[2.3] Improve dead client handling (#2501)

* improve dead client handling

* remove unused import

* fix docstring

---------

Co-authored-by: Yuan-Ting Hsieh (謝沅廷) <[email protected]>
Co-authored-by: Chester Chen <[email protected]>

Author: 3 people

nvflare/apis/fl_constant.py

@@ -236,6 +236,7 @@ class ServerCommandKey(object):
     CLIENTS = "clients"
     COLLECTOR = "collector"
     TURN_TO_COLD = "__turn_to_cold__"
+    REASON = "reason"
 
 
 class FedEventHeader(object):

nvflare/apis/impl/controller.py

@@ -80,7 +80,7 @@ def _get_client_task(target, task: Task):
 class _DeadClientStatus:
     def __init__(self):
         self.report_time = time.time()
-        self.dead_time = None
+        self.death_time = None
 
 
 class Controller(Responder, ControllerSpec, ABC):
@@ -356,8 +356,10 @@ def handle_dead_job(self, client_name: str, fl_ctx: FLContext):
         """
         # record the report and to be used by the task monitor
         with self._dead_clients_lock:
-            self.log_warning(fl_ctx, f"received dead job report for client {client_name}")
-            if not self._dead_clients.get(client_name):
+            if self._dead_clients.get(client_name):
+                # already on watch list
+                self.log_warning(fl_ctx, f"discarded dead job report for client {client_name}: already on watch list")
+            else:
                 self.log_warning(fl_ctx, f"client {client_name} is placed on dead client watch list")
                 self._dead_clients[client_name] = _DeadClientStatus()
 
@@ -856,12 +858,36 @@ def relay_and_wait(
         self.wait_for_task(task, abort_signal)
 
     def _monitor_tasks(self):
+        grace_period = ConfigService.get_float_var(name=_CONFIG_VAR_DEAD_CLIENT_GRACE_PERIOD, default=60.0)
         while not self._all_done:
-            clients_all_dead = self._check_dead_clients()
-            if not clients_all_dead:
+            self._determine_dead_clients(grace_period)
+            if self._all_clients_dead():
+                with self._engine.new_context() as fl_ctx:
+                    self.system_panic("All clients are dead", fl_ctx)
+                    return
+            else:
                 self._check_tasks()
             time.sleep(self._task_check_period)
 
+    def _determine_dead_clients(self, grace_period):
+        if not self._dead_clients:
+            return
+
+        now = time.time()
+        with self._dead_clients_lock:
+            for client_name, status in self._dead_clients.items():
+                if status.death_time:
+                    # already dead
+                    continue
+
+                if now - status.report_time < grace_period:
+                    # this report is still fresh - consider the client to be still alive
+                    continue
+
+                # consider client dead
+                status.death_time = now
+                self.logger.error(f"Client {client_name} is deemed dead!")
+
     def _check_tasks(self):
         with self._controller_lock:
             self._do_check_tasks()
@@ -897,7 +923,7 @@ def _do_check_tasks(self):
                 # check whether clients that the task is waiting are all dead
                 dead_clients = self._get_task_dead_clients(task)
                 if dead_clients:
-                    self.logger.info(f"client {dead_clients} is dead - set task {task.name} to TIMEOUT")
+                    self.logger.info(f"clients {dead_clients} dead - set task {task.name} to TIMEOUT")
                     task.completion_status = TaskCompletionStatus.CLIENT_DEAD
                     exit_tasks.append(task)
                     continue
@@ -949,22 +975,21 @@ def _get_task_dead_clients(self, task: Task):
             return None
 
         dead_clients = []
-        with self._dead_clients_lock:
-            for target in task.targets:
-                ct = _get_client_task(target, task)
-                if ct is not None and ct.result_received_time:
-                    # response has been received from this client
-                    continue
-
-                # either we have not sent the task to this client or we have not received response
-                # is the client already dead?
-                if self._client_still_alive(target):
-                    # this client is still alive
-                    # we let the task continue its course since we still have live clients
-                    return None
-                else:
-                    # this client is dead - remember it
-                    dead_clients.append(target)
+        for target in task.targets:
+            ct = _get_client_task(target, task)
+            if ct is not None and ct.result_received_time:
+                # response has been received from this client
+                continue
+
+            # either we have not sent the task to this client or we have not received response
+            # is the client already dead?
+            if self.get_client_death_time(target):
+                # this client is dead - remember it
+                dead_clients.append(target)
+            else:
+                # this client is still alive
+                # we let the task continue its course since we still have live clients
+                return None
 
         return dead_clients
 
@@ -993,46 +1018,18 @@ def wait_for_task(self, task: Task, abort_signal: Signal):
                 break
             time.sleep(self._task_check_period)
 
-    def _check_dead_clients(self):
+    def _all_clients_dead(self):
         if self._engine:
             clients = self._engine.get_clients()
-            dead_clients = []
-            with self._dead_clients_lock:
-                for client in clients:
-                    if not self._client_still_alive(client.name):
-                        dead_clients.append(client.name)
-
-            if dead_clients and len(clients) == len(dead_clients):
-                with self._engine.new_context() as fl_ctx:
-                    self.system_panic("All clients are dead", fl_ctx)
-                    return True
-        return False
-
-    def _client_still_alive(self, client_name):
-        now = time.time()
-        status = self._dead_clients.get(client_name, None)
-        grace_period = ConfigService.get_float_var(name=_CONFIG_VAR_DEAD_CLIENT_GRACE_PERIOD, default=60.0)
-
-        if not status:
-            # this client is still alive
-            return True
-
-        assert isinstance(status, _DeadClientStatus)
-        if status.dead_time:
-            return False
-
-        if now - status.report_time < grace_period:
-            # this report is still fresh - consider the client to be still alive
+            for client in clients:
+                if not self.get_client_death_time(client.name):
+                    # this client is still alive
+                    return False
             return True
-
-        # consider client dead
-        status.dead_time = now
-        self.logger.error(f"Client {client_name} is deemed dead!")
-        self.client_is_dead(client_name)
         return False
 
     def get_client_death_time(self, client_name: str):
-        """Get the time that the client was deemed dead
+        """Get the time that the client was deemed dead/disconnected
 
         Args:
             client_name: name of the client
@@ -1042,18 +1039,17 @@ def get_client_death_time(self, client_name: str):
         """
         status = self._dead_clients.get(client_name)
         if status:
-            assert isinstance(status, _DeadClientStatus)
-            return status.dead_time
+            return status.death_time
         return None
 
-    def process_job_heartbeat(self, fl_ctx: FLContext):
+    def process_job_heartbeat(self, fl_ctx: FLContext, reason: str):
         peer_ctx = fl_ctx.get_peer_context()
         assert isinstance(peer_ctx, FLContext)
         client_name = peer_ctx.get_identity_name()
         with self._dead_clients_lock:
             if client_name in self._dead_clients:
-                self.log_info(fl_ctx, f"Client {client_name} is removed from watch list")
+                self.log_info(fl_ctx, f"Client {client_name} is removed from watch list: {reason=}")
                 status = self._dead_clients.pop(client_name)
-                if status.dead_time:
-                    self.log_info(fl_ctx, f"Client {client_name} is revived")
+                if status.death_time:
+                    self.log_info(fl_ctx, f"Client {client_name} is revived: {reason=}")
                     self.client_is_revived(client_name)

nvflare/apis/responder.py

@@ -77,11 +77,12 @@ def process_task_check(self, task_id: str, fl_ctx: FLContext):
         pass
 
     @abstractmethod
-    def process_job_heartbeat(self, fl_ctx: FLContext):
+    def process_job_heartbeat(self, fl_ctx: FLContext, reason: str):
         """Called by the Engine to handle heartbeat received from clients.
 
         Args:
             fl_ctx: the FLContext
+            reason: reason of the HB
 
         Returns: None
 

nvflare/app_common/workflows/cyclic_ctl.py

@@ -139,15 +139,15 @@ def start_controller(self, fl_ctx: FLContext):
         self._last_client = None
 
     def _get_relay_orders(self, fl_ctx: FLContext) -> Union[List[Client], None]:
-        if len(self._participating_clients) <= 1:
-            self.system_panic(f"Not enough client sites ({len(self._participating_clients)}).", fl_ctx)
-            return None
-
         active_clients_map = {}
         for t in self._participating_clients:
             if not self.get_client_death_time(t.name):
                 active_clients_map[t.name] = t
 
+        if len(active_clients_map) <= 1:
+            self.system_panic(f"Not enough client sites (active_clients={len(active_clients_map)}).", fl_ctx)
+            return None
+
         if isinstance(self._order, list):
             targets = []
             for c_name in self._order:

nvflare/private/fed/server/fed_server.py

@@ -35,7 +35,6 @@
     WorkspaceConstants,
 )
 from nvflare.apis.fl_context import FLContext
-from nvflare.apis.shareable import Shareable
 from nvflare.apis.workspace import Workspace
 from nvflare.fuel.common.exit_codes import ProcessExitCode
 from nvflare.fuel.f3.cellnet.cell import Cell, Message
@@ -234,7 +233,7 @@ def fl_shutdown(self):
         self.shutdown = True
         start = time.time()
         while self.client_manager.clients:
-            # Wait for the clients to shutdown and quite first.
+            # Wait for the clients to shut down and quite first.
             time.sleep(0.1)
             if time.time() - start > self.shutdown_period:
                 self.logger.info("There are still clients connected. But shutdown the server after timeout.")
@@ -580,26 +579,17 @@ def _sync_client_jobs(self, request, client_token):
                     # this is a dict: token => nvflare.apis.client.Client
                     client = participating_clients.get(client_token, None)
                     if client:
-                        self._notify_dead_job(client, job_id)
+                        self._notify_dead_job(client, job_id, "missing job on client")
 
         return jobs_need_abort
 
-    def _notify_dead_job(self, client, job_id: str):
+    def _notify_dead_job(self, client, job_id: str, reason: str):
         try:
-            with self.engine.lock:
-                shareable = Shareable()
-                shareable.set_header(ServerCommandKey.FL_CLIENT, client.name)
-                fqcn = FQCN.join([FQCN.ROOT_SERVER, job_id])
-                request = new_cell_message({}, fobs.dumps(shareable))
-                self.cell.fire_and_forget(
-                    targets=fqcn,
-                    channel=CellChannel.SERVER_COMMAND,
-                    topic=ServerCommandNames.HANDLE_DEAD_JOB,
-                    message=request,
-                    optional=True,
-                )
-        except Exception:
-            self.logger.info("Could not connect to server runner process")
+            self.engine.notify_dead_job(job_id, client.name, reason)
+        except Exception as ex:
+            self.logger.info(
+                f"Failed to notify_dead_job to runner process of job {job_id}: {secure_format_exception(ex)}"
+            )
 
     def notify_dead_client(self, client):
         """Called to do further processing of the dead client
@@ -618,7 +608,7 @@ def notify_dead_client(self, client):
             assert isinstance(process_info, dict)
             participating_clients = process_info.get(RunProcessKey.PARTICIPANTS, None)
             if participating_clients and client.token in participating_clients:
-                self._notify_dead_job(client, job_id)
+                self._notify_dead_job(client, job_id, "client dead")
 
     def start_run(self, job_id, run_root, conf, args, snapshot):
         # Create the FL Engine

nvflare/private/fed/server/server_commands.py

@@ -264,6 +264,8 @@ def process(self, data: Shareable, fl_ctx: FLContext):
 
         """
         client_name = data.get_header(ServerCommandKey.FL_CLIENT)
+        reason = data.get_header(ServerCommandKey.REASON)
+        self.logger.warning(f"received dead job notification: {client_name=}; {reason=}")
         server_runner = fl_ctx.get_prop(FLContextKey.RUNNER)
         if server_runner:
             server_runner.handle_dead_job(client_name, fl_ctx)

nvflare/private/fed/server/server_engine.py

@@ -583,6 +583,19 @@ def update_job_run_status(self):
                 message=request,
             )
 
+    def notify_dead_job(self, job_id: str, client_name: str, reason: str):
+        shareable = Shareable()
+        shareable.set_header(ServerCommandKey.FL_CLIENT, client_name)
+        shareable.set_header(ServerCommandKey.REASON, reason)
+        self.send_command_to_child_runner_process(
+            job_id=job_id,
+            command_name=ServerCommandNames.HANDLE_DEAD_JOB,
+            command_data=shareable,
+            timeout=0.0,
+            optional=True,
+        )
+        self.logger.warning(f"notified SJ of dead-job: {job_id=}; {client_name=}; {reason=}")
+
     def send_command_to_child_runner_process(
         self, job_id: str, command_name: str, command_data, timeout=5.0, optional=False
     ):
@@ -594,7 +607,7 @@ def send_command_to_child_runner_process(
                     targets=fqcn,
                     channel=CellChannel.SERVER_COMMAND,
                     topic=command_name,
-                    request=request,
+                    message=request,
                     optional=optional,
                 )
                 return None

nvflare/private/fed/server/server_runner.py

@@ -329,7 +329,7 @@ def _try_to_get_task(self, client, fl_ctx, timeout=None, retry_interval=0.005):
                     return "", "", None
 
                 if self.current_wf.responder:
-                    self.current_wf.responder.process_job_heartbeat(fl_ctx=fl_ctx)
+                    self.current_wf.responder.process_job_heartbeat(fl_ctx=fl_ctx, reason="getTask")
 
                 task_name, task_id, task_data = self.current_wf.responder.process_task_request(client, fl_ctx)
 
@@ -371,7 +371,6 @@ def handle_dead_job(self, client_name: str, fl_ctx: FLContext):
             try:
                 if self.current_wf is None:
                     return
-
                 self.current_wf.responder.handle_dead_job(client_name=client_name, fl_ctx=fl_ctx)
             except Exception as e:
                 self.log_exception(
@@ -445,7 +444,7 @@ def process_submission(self, client: Client, task_name: str, task_id: str, resul
                     return
 
                 if self.current_wf.responder:
-                    self.current_wf.responder.process_job_heartbeat(fl_ctx)
+                    self.current_wf.responder.process_job_heartbeat(fl_ctx, "submitTask")
 
                 wf_id = result.get_cookie(ReservedHeaderKey.WORKFLOW, None)
                 if wf_id is not None and wf_id != self.current_wf.id:
@@ -508,7 +507,7 @@ def _handle_job_heartbeat(self, topic: str, request: Shareable, fl_ctx: FLContex
         self.log_debug(fl_ctx, "received client job_heartbeat")
         with self.wf_lock:
             if self.current_wf and self.current_wf.responder:
-                self.current_wf.responder.process_job_heartbeat(fl_ctx=fl_ctx)
+                self.current_wf.responder.process_job_heartbeat(fl_ctx=fl_ctx, reason="jobHeartbeat")
         return make_reply(ReturnCode.OK)
 
     def _handle_task_check(self, topic: str, request: Shareable, fl_ctx: FLContext) -> Shareable:
@@ -525,7 +524,7 @@ def _handle_task_check(self, topic: str, request: Shareable, fl_ctx: FLContext)
                 return make_reply(ReturnCode.TASK_UNKNOWN)
 
             if self.current_wf.responder:
-                self.current_wf.responder.process_job_heartbeat(fl_ctx)
+                self.current_wf.responder.process_job_heartbeat(fl_ctx, "taskCheck")
 
             # filter task result
             task = self.current_wf.responder.process_task_check(task_id=task_id, fl_ctx=fl_ctx)

nvflare/private/fed/server/sys_cmd.py

@@ -68,6 +68,14 @@ def get_spec(self):
                     authz_func=self.authorize_server_operation,
                     visible=True,
                 ),
+                CommandSpec(
+                    name="dead",
+                    description="send dead client msg to SJ",
+                    usage="dead <client-name>",
+                    handler_func=self.dead_client,
+                    authz_func=self.must_be_project_admin,
+                    visible=False,
+                ),
             ],
         )
 
@@ -156,3 +164,13 @@ def report_resources(self, conn: Connection, args: List[str]):
         table = conn.append_table(["Sites", "Resources"])
         for k, v in site_resources.items():
             table.add_row([str(k), str(v)])
+
+    def dead_client(self, conn: Connection, args: List[str]):
+        if len(args) != 3:
+            conn.append_error(f"Usage: {args[0]} client_name job_id")
+            return
+        client_name = args[1]
+        job_id = args[2]
+        engine = conn.app_ctx
+        engine.notify_dead_job(job_id, client_name, f"AdminCommand: {args[0]}")
+        conn.append_string(f"called notify_dead_job for client {client_name=} {job_id=}")