Query.as_sql() has side effects for select_related queries

[Erin Kelly]

Each time Query.as_sql() is called on a select_related query, the list of select columns grows again. Presumably Query.pre_sql_setup() should only get called once? For example:

In [39]: qs = Species.objects.all().select_related(depth=1)

In [40]: qs.query.as_sql()
Out[40]:
('SELECT "TEST_SPECIES"."ID", "TEST_SPECIES"."NAME", "TEST_SPECIES"."GENUS_ID", "TEST_GENUS"."ID", "TEST_GENUS"."NAME", "TEST_GENUS"."FAMILY_ID" FROM "TEST_SPECIES" INNER JOIN "TEST_GENUS" ON ("TEST_SPECIES"."GENUS_ID" = "TEST_GENUS"."ID")',
 ())

In [41]: qs.query.as_sql()
Out[41]:
('SELECT "TEST_SPECIES"."ID", "TEST_SPECIES"."NAME", "TEST_SPECIES"."GENUS_ID", "TEST_GENUS"."ID", "TEST_GENUS"."NAME", "TEST_GENUS"."FAMILY_ID", "TEST_SPECIES"."ID", "TEST_SPECIES"."NAME", "TEST_SPECIES"."GENUS_ID", "TEST_GENUS"."ID", "TEST_GENUS"."NAME", "TEST_GENUS"."FAMILY_ID" FROM "TEST_SPECIES" INNER JOIN "TEST_GENUS" ON ("TEST_SPECIES"."GENUS_ID" = "TEST_GENUS"."ID")',
 ())

[Malcolm Tredinnick]

We can't really arrange to call pre_sql_setup() just once, because if we did that and then added something to the query set, it could change the required select columns and we wouldn't know how to undo the previous select modifications from pre_sql_setup(). So the sequence

qs = ....
qs.query.as_sql()
qs = qs.select_related(depth=2)
qs.query.as_sql()

and similar usage patterns wouldn't be correct.

I've fixed this by keeping the extra select columns separate until we construct the query in as_sql().

[Malcolm Tredinnick]

(In [7445]) queryset-refactor: Removed an unwanted side-effect from the Query.as_sql() method.

Fixed #7056