Inconsistencies around Composite Row nullness

On Sunday, November 2, 2025, Chris Hanks <christopher.m.hanks@gmail.com>
wrote:

Hello -

I've experienced some logically inconsistent query output on my local
Postgres instance, version string: PostgreSQL 18.0 (Homebrew) on
aarch64-apple-darwin25.0.0, compiled by Apple clang version 17.0.0
(clang-1700.3.19.1), 64-bit

I also reproduced it on the most recent Postgres version available at
db-fiddle.com, version string: PostgreSQL 17.0 on x86_64-pc-linux-gnu,
compiled by gcc (GCC) 11.4.1 20230605 (Red Hat 11.4.1-2), 64-bit

This first statement resolves, reasonably, to NULL:
SELECT ROW(NULL::integer, 2) = ROW(NULL::integer, 2)

Yes, ROW constructed values within an equality resolve using SQL row
constructor comparison rules.

This next statement resolves to ROW(NULL, 2):
SELECT coalesce(ROW(NULL::integer, 2), ROW(1, 2))

This statement is technically impossible - nothing resolves to “ROW(…)” -
the fact that ROW (a row constructor) is involved is erased when passing
the result of the expression through a function such that a plain
composite/record is produced. It is necessary, for the rest of the system
to function correctly, that records are comparable using (null equals null
=> true) semantics (is distinct; composite type comparison).

These final two statements each resolve to true, which is inconsistent
with the previous statements (each should resolve to NULL):
SELECT coalesce(ROW(NULL::integer, 2), ROW(1, 2)) = ROW(NULL::integer, 2)
SELECT coalesce(ROW(NULL::integer, 2)) = ROW(NULL::integer, 2)

See specifically the commentary in row constructor comparison 9.25.5 and
composite type comparison 9.25.6 in the documentation.

There is a patch to further expound/consolidate discussion on this topic
(null handling in PostgreSQL) presently awaiting committer attention.

David J.

Chris Hanks <christopher.m.hanks@gmail.com> writes:

I've experienced some logically inconsistent query output on my local
Postgres instance, version string: PostgreSQL 18.0 (Homebrew) on
aarch64-apple-darwin25.0.0, compiled by Apple clang version 17.0.0
(clang-1700.3.19.1), 64-bit

Yeah, it's not terribly consistent, but neither is the SQL standard
in this area. I believe what is happening in your first example
is that the construct "ROW(a, b, ...) = ROW(x, y, ...)" is being
broken down into "(a = x) AND (b = y) AND ...", from which you can
get a NULL result as described. However, the insertion of coalesce()
stops that decomposition from happening, and then what you get is the
behavior of the native composite-type comparators (record_eq and
friends). Those functions adhere to the btree requirement of
producing a total order of the datatype, ie null results are not OK,
so they report that ROW(NULL, 2) = ROW(NULL, 2) is true not null.

We could get rid of some of the inconsistency by eliminating that
special treatment of equality of two row-constructors, but I'm
afraid there would be complaints from people who were relying on
that behavior for optimization purposes.

regards, tom lane