Clean up cluster process reaping (#6840)

distributed/utils_test.py

@@ -568,14 +568,14 @@ def security():
     return tls_only_security()
 
 
-def _kill_join(proc, timeout):
-    proc.kill()
-    proc.join(timeout)
-    if proc.is_alive():
-        raise multiprocessing.TimeoutError(
-            f"Process {proc} did not shut down within {timeout}s"
-        )
-    proc.close()
+def _kill_join_processes(processes):
+    # Join may hang or cause issues, so make sure all are killed first.
+    # Note that we don't use a timeout, but rely on the overall pytest timeout.
+    for proc in processes:
+        proc.kill()
+    for proc in processes:
+        proc.join()
+        proc.close()
 
 
 def _close_queue(q):
@@ -590,15 +590,13 @@ def cluster(
     nanny=False,
     worker_kwargs=None,
     active_rpc_timeout=10,
-    shutdown_timeout=20,
     scheduler_kwargs=None,
     config=None,
 ):
     worker_kwargs = worker_kwargs or {}
     scheduler_kwargs = scheduler_kwargs or {}
     config = config or {}
 
-    ws = weakref.WeakSet()
     enable_proctitle_on_children()
 
     with check_process_leak(check=True), check_instances(), config_for_cluster_tests():
@@ -608,6 +606,8 @@ def cluster(
             _run_worker = run_worker
 
         with contextlib.ExitStack() as stack:
+            processes = []
+            stack.callback(_kill_join_processes, processes)
             # The scheduler queue will receive the scheduler's address
             scheduler_q = get_mp_context().Queue()
             stack.callback(_close_queue, scheduler_q)
@@ -620,9 +620,8 @@ def cluster(
                 kwargs=scheduler_kwargs,
                 daemon=True,
             )
-            ws.add(scheduler)
             scheduler.start()
-            stack.callback(_kill_join, scheduler, shutdown_timeout)
+            processes.append(scheduler)
 
             # Launch workers
             workers_by_pid = {}
@@ -642,9 +641,8 @@ def cluster(
                     args=(q, scheduler_q, config),
                     kwargs=kwargs,
                 )
-                ws.add(proc)
                 proc.start()
-                stack.callback(_kill_join, proc, shutdown_timeout)
+                processes.append(proc)
                 workers_by_pid[proc.pid] = {"proc": proc}
 
             saddr_or_exception = scheduler_q.get()