Refs #33482 -- Fixed QuerySet selectin and filtering againts negated Exists() with empty queryset.

[felixxm]

Regression in b7d1da5.

[charettes]

Strange, I thought that sql='' translated to an elided expression.

django/django/db/models/sql/where.py

Lines 85 to 87 in 25514b6

	if sql:
	result.append(sql)
	result_params.extend(params)

Here's the query I get from expressions.tests.ExistsTests.test_negated_empty_exists from non-Oracle backends

SELECT "expressions_manager"."id", "expressions_manager"."name" FROM "expressions_manager" WHERE "expressions_manager"."id" = 1

Do we need to adapt select_format to special case sql = ""? Probably, but I didn't think that select_format would be involved in this kind of query?

[felixxm]

Yes, but we have a different select_format on Oracle: "CASE WHEN {} THEN 1 ELSE 0 END".format(sql), so an empty string is translated to CASE WHEN THEN 1 ELSE 0 END.

[charettes]

I think we should adapt select_format in this case to handle '' to encapsulate the Oracle logic in one location.

It's not clear why select_format is called here though as we're directly filtering against the expression and not selecting it?

django/tests/expressions/tests.py

Lines 1916 to 1918 in 25514b6

	qs = Manager.objects.filter(
	~Exists(Manager.objects.none()) & Q(pk=manager.pk)
	)

[felixxm]

Ahh, sorry, it's not about select_format() 🤦 Exists() is wrapped with Case() also in the WHERE clause:

django/django/db/models/lookups.py

Lines 136 to 140 in 25514b6

	for expr in (self.lhs, self.rhs):
	if connection.ops.conditional_expression_supported_in_where_clause(expr):
	expr = Case(When(expr, then=True), default=False)
	wrapped = True
	exprs.append(expr)

[charettes]

Ahh I see, ok let's go with this approach for now. There's likely a good way to reconcile the multiple boolean expression handling we have for Oracle across the ORM but this simple change should go a long way.

[charettes]

Looking back at ticket-33482 I think that a better way forward would be to have it drop its custom support for negation and have __invert__ simply return ~Q(self) as the latter already has all the proper machinery to deal with negation.

This adjustment should do for now though!

[felixxm]

@charettes Thanks for checking 👍 I added an extra test for selecting negated Exists().

[felixxm]

buildbot, test on oracle.

[felixxm]

@charettes It seems that I found a regression in selecting negated Exists() with empty querysets as well 😄

[charettes]

Ah I assume something along SELECT as empty_exists_negated FROM table 😅 ?

[felixxm]

Ah I assume something along SELECT as empty_exists_negated FROM table sweat_smile ?

Yes :)