE pg_conflict.pg_conflict_log_for_subid_oid
ADD COLUMN v20_new_col1 TEXT;

ALTER TABLE pg_conflict.pg_conflict_log_for_subid_oid
ADD COLUMN v20_new_col2 TEXT;

These changes are included in patch '0001'.
One important point here is that when 'ALTER TABLE ... ADD COLUMN' is
run, the server does not rewrite existing rows on disk. Instead, it
only updates the system catalog with the new column metadata.

While selecting data from the table, the server handles this as follows:
1. Deform what is physically present - 'slot_deform_heap_tuple()'
reads the raw tuple bytes from disk, but only up to 't_natts', which
is the number of columns recorded in the tuple header at the time that
row was inserted. It stops there because the tuple has no physical
data for columns added later.
2. Fill in what is missing - After deforming the tuple, if the
number of populated columns is still less than the number of columns
requested by the query, it calls 'slot_getmissingattrs()' to cover the
gap. Since the new columns were added with no default value,
'slot_getmissingattrs()' sets:
tts_isnull[attnum] = true;

This is how NULL is returned for the newly added columns in existing rows.

These changes were tested on a new server with the v40 version patch +
'0001' patch.
1. Pre-upgrade state using v40 version patches
Simulated conflicts using a setup where the schema does not include
the new columns:
postgres=# select * from pg_conflict.pg_conflict_log_for_subid_16396 ;
....
(4 rows)

2. Upgrade using 'pg_upgrade'
The upgrade was performed on a cluster initialized with patches v40 +
'0001', and it completed successfully.
Post-upgrade verification:
postgres=# select conflict_type, v20_new_col1, v20_new_col2 from
pg_conflict.pg_conflict_log_for_subid_16396 ;
conflict_type | v20_new_col1 | v20_new_col2
---------------+--------------+--------------
insert_exists | |
insert_exists | |
insert_exists | |
insert_exists | |
(4 rows)

Existing rows were preserved, and the newly added columns are visible
and populated with NULLs, as expected.

3. Post-upgrade conflict insertion
After starting the old publisher again to continue generating conflicts:
postgres=# select conflict_type, v20_new_col1, v20_new_col2 from
pg_conflict.pg_conflict_log_for_subid_16396 ;
conflict_type | v20_new_col1 | v20_new_col2
---------------+--------------+--------------
insert_exists | |
insert_exists | |
insert_exists | |
insert_exists | |
insert_exists | v20_new_col1 | v20_new_col2
insert_exists | v20_new_col1 | v20_new_col2
insert_exists | v20_new_col1 | v20_new_col2
(7 rows)

New conflicts are inserted successfully, and the newly added columns
are correctly populated for new entries.

Based on this testing, the current 'pg_upgrade' framework, along with
the additional dump-time adjustments, appears sufficient to support
schema evolution of conflict log tables, specifically for adding new
columns in future releases.

Thoughts?

Regards,
Vignesh