Fix new scheduled tasks jumping the queue (#17962)

Currently, when a new scheduled task is added and its scheduled time has
already passed, we set it to ACTIVE. This is problematic, because it
means it will jump the queue ahead of all other SCHEDULED tasks;
furthermore, if the Synapse process gets restarted, it will jump ahead
of any ACTIVE tasks which have been started but are taking a while to
run.

Instead, we leave it set to SCHEDULED, but kick off a call to
`_launch_scheduled_tasks`, which will decide if we actually have
capacity to start a new task, and start the newly-added task if so.

changelog.d/17962.misc

@@ -0,0 +1 @@
+Fix new scheduled tasks jumping the queue.

synapse/replication/tcp/commands.py

@@ -495,7 +495,7 @@ class LockReleasedCommand(Command):
 
 
 class NewActiveTaskCommand(_SimpleCommand):
-    """Sent to inform instance handling background tasks that a new active task is available to run.
+    """Sent to inform instance handling background tasks that a new task is ready to run.
 
     Format::
 

synapse/replication/tcp/handler.py

@@ -727,7 +727,7 @@ class ReplicationCommandHandler:
     ) -> None:
         """Called when get a new NEW_ACTIVE_TASK command."""
         if self._task_scheduler:
-            self._task_scheduler.launch_task_by_id(cmd.data)
+            self._task_scheduler.on_new_task(cmd.data)
 
     def new_connection(self, connection: IReplicationConnection) -> None:
         """Called when we have a new connection."""

synapse/util/task_scheduler.py

@@ -174,9 +174,10 @@ class TaskScheduler:
             The id of the scheduled task
         """
         status = TaskStatus.SCHEDULED
+        start_now = False
         if timestamp is None or timestamp < self._clock.time_msec():
             timestamp = self._clock.time_msec()
-            status = TaskStatus.ACTIVE
+            start_now = True
 
         task = ScheduledTask(
             random_string(16),
@@ -190,9 +191,11 @@ class TaskScheduler:
         )
         await self._store.insert_scheduled_task(task)
 
-        if status == TaskStatus.ACTIVE:
+        # If the task is ready to run immediately, run the scheduling algorithm now
+        # rather than waiting
+        if start_now:
             if self._run_background_tasks:
-                await self._launch_task(task)
+                self._launch_scheduled_tasks()
             else:
                 self._hs.get_replication_command_handler().send_new_active_task(task.id)
 
@@ -300,23 +303,13 @@ class TaskScheduler:
             raise Exception(f"Task {id} is currently ACTIVE and can't be deleted")
         await self._store.delete_scheduled_task(id)
 
-    def launch_task_by_id(self, id: str) -> None:
-        """Try launching the task with the given ID."""
-        # Don't bother trying to launch new tasks if we're already at capacity.
-        if len(self._running_tasks) >= TaskScheduler.MAX_CONCURRENT_RUNNING_TASKS:
-            return
+    def on_new_task(self, task_id: str) -> None:
+        """Handle a notification that a new ready-to-run task has been added to the queue"""
+        # Just run the scheduler
+        self._launch_scheduled_tasks()
 
-        run_as_background_process("launch_task_by_id", self._launch_task_by_id, id)
-
-    async def _launch_task_by_id(self, id: str) -> None:
-        """Helper async function for `launch_task_by_id`."""
-        task = await self.get_task(id)
-        if task:
-            await self._launch_task(task)
-
-    @wrap_as_background_process("launch_scheduled_tasks")
-    async def _launch_scheduled_tasks(self) -> None:
-        """Retrieve and launch scheduled tasks that should be running at that time."""
+    def _launch_scheduled_tasks(self) -> None:
+        """Retrieve and launch scheduled tasks that should be running at this time."""
         # Don't bother trying to launch new tasks if we're already at capacity.
         if len(self._running_tasks) >= TaskScheduler.MAX_CONCURRENT_RUNNING_TASKS:
             return
@@ -326,20 +319,26 @@ class TaskScheduler:
 
         self._launching_new_tasks = True
 
-        try:
-            for task in await self.get_tasks(
-                statuses=[TaskStatus.ACTIVE], limit=self.MAX_CONCURRENT_RUNNING_TASKS
-            ):
-                await self._launch_task(task)
-            for task in await self.get_tasks(
-                statuses=[TaskStatus.SCHEDULED],
-                max_timestamp=self._clock.time_msec(),
-                limit=self.MAX_CONCURRENT_RUNNING_TASKS,
-            ):
-                await self._launch_task(task)
+        async def inner() -> None:
+            try:
+                for task in await self.get_tasks(
+                    statuses=[TaskStatus.ACTIVE],
+                    limit=self.MAX_CONCURRENT_RUNNING_TASKS,
+                ):
+                    # _launch_task will ignore tasks that we're already running, and
+                    # will also do nothing if we're already at the maximum capacity.
+                    await self._launch_task(task)
+                for task in await self.get_tasks(
+                    statuses=[TaskStatus.SCHEDULED],
+                    max_timestamp=self._clock.time_msec(),
+                    limit=self.MAX_CONCURRENT_RUNNING_TASKS,
+                ):
+                    await self._launch_task(task)
 
-        finally:
-            self._launching_new_tasks = False
+            finally:
+                self._launching_new_tasks = False
+
+        run_as_background_process("launch_scheduled_tasks", inner)
 
     @wrap_as_background_process("clean_scheduled_tasks")
     async def _clean_scheduled_tasks(self) -> None:

tests/util/test_task_scheduler.py

@@ -18,8 +18,7 @@
 # [This file includes modifications made by New Vector Limited]
 #
 #
-
-from typing import Optional, Tuple
+from typing import List, Optional, Tuple
 
 from twisted.internet.task import deferLater
 from twisted.test.proto_helpers import MemoryReactor
@@ -104,33 +103,43 @@ class TestTaskScheduler(HomeserverTestCase):
                 )
             )
 
-        # This is to give the time to the active tasks to finish
-        self.reactor.advance(1)
-
-        # Check that only MAX_CONCURRENT_RUNNING_TASKS tasks has run and that one
-        # is still scheduled.
-        tasks = [
-            self.get_success(self.task_scheduler.get_task(task_id))
-            for task_id in task_ids
-        ]
+        def get_tasks_of_status(status: TaskStatus) -> List[ScheduledTask]:
+            tasks = (
+                self.get_success(self.task_scheduler.get_task(task_id))
+                for task_id in task_ids
+            )
+            return [t for t in tasks if t is not None and t.status == status]
 
+        # At this point, there should be MAX_CONCURRENT_RUNNING_TASKS active tasks and
+        # one scheduled task.
         self.assertEquals(
-            len(
-                [t for t in tasks if t is not None and t.status == TaskStatus.COMPLETE]
-            ),
+            len(get_tasks_of_status(TaskStatus.ACTIVE)),
             TaskScheduler.MAX_CONCURRENT_RUNNING_TASKS,
         )
+        self.assertEquals(
+            len(get_tasks_of_status(TaskStatus.SCHEDULED)),
+            1,
+        )
 
-        scheduled_tasks = [
-            t for t in tasks if t is not None and t.status == TaskStatus.ACTIVE
-        ]
-        self.assertEquals(len(scheduled_tasks), 1)
-
-        # We need to wait for the next run of the scheduler loop
-        self.reactor.advance((TaskScheduler.SCHEDULE_INTERVAL_MS / 1000))
+        # Give the time to the active tasks to finish
         self.reactor.advance(1)
 
-        # Check that the last task has been properly executed after the next scheduler loop run
+        # Check that MAX_CONCURRENT_RUNNING_TASKS tasks have run and that one
+        # is still scheduled.
+        self.assertEquals(
+            len(get_tasks_of_status(TaskStatus.COMPLETE)),
+            TaskScheduler.MAX_CONCURRENT_RUNNING_TASKS,
+        )
+        scheduled_tasks = get_tasks_of_status(TaskStatus.SCHEDULED)
+        self.assertEquals(len(scheduled_tasks), 1)
+
+        # The scheduled task should start 0.1s after the first of the active tasks
+        # finishes
+        self.reactor.advance(0.1)
+        self.assertEquals(len(get_tasks_of_status(TaskStatus.ACTIVE)), 1)
+
+        # ... and should finally complete after another second
+        self.reactor.advance(1)
         prev_scheduled_task = self.get_success(
             self.task_scheduler.get_task(scheduled_tasks[0].id)
         )