On Wed, 23 Sept 2026 at 23:41, Sami Imseih <[email protected]> wrote:
>
> > v2 fixes the issue you mentioned.
>
> Thanks for checking!
>
> > I think It also fixes another case. If the added column is in the replica
> > identity index, REPACK looks up the row with a NULL key. That fails the
> > Assert in find_target_tuple(), or gives "could not find target tuple"
> > without assertions. Adding this to the test covers it:
> >
> >     CREATE UNIQUE INDEX repack_test_i_c_idx ON repack_test (i, c);
> >     ALTER TABLE repack_test REPLICA IDENTITY USING INDEX 
> > repack_test_i_c_idx;
>
> ahh, good repro. It's worse for a pass-by-reference column, i.e. TEXT.
> find_target_tuple() only sets sk_argument, never SK_ISNULL, so the
> comparison function gets a NULL pointer rather than a NULL key.
> So a non assert build will actually segfault.
>
> v3 uses a TEXT column in the identity index, which covers the int case
> too since both go through the same sk_argument.
>
> --
> Sami Imseih
> Amazon Web Services (AWS)

Hi!
I reproduced this while stress testing REPACK (CONCURRENTLY) on HEAD.
Whats worse, this issue makes the database unrestorable via pg_dump/COPY.

I also checked DROP COLUMN, ALTER TYPE, SET STORAGE, SET COMPRESSION,
SET NOT NULL, ADD/DROP CONSTRAINT NOT VALID/VALIDATE, virtual
generated columns and found no related issue (both with and without
v3).

So patch LGTM


-- 
Best regards,
Kirill Reshke


Reply via email to