Add no worker timeout for scheduler

distributed/distributed-schema.yaml

@@ -76,7 +76,20 @@ properties:
             description: |
               Shut down the scheduler after this duration if no activity has occurred
 
-              This can be helpful to reduce costs and stop zombie processes from roaming the earth.
+          no-workers-timeout:
+            type:
+            - string
+            - "null"
+            description: |
+              Shut down the scheduler after this duration if there are pending tasks,
+              but no workers that can process them. This can either mean that there are
+              no workers running at all, or that there are idle workers but they've been
+              excluded through worker or resource restrictions.
+
+              In adaptive clusters, this timeout must be set to be safely higher than
+              the time it takes for workers to spin up.
+
+              Works in conjunction with idle-timeout.
 
           work-stealing:
             type: boolean

distributed/distributed.yaml

@@ -18,6 +18,7 @@ distributed:
     # after they have been removed from the scheduler
     events-cleanup-delay: 1h
     idle-timeout: null      # Shut down after this duration, like "1h" or "30 minutes"
+    no-workers-timeout: 20m # Shut down if there are tasks but no workers to process them
     work-stealing: True     # workers should steal tasks from each other
     work-stealing-interval: 100ms  # Callback time for work stealing
     worker-saturation: 1.1 # Send this fraction of nthreads root tasks to workers

distributed/scheduler.py

@@ -3517,6 +3517,12 @@
     default_port = 8786
     _instances: ClassVar[weakref.WeakSet[Scheduler]] = weakref.WeakSet()
 
+    worker_ttl: float | None
+    idle_since: float | None
+    idle_timeout: float | None
+    _no_workers_since: float | None  # Note: not None iff there are pending tasks
+    no_workers_timeout: float | None
+
     def __init__(
         self,
         loop=None,
@@ -3578,16 +3584,19 @@
         self.service_kwargs = service_kwargs or {}
         self.services = {}
         self.scheduler_file = scheduler_file
-        worker_ttl = worker_ttl or dask.config.get("distributed.scheduler.worker-ttl")
-        self.worker_ttl = parse_timedelta(worker_ttl) if worker_ttl else None
-        idle_timeout = idle_timeout or dask.config.get(
-            "distributed.scheduler.idle-timeout"
+
+        self.worker_ttl = parse_timedelta(
+            worker_ttl or dask.config.get("distributed.scheduler.worker-ttl")
+        )
+        self.idle_timeout = parse_timedelta(
+            idle_timeout or dask.config.get("distributed.scheduler.idle-timeout")
         )
-        if idle_timeout:
-            self.idle_timeout = parse_timedelta(idle_timeout)
-        else:
-            self.idle_timeout = None
         self.idle_since = time()
+        self.no_workers_timeout = parse_timedelta(
+            dask.config.get("distributed.scheduler.no-workers-timeout")
+        )
+        self._no_workers_since = None
+
         self.time_started = self.idle_since  # compatibility for dask-gateway
         self._replica_lock = RLock()
         self.bandwidth_workers = defaultdict(float)
@@ -3860,9 +3869,12 @@
             pc = PeriodicCallback(self.check_worker_ttl, self.worker_ttl * 1000)
             self.periodic_callbacks["worker-ttl"] = pc
 
-        pc = PeriodicCallback(self.check_idle, (self.idle_timeout or 1) * 1000 / 4)
+        pc = PeriodicCallback(self.check_idle, 250)
         self.periodic_callbacks["idle-timeout"] = pc
 
+        pc = PeriodicCallback(self._check_no_workers, 250)
+        self.periodic_callbacks["no-workers-timeout"] = pc
+
         if extensions is None:
             extensions = DEFAULT_EXTENSIONS.copy()
             if not dask.config.get("distributed.scheduler.work-stealing"):
@@ -8141,7 +8153,7 @@
 
     def check_idle(self) -> float | None:
         if self.status in (Status.closing, Status.closed):
-            return None
+            return None  # pragma: nocover
 
         if self.transition_counter != self._idle_transition_counter:
             self._idle_transition_counter = self.transition_counter
@@ -8178,6 +8190,33 @@
                 self._ongoing_background_tasks.call_soon(self.close)
         return self.idle_since
 
+    def _check_no_workers(self) -> None:
+        if self.status in (Status.closing, Status.closed):
+            return  # pragma: nocover
+
+        if (not self.queued and not self.unrunnable) or (self.queued and self.workers):
+            self._no_workers_since = None
+            return
+
+        # 1. There are queued or unrunnable tasks and no workers at all
+        # 2. There are unrunnable tasks and no workers satisfy their restrictions
+        # (Only rootish tasks can be queued, and rootish tasks can't have restrictions)
+
+        if not self._no_workers_since:
+            self._no_workers_since = time()
+            return
+
+        if (
+            self.no_workers_timeout
+            and time() > self._no_workers_since + self.no_workers_timeout
+        ):
+            logger.info(
+                "Tasks have been without any workers to run them for %s; "
+                "shutting scheduler down",
+                format_time(self.no_workers_timeout),
+            )
+            self._ongoing_background_tasks.call_soon(self.close)
+
     def adaptive_target(self, target_duration=None):
         """Desired number of workers based on the current workload
 

distributed/tests/test_scheduler.py

@@ -2405,11 +2405,11 @@ async def test_idle_timeout(c, s, a, b):
     pc.stop()
 
 
-@gen_cluster(
-    client=True,
-    nthreads=[],
-)
+@gen_cluster(client=True, nthreads=[])
 async def test_idle_timeout_no_workers(c, s):
+    """Test that idle-timeout is not triggered if there are no workers available
+    but there are tasks queued
+    """
     # Cancel the idle check periodic timeout so we can step through manually
     s.periodic_callbacks["idle-timeout"].stop()
 
@@ -2440,6 +2440,65 @@ async def test_idle_timeout_no_workers(c, s):
     assert s.check_idle()
 
 
+@gen_cluster(
+    client=True,
+    nthreads=[],
+    config={"distributed.scheduler.no-workers-timeout": None},
+)
+async def test_no_workers_timeout_disabled(c, s, a, b):
+    """no-workers-timeout has been disabled"""
+    future = c.submit(inc, 1, key="x")
+    await wait_for_state("x", ("queued", "no-worker"), s)
+    s._check_no_workers()
+    await asyncio.sleep(0.2)
+    assert s.status == Status.running
+
+
+@gen_cluster(
+    client=True,
+    nthreads=[],
+    config={"distributed.scheduler.no-workers-timeout": "100ms"},
+)
+async def test_no_workers_timeout_without_workers(c, s):
+    """Trip no-workers-timeout when there are no workers available"""
+    # Don't trip scheduler shutdown when there are no tasks
+    s._check_no_workers()
+    await asyncio.sleep(0.2)
+    assert s.status == Status.running
+
+    future = c.submit(inc, 1)
+    while s.status != Status.closed:
+        await asyncio.sleep(0.01)
+
+
+@gen_cluster(
+    client=True,
+    config={"distributed.scheduler.no-workers-timeout": "100ms"},
+)
+async def test_no_workers_timeout_bad_restrictions(c, s, a, b):
+    """Trip no-workers-timeout when there are workers available but none satisfies
+    task restrictions
+    """
+    future = c.submit(inc, 1, key="x", workers=["127.0.0.2:1234"])
+    while s.status != Status.closed:
+        await asyncio.sleep(0.01)
+
+
+@gen_cluster(client=True, nthreads=[("", 1)])
+async def test_no_workers_timeout_queued(c, s, a):
+    """Don't trip no-workers-timeout when there are queued tasks AND processing tasks"""
+    ev = Event()
+    futures = [c.submit(lambda ev: ev.wait(), ev, pure=False) for _ in range(3)]
+    while not a.state.tasks:
+        await asyncio.sleep(0.01)
+    assert s.queued or math.isinf(s.WORKER_SATURATION)
+
+    s._check_no_workers()
+    await asyncio.sleep(0.2)
+    assert s.status == Status.running
+    await ev.set()
+
+
 @gen_cluster(client=True, config={"distributed.scheduler.bandwidth": "100 GB"})
 async def test_bandwidth(c, s, a, b):
     start = s.bandwidth