From dcd20089ba9f5fa24acf04f2eac2f70e90c3e45c Mon Sep 17 00:00:00 2001
From: Jeremy Stretch <jstretch@netboxlabs.com>
Date: Wed, 2 Sep 2026 09:47:34 -0400
Subject: [PATCH] Fix cross-worker cache contamination in parallel test runs
 (#23106)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

RQQueueTestMixin cleared RQ queues with FLUSHALL, which empties every
database on the Redis server — including the caching database, whose
'config'/'config_version' keys are shared by all parallel test workers.
A flush landing mid-test forces an unrelated worker to re-read
core_configrevision, adding two queries to the affected request. Flush
only the queue's own database instead.

Also stop GraphQLDeferredColumnTestCase from comparing total query counts
between two requests, which is what surfaced the race as intermittent
"Query count grew from 7 to 9" failures in CI. Assert that the target
table is read exactly once per request instead, as #23034 did for the
equivalent custom fields test.
---
 netbox/netbox/tests/test_graphql.py | 32 ++++++++++++-----------------
 netbox/utilities/testing/mixins.py  |  5 ++++-
 2 files changed, 17 insertions(+), 20 deletions(-)

diff --git a/netbox/netbox/tests/test_graphql.py b/netbox/netbox/tests/test_graphql.py
index e4ebf6e2a..7e41a5411 100644
--- a/netbox/netbox/tests/test_graphql.py
+++ b/netbox/netbox/tests/test_graphql.py
@@ -11,7 +11,6 @@ from django.db import connection
 from django.test import override_settings
 from django.test.utils import CaptureQueriesContext
 from django.urls import reverse
-from django.utils import timezone
 from rest_framework import status
 from strawberry.extensions import QueryDepthLimiter
 from strawberry.schema.config import StrawberryConfig
@@ -39,7 +38,7 @@ from netbox.graphql.schema import Query, get_schema_extensions, schema
 from netbox.graphql.utils import register_model_graphql_type, splice_extension_bases, validate_extension_final_names
 from netbox.registry import registry
 from netbox.tests.dummy_plugin.models import DummySiteAttachment
-from users.models import ObjectPermission, Token, User
+from users.models import ObjectPermission, User
 from utilities.tables import get_table_for_model
 from utilities.testing import APITestCase, APIViewTestCases, TestCase, disable_warnings
 
@@ -1171,8 +1170,8 @@ class GraphQLDeferredColumnTestCase(APITestCase):
     reads must therefore be declared via an `only` hint; otherwise the column is deferred and reading it
     reloads the row from the database once per object returned (see #22813).
 
-    Each test below asserts both that no single-row reload occurs and that the total query count does not
-    grow with the number of objects returned.
+    Each test below asserts both that no single-row reload occurs and that the model's table is read
+    exactly once per request, regardless of the number of objects returned.
     """
     OBJECT_COUNT = 10
 
@@ -1241,14 +1240,13 @@ class GraphQLDeferredColumnTestCase(APITestCase):
         """
         Execute `query_template` (which must accept a `limit` interpolation) for a single object and for
         OBJECT_COUNT objects, asserting that no row of `table` is re-fetched by primary key and that the
-        total query count is identical for both. `validate` is called with the returned objects.
-        """
-        # Token authentication updates Token.last_used at most once per minute, so the first API request
-        # made by a test issues an additional UPDATE (see netbox/api/authentication.py). Stamp the token
-        # up front to keep that one-off write out of the counts compared below.
-        Token.objects.filter(pk=self.token.pk).update(last_used=timezone.now())
+        table is read exactly once per request regardless of the number of objects returned. `validate`
+        is called with the returned objects.
 
-        query_counts = []
+        Assert on the number of reads of `table` rather than on the total query count: the capture window
+        also holds request processing queries which may be repeated when a cache key shared by the other
+        parallel test workers is evicted mid-test, which would otherwise fail the test spuriously.
+        """
         for limit in (1, self.OBJECT_COUNT):
             data, queries = self._execute(query_template % {'limit': limit})
             objects = data[list_field]
@@ -1259,16 +1257,12 @@ class GraphQLDeferredColumnTestCase(APITestCase):
             self.assertEqual(
                 reloads, [], msg=f'{len(reloads)} deferred-column reload(s) for {limit} object(s): {reloads[:1]}'
             )
-            query_counts.append(len(queries))
 
-        self.assertEqual(
-            query_counts[0],
-            query_counts[1],
-            msg=(
-                f'Query count grew from {query_counts[0]} to {query_counts[1]} when the number of objects '
-                f'returned grew from 1 to {self.OBJECT_COUNT}'
+            reads = [q['sql'] for q in queries if f'FROM "{table}"' in q['sql']]
+            self.assertEqual(
+                len(reads), 1,
+                msg=f'{table} must be read once for {limit} object(s), got {len(reads)} read(s): {reads}'
             )
-        )
 
     def test_custom_fields(self):
         """
diff --git a/netbox/utilities/testing/mixins.py b/netbox/utilities/testing/mixins.py
index 55088a664..c515fe84b 100644
--- a/netbox/utilities/testing/mixins.py
+++ b/netbox/utilities/testing/mixins.py
@@ -24,7 +24,10 @@ class RQQueueTestMixin(SerializeMixin):
     @classmethod
     def clear_rq_queues(cls):
         for queue_name in cls.rq_queue_names:
-            get_queue(queue_name).connection.flushall()
+            # Flush only the queue's own database. FLUSHALL would empty every database on the Redis
+            # server, including the caching database, whose keys (e.g. the cached config revision)
+            # are shared by the other parallel test workers.
+            get_queue(queue_name).connection.flushdb()
 
     def run_rq_jobs(self, *queue_names, burst=True):
         """