[Bug]Vacuum full silently NULL out fast default columns

Started by SATYANARAYANA NARLAPURAMabout 6 hours ago2 messageshackers
Jump to latest
#1SATYANARAYANA NARLAPURAM
satyanarlapuram@gmail.com

Hi Hackers,

VACUUM FULL silently turns columns added via ALTER TABLE ... ADD COLUMN ...
DEFAULT <const> into NULL
on all pre-existing rows. The issue exists for other operations like
CLUSTER, REPACK.

Repro:

CREATE TABLE t (id int PRIMARY KEY);
INSERT INTO t SELECT generate_series(1,3);
ALTER TABLE t ADD COLUMN x int DEFAULT 42;
SELECT * FROM t; -- (1,42),(2,42),(3,42)
VACUUM FULL t;
SELECT * FROM t; -- (1,NULL),(2,NULL),(3,NULL)

If the column is NOT NULL, the value becomes the type's zero value
instead of NULL, silently bypassing both NOT NULL and any CHECK
constraint declared on it.

Root Cause: fast path in reform_tuple() in heapam_handler.c returns a copy
of the source tuple when no dropped columns need fixing up. The check
doesn't account for short tuples (HeapTupleHeaderGetNatts(t) <
relnatts) that rely on attmissingval to materialize the default. After
the rewrite, finish_heap_swap() calls RelationClearMissing(), clearing
the only source of those values, and the short tuples then read as
NULL.

Fix: force reform when the source tuple is shorter than the new tuple
descriptor.

Patch attached. Added a regression test in fast_default.sql covering
VACUUM FULL, CLUSTER, and REPACK on a table with fast-default columns
including a NOT NULL CHECK column.

Thanks,
Satya

Attachments:

0001-VACUUM-FULL-silently-NULL-out-fast-default-columns.patchapplication/octet-stream; name=0001-VACUUM-FULL-silently-NULL-out-fast-default-columns.patchDownload+63-2
#2SATYANARAYANA NARLAPURAM
satyanarlapuram@gmail.com
In reply to: SATYANARAYANA NARLAPURAM (#1)
Re: [Bug]Vacuum full silently NULL out fast default columns

Hi,

On Mon, May 4, 2026 at 6:40 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:

SATYANARAYANA NARLAPURAM <satyanarlapuram@gmail.com> writes:

VACUUM FULL silently turns columns added via ALTER TABLE ... ADD COLUMN

...

DEFAULT <const> into NULL
on all pre-existing rows. The issue exists for other operations like
CLUSTER, REPACK.

That is a seriously awful bug. Fortunately it is not in any shipping
release. A quick bisect run agrees that it broke here:

28d534e2ae0ac888b5460f977a10cd9bb017ef98 is the first bad commit
commit 28d534e2ae0ac888b5460f977a10cd9bb017ef98 (HEAD)
Author: Álvaro Herrera <alvherre@kurilemu.de>
Date: Mon Apr 6 21:55:08 2026 +0200

Add CONCURRENTLY option to REPACK

Patch attached. Added a regression test in fast_default.sql covering
VACUUM FULL, CLUSTER, and REPACK on a table with fast-default columns
including a NOT NULL CHECK column.

I don't know if this is the best code fix (I don't like putting extra
checks into a loop condition like this). But I agree we need some
more tests covering this area.

Thanks for reviewing! Attached v2 patch. Agreed, tried to optimize LOC in
V1. Before the change
loop was not breaking early, I fixed that as well in V2.

Thanks,
Satya

Attachments:

v2-0001-VACUUM-FULL-silently-NULL-out-fast-default-columns.patchapplication/octet-stream; name=v2-0001-VACUUM-FULL-silently-NULL-out-fast-default-columns.patchDownload+72-5