Use default scheduler port for LocalCluster

1. Unify multiple occurrences of 8786 into a single constant
   `DEFAULT_SCHEDULER_PORT`.
2. Change the default for the local cluster to use that instead
   of `0` (random port).
3. Introduce a fallback address for scheduler start in case of no
   port having been given.

distributed/cli/dask_scheduler.py

@@ -20,6 +20,7 @@
     enable_proctitle_on_children,
     enable_proctitle_on_current,
 )
+from distributed.scheduler import DEFAULT_SCHEDULER_PORT
 
 logger = logging.getLogger("distributed.scheduler")
 
@@ -165,9 +166,9 @@ def main(
 
     if port is None and (not host or not re.search(r":\d", host)):
         if isinstance(protocol, list):
-            port = [8786] + [0] * (len(protocol) - 1)
+            port = [DEFAULT_SCHEDULER_PORT] + [0] * (len(protocol) - 1)
         else:
-            port = 8786
+            port = DEFAULT_SCHEDULER_PORT
 
     if isinstance(protocol, list) or isinstance(port, list):
         if (not isinstance(protocol, list) or not isinstance(port, list)) or len(

distributed/cli/dask_ssh.py

@@ -8,6 +8,7 @@
 import click
 
 from distributed.deploy.old_ssh import SSHCluster
+from distributed.scheduler import DEFAULT_SCHEDULER_PORT
 
 logger = logging.getLogger("distributed.dask_ssh")
 
@@ -30,7 +31,7 @@
 )
 @click.option(
     "--scheduler-port",
-    default=8786,
+    default=DEFAULT_SCHEDULER_PORT,
     show_default=True,
     type=int,
     help="Specify scheduler port number.",

distributed/core.py

@@ -859,13 +859,30 @@ async def listen(self, port_or_addr=None, allow_offload=True, **kwargs):
         else:
             addr = port_or_addr
             assert isinstance(addr, str)
-        listener = await listen(
-            addr,
-            self.handle_comm,
-            deserialize=self.deserialize,
-            allow_offload=allow_offload,
-            **kwargs,
-        )
+        try:
+            listener = await listen(
+                addr,
+                self.handle_comm,
+                deserialize=self.deserialize,
+                allow_offload=allow_offload,
+                **kwargs,
+            )
+        except OSError:
+            fallback_port_or_addr = kwargs.get("fallback_port_or_addr", None)
+            if not fallback_port_or_addr:
+                raise
+            warnings.warn(
+                f"Address {addr} is already in use.\n"
+                f"Falling back to {fallback_port_or_addr} instead"
+            )
+            listener = await listen(
+                fallback_port_or_addr,
+                self.handle_comm,
+                deserialize=self.deserialize,
+                allow_offload=allow_offload,
+                **kwargs,
+            )
+
         self.listeners.append(listener)
 
     def handle_comm(self, comm: Comm) -> NoOpAwaitable:

distributed/deploy/local.py

@@ -120,7 +120,7 @@ def __init__(
         start=None,
         host=None,
         ip=None,
-        scheduler_port=0,
+        scheduler_port=None,
         silence_logs=logging.WARN,
         dashboard_address=":8787",
         worker_dashboard_address=None,

distributed/scheduler.py

@@ -195,6 +195,8 @@
     "stealing": WorkStealing,
 }
 
+DEFAULT_SCHEDULER_PORT = 8786
+
 
 class ClientState:
     """A simple object holding information about a client."""
@@ -3574,7 +3576,6 @@ class Scheduler(SchedulerState, ServerNode):
         Time we expect certain functions to take, e.g. ``{'sum': 0.25}``
     """
 
-    default_port = 8786
     _instances: ClassVar[weakref.WeakSet[Scheduler]] = weakref.WeakSet()
 
     worker_ttl: float | None
@@ -3686,8 +3687,18 @@ def __init__(
             interface=interface,
             protocol=protocol,
             security=security,
-            default_port=self.default_port,
+            default_port=DEFAULT_SCHEDULER_PORT,
         )
+        if port is None:
+            self._fallback_start_addresses = addresses_from_user_args(
+                host=host,
+                port=0,
+                interface=interface,
+                protocol=protocol,
+                security=security,
+            )
+        else:
+            self._fallback_start_addresses = []
 
         http_server_modules = dask.config.get("distributed.scheduler.http.routes")
         show_dashboard = dashboard or (dashboard is None and dashboard_address)
@@ -4094,11 +4105,14 @@ async def start_unsafe(self) -> Self:
 
         self._clear_task_state()
 
-        for addr in self._start_address:
+        for addr, fallback_addr in itertools.zip_longest(
+            self._start_address, self._fallback_start_addresses
+        ):
             await self.listen(
                 addr,
                 allow_offload=False,
                 handshake_overrides={"pickle-protocol": 4, "compression": None},
+                fallback_port_or_addr=fallback_addr,
                 **self.security.get_listen_args("scheduler"),
             )
             self.ip = get_address_host(self.listen_address)

distributed/utils_test.py

@@ -2530,7 +2530,7 @@ def _bind_port(port):
             s.listen(1)
             yield s
 
-    default_ports = [8786]
+    default_ports = [Scheduler.DEFAULT_SCHEDULER_PORT]
 
     while time() - start < _TEST_TIMEOUT:
         try: