"David G. Johnston" <david.g.johnston@gmail.com> writes:
> On Tuesday, January 23, 2024, PG Bug reporting form <noreply@postgresql.org>
> wrote:
>> create table t(i int);
>>
>> explain (costs off) select * from t join t tt on t.xmin = tt.xmin;
>> QUERY PLAN
>> ---------------------------------
>> Hash Join
>> Hash Cond: (t.xmin = tt.xmin)
>> -> Seq Scan on t
>> -> Hash
>> -> Seq Scan on t tt
>> (5 rows)
>>
>> explain (costs off) select * from t join t tt using (xmin);
>> ERROR: column "xmin" specified in USING clause does not exist in left
>> table
> I don’t this being worth the effort to change, and really seems like
> completely expected behavior. “Select *” doesn’t output xmin, it requires
> explicit table qualification to see it. This is the same thing.
Well, it is odd that "using (xmin)" isn't equivalent to the allegedly
equivalent "on t.xmin = tt.xmin". This is down to the infrastructure
in transformFromClauseItem(), which searches the lists of (regular,
non-system) relation output column names to expand USING(). But like
you, I can't get excited about changing it. There are a couple of
practical reasons why not:
* NATURAL JOIN is defined in terms of USING. But we *certainly* don't
want "x NATURAL JOIN y" deciding that it should equate all the system
columns of x to those of y. So there's going to be inconsistency at
one level or the other no matter what.
* I really find it hard to imagine a valid use case for joining on any
system column. There are use-cases for joining on TID in an UPDATE
involving a self-join to the target table; but you can't write that
with JOIN USING syntax.
regards, tom lane
