PYTHON-3689 Fix flaky test_load_balancing without @flaky decorator

test/asynchronous/test_server_selection_in_window.py

@@ -21,7 +21,6 @@
 from pathlib import Path
 from test.asynchronous import AsyncIntegrationTest, async_client_context, unittest
 from test.asynchronous.helpers import ConcurrentRunner
-from test.asynchronous.utils import flaky
 from test.asynchronous.utils_selection_tests import create_topology
 from test.asynchronous.utils_spec_runner import AsyncSpecTestCreator
 from test.utils_shared import (
@@ -138,7 +137,6 @@ async def frequencies(self, client, listener, n_finds=10):
 
     @async_client_context.require_failCommand_appName
     @async_client_context.require_multiple_mongoses
-    @flaky(reason="PYTHON-3689")
     async def test_load_balancing(self):
         listener = OvertCommandListener()
         cmap_listener = CMAPListener()
@@ -165,12 +163,32 @@ async def test_load_balancing(self):
                 "appName": "loadBalancingTest",
             },
         }
+        coll = client.test.test
+        N_TASKS = 10
         async with self.fail_point(delay_finds):
             nodes = async_client_context.client.nodes
             self.assertEqual(len(nodes), 1)
             delayed_server = next(iter(nodes))
+            # Start background tasks to build up op_count on the delayed server.
+            # This ensures the measurement phase sees a stable op_count imbalance
+            # rather than a 50/50 random distribution from equal initial counts.
+            background_tasks = [FinderTask(coll, 1) for _ in range(N_TASKS)]
+            for task in background_tasks:
+                await task.start()
+            # Wait until all background finds are dispatched so the delayed
+            # server's finds are in-flight (each blocked for 500ms).
+            await async_wait_until(
+                lambda: len(listener.started_events) >= N_TASKS,
+                "background tasks to dispatch finds",
+            )
+            # Measure distribution while the delayed server is busy.
+            listener.reset()
             freqs = await self.frequencies(client, listener)
             self.assertLessEqual(freqs[delayed_server], 0.25)
+            for task in background_tasks:
+                await task.join()
+            for task in background_tasks:
+                self.assertTrue(task.passed)
         listener.reset()
         freqs = await self.frequencies(client, listener, n_finds=150)
         self.assertAlmostEqual(freqs[delayed_server], 0.50, delta=0.15)

test/asynchronous/unified_format.py

@@ -1470,7 +1470,6 @@ async def run_scenario(self, spec, uri=None):
             ("PYTHON-5174", ".*Driver_extends_timeout_while_streaming"),
             ("PYTHON-5315", ".*TestSrvPolling.test_recover_from_initially_.*"),
             ("PYTHON-4987", ".*UnknownTransactionCommitResult_labels_to_connection_errors"),
-            ("PYTHON-3689", ".*TestProse.test_load_balancing"),
             ("PYTHON-3522", ".*csot.*"),
         ]
         for reason, flaky_test in flaky_tests:

test/test_server_selection_in_window.py

@@ -21,7 +21,6 @@
 from pathlib import Path
 from test import IntegrationTest, client_context, unittest
 from test.helpers import ConcurrentRunner
-from test.utils import flaky
 from test.utils_selection_tests import create_topology
 from test.utils_shared import (
     CMAPListener,
@@ -138,7 +137,6 @@ def frequencies(self, client, listener, n_finds=10):
 
     @client_context.require_failCommand_appName
     @client_context.require_multiple_mongoses
-    @flaky(reason="PYTHON-3689")
     def test_load_balancing(self):
         listener = OvertCommandListener()
         cmap_listener = CMAPListener()
@@ -165,12 +163,32 @@ def test_load_balancing(self):
                 "appName": "loadBalancingTest",
             },
         }
+        coll = client.test.test
+        N_TASKS = 10
         with self.fail_point(delay_finds):
             nodes = client_context.client.nodes
             self.assertEqual(len(nodes), 1)
             delayed_server = next(iter(nodes))
+            # Start background tasks to build up op_count on the delayed server.
+            # This ensures the measurement phase sees a stable op_count imbalance
+            # rather than a 50/50 random distribution from equal initial counts.
+            background_tasks = [FinderTask(coll, 1) for _ in range(N_TASKS)]
+            for task in background_tasks:
+                task.start()
+            # Wait until all background finds are dispatched so the delayed
+            # server's finds are in-flight (each blocked for 500ms).
+            wait_until(
+                lambda: len(listener.started_events) >= N_TASKS,
+                "background tasks to dispatch finds",
+            )
+            # Measure distribution while the delayed server is busy.
+            listener.reset()
             freqs = self.frequencies(client, listener)
             self.assertLessEqual(freqs[delayed_server], 0.25)
+            for task in background_tasks:
+                task.join()
+            for task in background_tasks:
+                self.assertTrue(task.passed)
         listener.reset()
         freqs = self.frequencies(client, listener, n_finds=150)
         self.assertAlmostEqual(freqs[delayed_server], 0.50, delta=0.15)

test/unified_format.py

@@ -1457,7 +1457,6 @@ def run_scenario(self, spec, uri=None):
             ("PYTHON-5174", ".*Driver_extends_timeout_while_streaming"),
             ("PYTHON-5315", ".*TestSrvPolling.test_recover_from_initially_.*"),
             ("PYTHON-4987", ".*UnknownTransactionCommitResult_labels_to_connection_errors"),
-            ("PYTHON-3689", ".*TestProse.test_load_balancing"),
             ("PYTHON-3522", ".*csot.*"),
         ]
         for reason, flaky_test in flaky_tests: