(record = record) inconsistent with record_eq(arg1, arg2)

On Monday, July 31, 2023, jian he <jian.universality@gmail.com> wrote:

hi.
pg15: https://dbfiddle.uk/3_GxPdj4
not sure if it's a bug or I misunderstood.

select row(1,null::int) = (1,1::int); -- return null
select record_eq(row(1,null::int),(1,1::int)); --return false.

This isn’t a bug though I can’t point to the exact code where the
difference manifests.

The SQL standard mandates the query behavior of the equals operator while
internal indexing and hash details require that the equality function not
return null.

Calling record_eq directly in SQL is unsupported so the claim that the
second query is buggy is incorrect altogether. We don’t promise any
particular result to the user in that situation, and could indeed remove
the function and the query could produce an error and we’d be technically
bug-free.

David J.

"David G. Johnston" <david.g.johnston@gmail.com> writes:

On Monday, July 31, 2023, jian he <jian.universality@gmail.com> wrote:

select row(1,null::int) = (1,1::int); -- return null
select record_eq(row(1,null::int),(1,1::int)); --return false.

This isn’t a bug though I can’t point to the exact code where the
difference manifests.

Yeah, the difference in behavior is deliberate. See

https://www.postgresql.org/docs/current/functions-comparisons.html#ROW-WISE-COMPARISON

which documents the behavior of ROW(...) = ROW(...), and

https://www.postgresql.org/docs/current/functions-comparisons.html#COMPOSITE-TYPE-COMPARISON

which documents the behavior of record comparisons that are written
in other ways, and explains the reason for the difference. The
SQL standard governs the behavior of "ROW(...) = ROW(...)", but last
I looked they don't really have a notion of composite-valued scalars,
so the spec has nothing to say about what "row-valued-expression =
row-valued-expression" should do. Even if it did, we would not
change this, because it's too important to have a total ordering
for composite columns.

regards, tom lane