Django 4.2 casts text fields when testing IS NULL, preventing use of partial indexes

[Alex Vandiver]

The Zulip project has a model with unique constraints which are expressed as two non-overlapping partial indexes:

class UserCount(models.Model):
    user = models.ForeignKey(UserProfile, on_delete=models.CASCADE)
    realm = models.ForeignKey(Realm, on_delete=models.CASCADE)
    property = models.CharField(max_length=32)
    subgroup = models.CharField(max_length=16, null=True)
    end_time = models.DateTimeField()
    value = models.BigIntegerField()

    class Meta:
        constraints = [
            UniqueConstraint(
                fields=["user", "property", "subgroup", "end_time"],
                condition=Q(subgroup__isnull=False),
                name="unique_user_count",
            ),
            UniqueConstraint(
                fields=["user", "property", "end_time"],
                condition=Q(subgroup__isnull=True),
                name="unique_user_count_null_subgroup",
            ),
        ]

However, since commit 09ffc5c1212d4ced58b708cbbf3dfbfb77b782ca in Django 4.2, a query of the form:

UserCount.objects.get(
    property="messages_read::hour",
    subgroup=None,
    end_time=timezone_now(),
    user_id=user_profile.id,
    realm_id=realm.id,
)

...generates this SQL:

SELECT "analytics_usercount"."id", "analytics_usercount"."property", "analytics_usercount"."subgroup", "analytics_usercount"."end_time", "analytics_usercount"."value", "analytics_usercount"."user_id", "analytics_usercount"."realm_id" FROM "analytics_usercount" WHERE ("analytics_usercount"."end_time" = '2023-09-13T19:16:34.195355+00:00'::timestamptz AND "analytics_usercount"."property" = 'messages_read::hour' AND "analytics_usercount"."realm_id" = 4715 AND "analytics_usercount"."subgroup"::text IS NULL AND "analytics_usercount"."user_id" = 428054) LIMIT 21

The cast of "analytics_usercount"."subgroup"::text IS NULL causes PostgreSQL to not be able to use the unique partial index:

 Limit  (cost=48.30..49.42 rows=1 width=61)
   ->  Bitmap Heap Scan on analytics_usercount  (cost=48.30..49.42 rows=1 width=61)
         Recheck Cond: (((property)::text = 'messages_read::hour'::text) AND (realm_id = 4715) AND (end_time = '2023-09-13 19:00:00+00'::timestamp with time zone) AND (user_id = 428054))
         Filter: ((subgroup)::text IS NULL)
         ->  BitmapAnd  (cost=48.30..48.30 rows=1 width=0)
               ->  Bitmap Index Scan on analytics_usercount_property_591dbec1_idx  (cost=0.00..4.88 rows=158 width=0)
                     Index Cond: (((property)::text = 'messages_read::hour'::text) AND (realm_id = 4715) AND (end_time = '2023-09-13 19:00:00+00'::timestamp with time zone))
               ->  Bitmap Index Scan on analytics_usercount_e8701ad4  (cost=0.00..43.17 rows=3626 width=0)
                     Index Cond: (user_id = 428054)

Dropping the explicit cast causes it to use the index:

 Limit  (cost=0.57..2.80 rows=1 width=61)
   ->  Index Scan using unique_user_count_null_subgroup on analytics_usercount  (cost=0.57..2.80 rows=1 width=61)
         Index Cond: ((user_id = 428054) AND ((property)::text = 'messages_read::hour'::text) AND (end_time = '2023-09-13 19:00:00+00'::timestamp with time zone))
         Filter: (realm_id = 4715)

...and improving the query runtime significantly.

It's not clear to me from 09ffc5c1212d4ced58b708cbbf3dfbfb77b782ca what about psycopg3 requires this cast.

[Simon Charette]

We should try to avoid casts if possible (IIRC they are only required for server-side bindings).

Out of curiosity, if you try re-creating your UniqueConstraint on 4.2, are they created with casts? By that I'm in no mean implying that a solution to this problem is to force the re-creation of constraints on each releases but I want to have an idea of the scope of the problem.

If that's the case it means that reverting to no cast in a minor release might break Postgres query planer the other way around as constraint created on 4.2 include ::text cast but queries on a future version removing them don't.

[Alex Vandiver]

Yes, the indexes as created on 4.2 do have the cast in them. The diff of our schema across the 4.2 upgrade is:

$ diff -U0 <(rg -U '(?s)CREATE[^;]*;' before.sql) <(rg -U '(?s)CREATE[^;]*;' after.sql)
--- /dev/fd/63  2023-09-13 12:55:09.595790167 -0700
+++ /dev/fd/62  2023-09-13 12:55:09.595790167 -0700
@@ -1095,4 +1095,4 @@
-CREATE UNIQUE INDEX unique_installation_count ON zulip.analytics_installationcount USING btree (property, subgroup, end_time) WHERE (subgroup IS NOT NULL);
-CREATE UNIQUE INDEX unique_installation_count_null_subgroup ON zulip.analytics_installationcount USING btree (property, end_time) WHERE (subgroup IS NULL);
-CREATE UNIQUE INDEX unique_realm_count ON zulip.analytics_realmcount USING btree (realm_id, property, subgroup, end_time) WHERE (subgroup IS NOT NULL);
-CREATE UNIQUE INDEX unique_realm_count_null_subgroup ON zulip.analytics_realmcount USING btree (realm_id, property, end_time) WHERE (subgroup IS NULL);
+CREATE UNIQUE INDEX unique_installation_count ON zulip.analytics_installationcount USING btree (property, subgroup, end_time) WHERE ((subgroup)::text IS NOT NULL);
+CREATE UNIQUE INDEX unique_installation_count_null_subgroup ON zulip.analytics_installationcount USING btree (property, end_time) WHERE ((subgroup)::text IS NULL);
+CREATE UNIQUE INDEX unique_realm_count ON zulip.analytics_realmcount USING btree (realm_id, property, subgroup, end_time) WHERE ((subgroup)::text IS NOT NULL);
+CREATE UNIQUE INDEX unique_realm_count_null_subgroup ON zulip.analytics_realmcount USING btree (realm_id, property, end_time) WHERE ((subgroup)::text IS NULL);
@@ -1100,4 +1100,4 @@
-CREATE UNIQUE INDEX unique_stream_count ON zulip.analytics_streamcount USING btree (stream_id, property, subgroup, end_time) WHERE (subgroup IS NOT NULL);
-CREATE UNIQUE INDEX unique_stream_count_null_subgroup ON zulip.analytics_streamcount USING btree (stream_id, property, end_time) WHERE (subgroup IS NULL);
-CREATE UNIQUE INDEX unique_user_count ON zulip.analytics_usercount USING btree (user_id, property, subgroup, end_time) WHERE (subgroup IS NOT NULL);
-CREATE UNIQUE INDEX unique_user_count_null_subgroup ON zulip.analytics_usercount USING btree (user_id, property, end_time) WHERE (subgroup IS NULL);
+CREATE UNIQUE INDEX unique_stream_count ON zulip.analytics_streamcount USING btree (stream_id, property, subgroup, end_time) WHERE ((subgroup)::text IS NOT NULL);
+CREATE UNIQUE INDEX unique_stream_count_null_subgroup ON zulip.analytics_streamcount USING btree (stream_id, property, end_time) WHERE ((subgroup)::text IS NULL);
+CREATE UNIQUE INDEX unique_user_count ON zulip.analytics_usercount USING btree (user_id, property, subgroup, end_time) WHERE ((subgroup)::text IS NOT NULL);
+CREATE UNIQUE INDEX unique_user_count_null_subgroup ON zulip.analytics_usercount USING btree (user_id, property, end_time) WHERE ((subgroup)::text IS NULL);

So if the application's schema was inserted with 4.2, then its partial indexes work with the current code. If it was inserted with Django < 4.2, then the indexes don't work with >= 4.2.

This also means that there's a bit of a pickle -- you can't just strip off the ::text in the query in some later 4.2 bugfix release, because that will also make any 4.2-inserted indexes with ::text not work with the newly ::text-less query.

Replying to Simon Charette:

We should try to avoid casts if possible (IIRC they are only required for server-side bindings).

Out of curiosity, if you try re-creating your UniqueConstraint on 4.2, are they created with casts? By that I'm in no mean implying that a solution to this problem is to force the re-creation of constraints on each releases but I want to have an idea of the scope of the problem.

If that's the case it means that reverting to no cast in a minor release might break Postgres query planer the other way around as constraint created on 4.2 include ::text cast but queries on a future version removing them don't.

[Simon Charette]

This also means that there's a bit of a pickle -- you can't just strip off the ::text in the query in some later 4.2 bugfix release, because that will also make any 4.2-inserted indexes with ::text not work with the newly ::text-less query.

Right, that's what I feared. I think we should favour reverting to the previous behaviour if possible though as there are likely to be more constraints created prior to 4.2 in the wild than after even at this point. This is definitely something that should be mentioned in the release notes though.

Any thoughts here Florian and Mariusz?

[Mariusz Felisiak]

Replying to Simon Charette:

Any thoughts here Florian and Mariusz?

I'd prefer to avoid casting, however, it is necessary when server-side binding is enabled.

[Simon Charette]

What do you think of making the casting conditional to the enablement of the server_side_binding option in this case?

That would contain the issue to the already existing buckets or problems that come with enabling this option that we're already aware of (e.g. grouping by parameterized annotation).

Happy to prepare a patch if you believe that it's worth pursuing.

[Mariusz Felisiak]

Replying to Simon Charette:

What do you think of making the casting conditional to the enablement of the server_side_binding option in this case?

Yeah, I was thinking the same thing. This is probably the only reasonable option we have.

[Florian Apolloner]

This doesn't just affect ::text but all casts?

[Mariusz Felisiak]

Replying to Florian Apolloner:

This doesn't just affect ::text but all casts?

As far as I'm aware, from the lookup perspective it ​affects mainly __isnull.

[murtatah]

Hey everyone,

I'm excited to tackle this one! The issue (#34840) looks like a great challenge, and I'd love to help resolve it. 😊

I'll start working on a solution and keep you posted on my progress. If you have any insights or suggestions along the way, please feel free to share them.while this is the first time I'm working on something like this so I hope I can do it right as much as possible!

Thanks for the opportunity!

[Florian Apolloner]

Replying to Mariusz Felisiak:

Replying to Florian Apolloner:

This doesn't just affect ::text but all casts?

As far as I'm aware, from the lookup perspective it ​affects mainly __isnull.

Oh right, and that is mostly a result of handling text columns as unknown oids instead of text which has/had it's own share of problems.

Replying to Simon Charette:

What do you think of making the casting conditional to the enablement of the server_side_binding option in this case?

Thinking about this more I am not sure this is a good idea; after all people should be free to change that setting and Django should behave the same. I feel like we are running into limitations of our ORM here, do we really need the ::text cast here for server side bindings? Isn't postgresql able to properly deduce the types? After all, in the query as well as in the index we reference the column subgroup which has a well defined type (or am I mistaken here). We should only need type casts if we involve bind parameters.

Though that begs the next question: Assume you generate an index for a text column named x. Would a query with .filter(x__startswith='test') use the index if the index is created on x but the query uses x::text like 'test%'?

EDIT:// The next question might be a non issue generate x like 'test%'::text -- Nevertheless it feels a bit unfortunate that postgresql cannot use the index with the cast.

[Florian Apolloner]

Running the tests with server side param binding enabled and having removed the cast for IS NULL results in three failures:

======================================================================
ERROR: test_model_validation_with_condition (constraints.tests.UniqueConstraintTests.test_model_validation_with_condition)
Partial unique constraints are not ignored by
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/usr/lib64/python3.11/unittest/case.py", line 57, in testPartExecutor
    yield
  File "/usr/lib64/python3.11/unittest/case.py", line 623, in run
    self._callTestMethod(testMethod)
    ^^^^^^^^^^^^^^^^^
  File "/usr/lib64/python3.11/unittest/case.py", line 579, in _callTestMethod
    if method() is not None:
    ^^^^^^^^^^^^^^^^^
  File "/home/florian/sources/django.git/django/test/testcases.py", line 1425, in skip_wrapper
    return test_func(*args, **kwargs)
    ^^^^^^^^^^^^^^^^^
  File "/home/florian/sources/django.git/tests/constraints/tests.py", line 804, in test_model_validation_with_condition
    ).validate_constraints()
    ^^^^^^^^^^^^^^^^^
  File "/home/florian/sources/django.git/django/db/models/base.py", line 1497, in validate_constraints
    raise ValidationError(errors)
    ^^^^^^^^^^^^^^^^^
django.core.exceptions.ValidationError: {'__all__': ['Constraint “name_without_color_uniq” is violated.']}

======================================================================
ERROR: test_validate_condition (constraints.tests.UniqueConstraintTests.test_validate_condition)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/usr/lib64/python3.11/unittest/case.py", line 57, in testPartExecutor
    yield
  File "/usr/lib64/python3.11/unittest/case.py", line 623, in run
    self._callTestMethod(testMethod)
    ^^^^^^^^^^^^^^^^^
  File "/usr/lib64/python3.11/unittest/case.py", line 579, in _callTestMethod
    if method() is not None:
    ^^^^^^^^^^^^^^^^^
  File "/home/florian/sources/django.git/django/test/testcases.py", line 1425, in skip_wrapper
    return test_func(*args, **kwargs)
    ^^^^^^^^^^^^^^^^^
  File "/home/florian/sources/django.git/tests/constraints/tests.py", line 879, in test_validate_condition
    constraint.validate(
    ^^^^^^^^^^^^^^^^^
  File "/home/florian/sources/django.git/django/db/models/constraints.py", line 461, in validate
    raise ValidationError(
    ^^^^^^^^^^^^^^^^^
django.core.exceptions.ValidationError: ['Constraint “name_without_color_uniq” is violated.']

======================================================================
ERROR: test_validate_expression_condition (constraints.tests.UniqueConstraintTests.test_validate_expression_condition)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/usr/lib64/python3.11/unittest/case.py", line 57, in testPartExecutor
    yield
  File "/usr/lib64/python3.11/unittest/case.py", line 623, in run
    self._callTestMethod(testMethod)
    ^^^^^^^^^^^^^^^^^
  File "/usr/lib64/python3.11/unittest/case.py", line 579, in _callTestMethod
    if method() is not None:
    ^^^^^^^^^^^^^^^^^
  File "/home/florian/sources/django.git/tests/constraints/tests.py", line 965, in test_validate_expression_condition
    constraint.validate(
    ^^^^^^^^^^^^^^^^^
  File "/home/florian/sources/django.git/django/db/models/constraints.py", line 461, in validate
    raise ValidationError(
    ^^^^^^^^^^^^^^^^^
django.core.exceptions.ValidationError: ['Constraint “name_lower_without_color_uniq” is violated.']

----------------------------------------------------------------------
Ran 16695 tests in 201.401s

[murtatah]

fix: Correct value format for CharArrayModel in test_isnull_lookup_cast

In the test_isnull_lookup_cast method, the CharArrayModel instance was created with an incorrect value format for the 'field' attribute. This caused a PostgreSQL error when trying to create the instance.

This commit fixes the issue by using a list to represent an array and encloses the value properly in curly braces, ensuring the correct format for the 'field' attribute.i made some test check the status of it.

[Florian Apolloner]

Please don't see tickets with your own merge requests to "ready for checkin". Additionally, simply removing the cast is not the correct fix (at the very least it is most likely possible to just drop the if etc…)

[Florian Apolloner]

Regarding my comment https://code.djangoproject.com/ticket/34840#comment:11 -- The validation error is wrong there, it is actually a database area that is masked (​https://github.com/django/django/pull/15659#pullrequestreview-1629960210). I wonder if we could teach the ORM to only add casts when using bind params, there is no reason to add those for column refs. That is unless our testsuite doesn't miss some further important edge cases :)

At the very least we will need instructions (release notes, maybe a blog post) for people on how to check whether they are affected and how to fix it.

[Mariusz Felisiak]

I wonder if we could extend Simon's solution for #34754 (3434dbd39d373df7193ad006b970c09c1a909ea3). The following patch passes all tests (with and without server side binding parameters):

django/db/backends/postgresql/operations.py

@@ -155 +155 @@
     def lookup_cast(self, lookup_type, internal_type=None):
         lookup = "%s"
 
-        if lookup_type == "isnull" and internal_type in (
-            "CharField",
-            "EmailField",
-            "TextField",
-        ):
-            return "%s::text"
-
         # Cast text lookups to text to allow things like filter(x__contains=4)
         if lookup_type in (
             "iexact",

django/db/models/lookups.py

@@ -624 +624 @@
             raise ValueError(
                 "The QuerySet value for an isnull lookup must be True or False."
             )
-        if isinstance(self.lhs, Value) and self.lhs.value is None:
+        if isinstance(self.lhs, Value):
             if self.rhs:
-                raise FullResultSet
+                if self.lhs.value is None:
+                    raise FullResultSet
+                else:
+                    raise EmptyResultSet
             else:
-                raise EmptyResultSet
+                if self.lhs.value is None:
+                    raise EmptyResultSet
+                else:
+                    raise FullResultSet
         sql, params = self.process_lhs(compiler, connection)
         if self.rhs:
             return "%s IS NULL" % sql, params

django/db/models/query_utils.py

@@ -10 +10 @@
 import logging
 from collections import namedtuple
 
-from django.core.exceptions import FieldError
+from django.core.exceptions import EmptyResultSet, FieldError
 from django.db import DEFAULT_DB_ALIAS, DatabaseError, connections
 from django.db.models.constants import LOOKUP_SEP
 from django.utils import tree
@@ -140 +140 @@
         except DatabaseError as e:
             logger.warning("Got a database error calling check() on %r: %s", self, e)
             return True
+        except EmptyResultSet:
+            return False
 
     def deconstruct(self):
         path = "%s.%s" % (self.__class__.__module__, self.__class__.__name__)

[Mariusz Felisiak]

Alex, does it work for you?

[Florian Apolloner]

That is some creative thinking Mariusz. So just so I get this right: Django "optimizes" away part of the asked query if it can determine the outcome of the clause/lookup/etc? If that is the case: wow, never realized that :)

[Simon Charette]

Florian, yeah that's the gist of it.

This approach has existed for a while in the ORM to avoid performing a query for filter(somefield__in=[]) as SQL doesn't allow "column" IN () anyway and we've leaned on it recently to prevent some full table scan on Postgres (#27397) and issues with NULL handling when dealing with JSONField (#34754).

I think it makes sense to use this approach in cases where Python input poses type ambiguity at the database level like it did with the IN operator, after all we were forced to add ::text cast because Postgres doesn't know what the type of %s IS NULL should be.

In all cases the adjustments to Q.check to handle EmptyResultSet should be added as other lookups could generate it, the same can be said with FullResultSet.

I think we should not let the discussion about the silencing of DatabaseError die off though and consider a deprecation path to raise the awareness of the issue. Maybe a possible alternative solution that doesn't warrant expression introspection (e.g. special casing RawSQL) could be a Constraint kwarg that allows disabling model level validation entirely? That could be a way to silence the warning during the deprecation period.

[Florian Apolloner]

Replying to Simon Charette:

This approach has existed for a while in the ORM to avoid performing a query for filter(somefield__in=[]) as SQL doesn't allow "column" IN () anyway and we've leaned on it recently to prevent some full table scan on Postgres (#27397) and issues with NULL handling when dealing with JSONField (#34754).

I think it makes sense to use this approach in cases where Python input poses type ambiguity at the database level like it did with the IN operator, after all we were forced to add ::text cast because Postgres doesn't know what the type of %s IS NULL should be.

Makes sense. And as we see we only need the ::text cast in very special situations, a .filter(field__isnull=True) won't require it…

In all cases the adjustments to Q.check to handle EmptyResultSet should be added as other lookups could generate it, the same can be said about FullResultSet.

ACK, and we need tests for both cases. I see that the WhereNode properly propagates EmptyResultSet/FullResultSet -- I wonder if there are other cases where we don't do that yet (but I guess we will need to find them on a case by case basis).

I think we should not let the discussion about the silencing of DatabaseError die off though and consider a deprecation path to raise the awareness of the issue. Maybe a possible alternative solution that doesn't warrant expression introspection (e.g. special casing RawSQL) could be a Constraint kwarg that allows disabling model level validation entirely? That could be a way to silence the warning during the deprecation period.

+1. Though that should probably be a separate PR/ticket?

As for the indexing issues for people which already created indexes on 4.2 -- we should probably (aside from the release notes) also mention it explicitly in a blog post? This is something that can be very hard to find.

[Alex Vandiver]

Replying to Mariusz Felisiak:

Alex, does it work for you?

My parse of it is that it'll drop the ::text cast in the query, which means that the partial index, as created pre-4.2, work as expected -- so yup!

(For the meantime, we've manually recreated the indexes with the cast in them in production, so we get their utility -- but we'll happily swap those back to the cast-less variety when we upgrade to a Django with this fix)

[Mariusz Felisiak]

​PR

[GitHub <noreply@…>]

In 779cd28:

Fixed #34840 -- Avoided casting string base fields on PostgreSQL.

Thanks Alex Vandiver for the report.

Regression in 09ffc5c1212d4ced58b708cbbf3dfbfb77b782ca.

[Mariusz Felisiak <felisiak.mariusz@…>]

In 2566f92:

[5.0.x] Fixed #34840 -- Avoided casting string base fields on PostgreSQL.

Thanks Alex Vandiver for the report.

Regression in 09ffc5c1212d4ced58b708cbbf3dfbfb77b782ca.

Backport of 779cd28acb1f7eb06f629c0ea4ded99b5ebb670a from main.

[Mariusz Felisiak <felisiak.mariusz@…>]

In a148461f:

[4.2.x] Fixed #34840 -- Avoided casting string base fields on PostgreSQL.

Thanks Alex Vandiver for the report.

Regression in 09ffc5c1212d4ced58b708cbbf3dfbfb77b782ca.

Backport of 779cd28acb1f7eb06f629c0ea4ded99b5ebb670a from main.

[GitHub <noreply@…>]

In 9fd3a0f:

Refs #34840 -- Fixed test_validate_nullable_textfield_with_isnull_true() on databases that don's support table check constraints.

Thanks Tim Graham for the report.

[Mariusz Felisiak <felisiak.mariusz@…>]

In c3cf518:

[5.0.x] Refs #34840 -- Fixed test_validate_nullable_textfield_with_isnull_true() on databases that don's support table check constraints.

Thanks Tim Graham for the report.
Backport of 9fd3a0ffc8fb8e532b0d5717f99fb15aff175a6c from main

[Mariusz Felisiak <felisiak.mariusz@…>]

In b6bb2f80:

[4.2.x] Refs #34840 -- Fixed test_validate_nullable_textfield_with_isnull_true() on databases that don's support table check constraints.

Thanks Tim Graham for the report.
Backport of 9fd3a0ffc8fb8e532b0d5717f99fb15aff175a6c from main

[GitHub <noreply@…>]

In 73b2c63:

Refs #34840 -- Improved release note describing index regression.

[Mariusz Felisiak <felisiak.mariusz@…>]

In 12a1932:

[5.0.x] Refs #34840 -- Improved release note describing index regression.

Backport of 73b2c63127297e214d3bfdc8d9a96837691fc2a0 from main

[Mariusz Felisiak <felisiak.mariusz@…>]

In caec4f4a:

[4.2.x] Refs #34840 -- Improved release note describing index regression.

Backport of 73b2c63127297e214d3bfdc8d9a96837691fc2a0 from main