Exists subquery ignores ordering when it's distinct

[Kevin Marsh]

I ran into a bug in the Exists expression, I'm using it to do some filtering where I just want the latest object grouped by a certain value (eg. latest financial statement object for each company). See the patch for a failing test, but using the Manager/Employee models from the test suite you can see in the slightly contrived example that for an Exists query like

# ... filtering by highest paid employee per manager
Exists(
  Employee.objects.order_by("manager", "-salary").distinct("manager").filter(pk=OuterRef("pk"))
)
# ...

Gets transformed into SQL (Postgresql) like this where the ordering has been stripped out

-- ...
EXISTS(
 SELECT DISTINCT ON (U0."manager_id")
        1 AS "a"
 FROM "expressions_employee" U0
 WHERE U0."id" = ("expressions_employee"."id")
 -- Missing ordering which is required for distinct
 LIMIT 1
)
-- ...

---

Obviously we want to call clear_ordering most of the time on Exists subqueries for performance reasons, but in this case we either shouldn't clear them or loudly raise an exception saying that distinct Exists queries are not supported

[Kevin Marsh]

Failing test

[Simon Charette]

The report makes sense. We could go two ways here

1. Clear distinct_fields as well as it's unnecessary (we only care about the existence of a single row after all)
2. Avoid ​forcing the removal of ordering since the clear_ordering method ​is a noop when distinct fields are defined.

[Simon Charette]

Actually something is off in your test here. Even if the ordering is not cleared all employees will match because your reduce the number of row by employee

e.g.

SELECT "expressions_employee"."id",
       "expressions_employee"."firstname",
       "expressions_employee"."lastname",
       "expressions_employee"."salary",
       "expressions_employee"."manager_id",
       "expressions_employee"."based_in_eu"
FROM "expressions_employee"
WHERE (EXISTS
         (SELECT DISTINCT ON (U0."manager_id") 1 AS "a"
          FROM "expressions_employee" U0
          WHERE (U0."manager_id" IS NOT NULL
                 AND U0."id" = ("expressions_employee"."id"))
          ORDER BY U0."manager_id" ASC, U0."salary" DESC
          LIMIT 1)
       AND "expressions_employee"."manager_id" = 2)
LIMIT 21;

In other words, your distinct group is by U0."id" = ("expressions_employee"."id") so you'll always have matching rows even if the ordering is maintained. Try it out for yourself with the following patch

django/db/models/sql/query.py

@@ -648 +648 @@
                 combined_query.exists(limit=False)
                 for combined_query in q.combined_queries
             )
-        q.clear_ordering(force=True)
+        q.clear_ordering()
         if limit:
             q.set_limits(high=1)
         q.add_annotation(Value(1), "a")