N+1 queries from RelatedManager + only("pk")

[REGNIER Guillaume]

When iterating over a queryset constructed from a RelatedManager and a .only(...) call that does not include the related field, a query occurs when instances are produced from the queryset.

Steps to Reproduce:

class Company(models.Model):
    pass

class Employee(models.Model):
    company = models.ForeignKey(Company, on_delete=models.CASCADE, related_name="employees")

company = Company.objects.create()
Employee.objects.bulk_create(Employee(company=company) for _ in range(10))

for employee in company.employees.only("pk"):
    # Some code that only access pk
    _ = employee.pk

Expected Behavior:

One query like

SELECT "employee"."id" FROM "employee" WHERE "employee"."company_id" = {COMPANY_ID}

Actual Behavior:

10 additional queries like:

SELECT "employee"."id", "employee"."company_id" FROM "employee" WHERE "employee"."id" = {EMPLOYEE_ID}

Analysis:

My understanding is that there is an optimization that fills the parent model on related instances without needing additional SQL join/query. However, when only a subset of fields is selected (in this case, only the primary key), the parent ID might not be loaded from the database, resulting in additional queries to perform said optimization.

Workaround:

company = Company.objects.create()
Employee.objects.bulk_create(Employee(company=company) for _ in range(10))

for employee in Employee.objects.filter(company=company).only("pk"):
    # Some code that only access pk
    _ = employee.pk

[Simon Charette]

This relates to #20927, #18177, and at least another ticket that asked for "company_id" to be implicitly included in the only call that I just can't find anymore.

I think we should try to resolve that in a different way and a potential solution could be something like

django/db/models/query.py

@@ -119 +119 @@
                 ),
             )
             for field, related_objs in queryset._known_related_objects.items()
+            # if not is_deferred(field)
         ]
         for row in compiler.results_iter(results):
             obj = model_cls.from_db(

But in order to work properly it would require a refactor of known_related_objects as the current approach is naive wrt/ to how it namespace related objects by field as the same model field can be included multiple times in the the same queryset (e.g. see #35356)

[Natalia Bidart]

Hello Guillaume!

I'm having a hard time understanding the goal of your code. I assume you have provided a simplified version to be used as a small example, which I appreciate, but with the information given, the proper way to fetch records from the database without generating N+1 queries would be:

>>> from django.db import connection, reset_queries
>>> reset_queries(); connection.queries
[]
>>> company = Company.objects.prefetch_related("employees").last()
>>> for e in company.employees.all():
>>>     print(e.pk)
7
8
9
10
>>> connection.queries
[{'sql': 'SELECT "ticket_35442_company"."id", "ticket_35442_company"."name" FROM "ticket_35442_company" ORDER BY "ticket_35442_company"."id" DESC LIMIT 1',
  'time': '0.002'},
 {'sql': 'SELECT "ticket_35442_employee"."id", "ticket_35442_employee"."company_id" FROM "ticket_35442_employee" WHERE "ticket_35442_employee"."company_id" IN (5)',
  'time': '0.001'}]

Or, my preferred (a single query and the same workaround as was posted originally):

>>> for e in Employee.objects.filter(company=company).only("pk"): print(e.pk)
7
8
9
10
>>> connection.queries
[{'sql': 'SELECT "ticket_35442_employee"."id" FROM "ticket_35442_employee" WHERE "ticket_35442_employee"."company_id" = 5',
  'time': '0.000'}]

[Natalia Bidart]

Hey Simon, thank you for your message. I fail to see how Django is at fault here, considering that there are alternative usages of the ORM and its queries to accomplish the same result (as far as we understand the use case) that would not execute N+1 queries. This seems like a (likely?) dupe of #20923 which was closed as wontfix.

[Natalia Bidart]

I'll close as duplicate of #20923 for now following the ​ticket triaging process, but any further commenting is welcomed and I'm happy to reconsider the ticket status with more information.

[REGNIER Guillaume]

Thank you both for answering!

Yes, i did provide a minimal exemple. Here i just wanted to use company.employees.all() instead of Employee.objects.filter(company=company).
I often use RelatedManagers this way to avoid importing the related model and i do find the final code easier to read.
I thought the two pieces of code above were meant to be equivalent.

I might have understood the RelatedManager wrong but when i'm passing a queryset along, i have no way to know if it originated from a RelatedManager or a regular Manager, meaning i have no way to know if .only() is safe to use or if it'll cause N+1 queries.

If this behavior is intended/not an issue, it means that i either need to ban .only() or RelatedManagers from my coding habbits since i don't have any way to ensure this won't happen again.

[Simon Charette]

Accessing objects though a related manager should still allow for only to work as expected. The fact that the ORM doesn't even warn you when it silently issues queries on field deferral leaks (#22492) makes this behavior really insidious and prevents safe usage of related managers as Guillaume brought up.

I think the the ORM should either have RelatedManager.only include the reverse field implicitly (which we've kind of rules out against in #33835) or we should find a way to more safely assign known related objects in the face of deferred fields (e.g. make sure required attrs are in obj.__dict__).

Potentially naive patch

django/db/models/query.py

@@ -107 +107 @@
             (
                 field,
                 related_objs,
-                operator.attrgetter(
-                    *[
-                        (
-                            field.attname
-                            if from_field == "self"
-                            else queryset.model._meta.get_field(from_field).attname
-                        )
-                        for from_field in field.from_fields
-                    ]
-                ),
+                attnames := [
+                    (
+                        field.attname
+                        if from_field == "self"
+                        else queryset.model._meta.get_field(from_field).attname
+                    )
+                    for from_field in field.from_fields
+                ],
+                operator.attrgetter(*attnames),
             )
             for field, related_objs in queryset._known_related_objects.items()
         ]
         for row in compiler.results_iter(results):
             obj = model_cls.from_db(
@@ -131 +131 @@
                     setattr(obj, attr_name, row[col_pos])
 
             # Add the known related objects to the model.
-            for field, rel_objs, rel_getter in known_related_objects:
+            for field, rel_objs, attnames, rel_getter in known_related_objects:
                 # Avoid overwriting objects loaded by, e.g., select_related().
                 if field.is_cached(obj):
                     continue
+                # Avoid fetching potentially deferred attributes that would
+                # result in unexpected queries.
+                if any(attname not in obj.__dict__ for attname in attnames):
+                    continue
                 rel_obj_id = rel_getter(obj)
                 try:
                     rel_obj = rel_objs[rel_obj_id]

[Simon Charette]

So I tried running the suite against the above patch and it delivered!

I was pointed at a single failure being defer.tests.DeferTests.test_only

======================================================================
FAIL: test_only (defer.tests.DeferTests)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/Users/charettes/Workspace/django/tests/defer/tests.py", line 52, in test_only
    self.assert_delayed(self.s1.primary_set.only("pk")[0], 2)
  File "/Users/charettes/Workspace/django/tests/defer/tests.py", line 24, in assert_delayed
    self.assertEqual(count, num)
AssertionError: 3 != 2

What's interesting though is that ​it might have a been a sneaky regression caused by the field deferral refactor in Django 1.10 that flew under the radar.

It was not discussed at all in the two PRs where #26207 was resolved

• ​https://github.com/django/django/pull/6118
• ​https://github.com/django/django/pull/6491

[Simon Charette]

For the record, I checked out I checked out Django 1.9 and it seems the issue is present there so while the commit mentioned above that landed in 1.10 did alter a related test it didn't actually break anything in this regard.

[Sarah Boyce]

Thank you for the investigation Simon, reopening 👍