>>>>> "Pavel" == Pavel Stehule <pavel.stehule@gmail.com> writes:
Pavel> I am strongly sure, so current default is best and any change of
Pavel> this behave (it is simply - just use #option) is strongly wrong.
>> I don't buy it; I call this a bug.
>>
>> Here's why: in an ON CONFLICT (col) clause, the (col) is not a list of
>> expressions or even really a list of columns, what it is is an index
>> definition (i.e. the same thing that would appear in CREATE INDEX). One
>> consequence of this is that _qualified_ column names, which are a usual
>> solution to variable name vs column conflicts, are not allowed here.
>> There is already special processing done on the clause for this reason
>> (the hiding of other tables that might be visible at this point in the
>> query), and I would say that this simply doesn't go far enough and that
>> parameters should be hidden too (by suppressing the columnref hooks
>> while the arbiter clause is being analyzed).
>>
Pavel> If there a expressions are not accepted there, then I can accept
Pavel> your argumentation. Second hand is a implementation.
Expressions are allowed there on the same basis that they are allowed in
CREATE INDEX (i.e. with extra parens) - but they can still only refer to
columns of the specific table for which we're trying to identify an
arbiter index, not to anything whatsoever from the ordinary context of
the query.
If I were writing a patch to change this, I would expand on this comment:
/*
* While we process the arbiter expressions, accept only non-qualified
* references to the target table. Hide any other relations.
*/
to say something like:
/*
* While we process the arbiter expressions, accept only non-qualified
* references to the target table. Hide any other relations, and also
* suppress the columnref hooks. This is justified by the fact that
* these expressions and column references are not part of query
* execution, but rather are a way to identify a constraint. As such,
* the parser environment has no business trying to change the
* interpretation of column names here (just as it would have no
* business changing a column name list in the INSERT proper).
*/
--
Andrew (irc:RhodiumToad)
