def stop_workers(self, actor_id, worker_ids): """Stop existing workers; used when updating an actor's image.""" logger.debug("Top of stop_workers() for actor: {}.".format(actor_id)) try: workers_dict = workers_store[actor_id] except KeyError: logger.debug("workers_store had no workers for actor: {}".format(actor_id)) workers_dict = {} # if there are existing workers, we need to close the actor message channel and # gracefully shutdown the existing worker processes. if len(workers_dict.items()) > 0: logger.info("Found {} workers to stop.".format(len(workers_dict.items()))) # first, close the actor msg channel to prevent any new messages from being pulled # by the old workers. actor_ch = ActorMsgChannel(actor_id) actor_ch.close() logger.info("Actor channel closed for actor: {}".format(actor_id)) # now, send messages to workers for a graceful shutdown: for _, worker in workers_dict.items(): # don't stop the new workers: if worker['id'] not in worker_ids: ch = WorkerChannel(worker_id=worker['id']) # since this is an update, there are new workers being started, so # don't delete the actor msg channel: ch.put('stop-no-delete') logger.info("Sent 'stop-no-delete' message to worker_id: {}".format(worker['id'])) ch.close() else: logger.debug("skipping worker {} as it it not in worker_ids.".format(worker)) else: logger.info("No workers to stop.")
def shutdown_worker(worker_id): """Gracefully shutdown a single worker.""" logger.debug("top of shutdown_worker for worker_id: {}".format(worker_id)) ch = WorkerChannel(worker_id=worker_id) ch.put("stop") logger.info("A 'stop' message was sent to worker: {}".format(worker_id)) ch.close()
def shutdown_worker(worker_id, delete_actor_ch=True): """Gracefully shutdown a single worker.""" logger.debug("top of shutdown_worker for worker_id: {}".format(worker_id)) ch = WorkerChannel(worker_id=worker_id) if not delete_actor_ch: ch.put("stop-no-delete") else: ch.put("stop") logger.info("A 'stop' message was sent to worker: {}".format(worker_id)) ch.close()
def shutdown_worker(actor_id, worker_id, delete_actor_ch=True): """Gracefully shutdown a single worker." actor_id (str) - the dbid of the associated actor. """ logger.debug("top of shutdown_worker for worker_id: {}".format(worker_id)) # set the worker status to SHUTDOWN_REQUESTED: try: Worker.update_worker_status(actor_id, worker_id, SHUTDOWN_REQUESTED) except Exception as e: logger.error(f"worker got exception trying to update status to SHUTODWN_REQUESTED. actor_id: {actor_id};" f"worker_id: {worker_id}; exception: {e}") ch = WorkerChannel(worker_id=worker_id) if not delete_actor_ch: ch.put("stop-no-delete") else: ch.put("stop") logger.info("A 'stop' message was sent to worker: {}".format(worker_id)) ch.close()
def check_workers(actor_id, ttl): """Check health of all workers for an actor.""" logger.info("Checking health for actor: {}".format(actor_id)) try: workers = Worker.get_workers(actor_id) except Exception as e: logger.error("Got exception trying to retrieve workers: {}".format(e)) return None logger.debug("workers: {}".format(workers)) host_id = os.environ.get('SPAWNER_HOST_ID', Config.get('spawner', 'host_id')) logger.debug("host_id: {}".format(host_id)) for worker in workers: worker_id = worker['id'] worker_status = worker.get('status') # if the worker has only been requested, it will not have a host_id. it is possible # the worker will ultimately get scheduled on a different host; however, if there is # some issue and the worker is "stuck" in the early phases, we should remove it.. if 'host_id' not in worker: # check for an old create time worker_create_t = worker.get('create_time') # in versions prior to 1.9, worker create_time was not set until after it was READY if not worker_create_t: hard_delete_worker( actor_id, worker_id, reason_str= 'Worker did not have a host_id or create_time field.') # if still no host after 5 minutes, delete it if worker_create_t < get_current_utc_time() - datetime.timedelta( minutes=5): hard_delete_worker( actor_id, worker_id, reason_str='Worker did not have a host_id and had ' 'old create_time field.') # ignore workers on different hosts because this health agent cannot interact with the # docker daemon responsible for the worker container.. if not host_id == worker['host_id']: continue # we need to delete any worker that is in SHUTDOWN REQUESTED or SHUTTING down for too long if worker_status == codes.SHUTDOWN_REQUESTED or worker_status == codes.SHUTTING_DOWN: worker_last_health_check_time = worker.get( 'last_health_check_time') if not worker_last_health_check_time: worker_last_health_check_time = worker.get('create_time') if not worker_last_health_check_time: hard_delete_worker( actor_id, worker_id, reason_str='Worker in SHUTDOWN and no health checks.') elif worker_last_health_check_time < get_current_utc_time( ) - datetime.timedelta(minutes=5): hard_delete_worker( actor_id, worker_id, reason_str='Worker in SHUTDOWN for too long.') # check if the worker has not responded to a health check recently; we use a relatively long period # (60 minutes) of idle health checks in case there is an issue with sending health checks through rabbitmq. # this needs to be watched closely though... worker_last_health_check_time = worker.get('last_health_check_time') if not worker_last_health_check_time or \ (worker_last_health_check_time < get_current_utc_time() - datetime.timedelta(minutes=60)): hard_delete_worker( actor_id, worker_id, reason_str='Worker has not health checked for too long.') # first send worker a health check logger.info(f"sending worker {worker_id} a health check") ch = WorkerChannel(worker_id=worker_id) try: logger.debug("Issuing status check to channel: {}".format( worker['ch_name'])) ch.put('status') except (channelpy.exceptions.ChannelTimeoutException, Exception) as e: logger.error( f"Got exception of type {type(e)} trying to send worker {worker_id} a " f"health check. e: {e}") finally: try: ch.close() except Exception as e: logger.error( "Got an error trying to close the worker channel for dead worker. Exception: {}" .format(e)) # now check if the worker has been idle beyond the max worker_ttl configured for this abaco: if ttl < 0: # ttl < 0 means infinite life logger.info("Infinite ttl configured; leaving worker") continue # we don't shut down workers that are currently running: if not worker['status'] == codes.BUSY: last_execution = worker.get('last_execution_time', 0) # if worker has made zero executions, use the create_time if last_execution == 0: last_execution = worker.get('create_time', datetime.datetime.min) logger.debug("using last_execution: {}".format(last_execution)) try: assert type(last_execution) == datetime.datetime except: logger.error( "Time received for TTL measurements is not of type datetime." ) last_execution = datetime.datetime.min if last_execution + datetime.timedelta( seconds=ttl) < datetime.datetime.utcnow(): # shutdown worker logger.info("Shutting down worker beyond ttl.") shutdown_worker(actor_id, worker['id']) else: logger.info("Still time left for this worker.") if worker['status'] == codes.ERROR: # shutdown worker logger.info("Shutting down worker in error status.") shutdown_worker(actor_id, worker['id'])
def check_workers(actor_id, ttl): """Check health of all workers for an actor.""" logger.info("Checking health for actor: {}".format(actor_id)) try: workers = Worker.get_workers(actor_id) except Exception as e: logger.error("Got exception trying to retrieve workers: {}".format(e)) return None logger.debug("workers: {}".format(workers)) host_id = os.environ.get('SPAWNER_HOST_ID', Config.get('spawner', 'host_id')) logger.debug("host_id: {}".format(host_id)) for worker in workers: # if the worker has only been requested, it will not have a host_id. if 'host_id' not in worker: # @todo- we will skip for now, but we need something more robust in case the worker is never claimed. continue # ignore workers on different hosts if not host_id == worker['host_id']: continue # first check if worker is responsive; if not, will need to manually kill logger.info("Checking health for worker: {}".format(worker)) ch = WorkerChannel(worker_id=worker['id']) worker_id = worker.get('id') result = None try: logger.debug("Issuing status check to channel: {}".format( worker['ch_name'])) result = ch.put_sync('status', timeout=5) except channelpy.exceptions.ChannelTimeoutException: logger.info( "Worker did not respond, removing container and deleting worker." ) try: rm_container(worker['cid']) except DockerError: pass try: Worker.delete_worker(actor_id, worker_id) logger.info("worker {} deleted from store".format(worker_id)) except Exception as e: logger.error( "Got exception trying to delete worker: {}".format(e)) # if the put_sync timed out and we removed the worker, we also need to delete the channel # otherwise the un-acked message will remain. try: ch.delete() except Exception as e: logger.error( "Got exception: {} while trying to delete worker channel for worker: {}" .format(e, worker_id)) finally: try: ch.close() except Exception as e: logger.error( "Got an error trying to close the worker channel for dead worker. Exception: {}" .format(e)) if result and not result == 'ok': logger.error( "Worker responded unexpectedly: {}, deleting worker.".format( result)) try: rm_container(worker['cid']) Worker.delete_worker(actor_id, worker_id) except Exception as e: logger.error( "Got error removing/deleting worker: {}".format(e)) else: # worker is healthy so update last health check: Worker.update_worker_health_time(actor_id, worker_id) logger.info("Worker ok.") # now check if the worker has been idle beyond the ttl: if ttl < 0: # ttl < 0 means infinite life logger.info("Infinite ttl configured; leaving worker") return # we don't shut down workers that are currently running: if not worker['status'] == codes.BUSY: last_execution = worker.get('last_execution_time', 0) # if worker has made zero executions, use the create_time if last_execution == 0: last_execution = worker.get('create_time', datetime.datetime.min) logger.debug("using last_execution: {}".format(last_execution)) try: assert type(last_execution) == datetime.datetime except: logger.error( "Time received for TTL measurements is not of type datetime." ) last_execution = datetime.datetime.min if last_execution + datetime.timedelta( seconds=ttl) < datetime.datetime.utcnow(): # shutdown worker logger.info("Shutting down worker beyond ttl.") shutdown_worker(actor_id, worker['id']) else: logger.info("Still time left for this worker.") if worker['status'] == codes.ERROR: # shutdown worker logger.info("Shutting down worker in error status.") shutdown_worker(actor_id, worker['id'])
try: main() except Exception as e: try: worker_id = os.environ.get('worker_id') except: logger.error( f"worker main thread got exception trying to get worker id from environment." f"not able to send stop-no-delete message to itself." f"worker_id: {worker_id}.") worker_id = '' if worker_id: try: ch = WorkerChannel(worker_id=worker_id) # since this is an exception, we don't know that the actor has been deleted # don't delete the actor msg channel: ch.put('stop-no-delete') logger.info( f"Worker main loop sent 'stop-no-delete' message to itself; worker_id: {worker_id}." ) ch.close() msg = "worker caught exception from main loop. worker exiting. e" \ "Exception: {} worker_id: {}".format(e, worker_id) logger.info(msg) except Exception as e: logger.error( f"worker main thread got exception trying to send stop-no-delete message to itself;" f"worker_id: {worker_id}.") keep_running = False sys.exit()
def check_workers(actor_id, ttl): """Check health of all workers for an actor.""" logger.info("Checking health for actor: {}".format(actor_id)) try: workers = Worker.get_workers(actor_id) except Exception as e: logger.error("Got exception trying to retrieve workers: {}".format(e)) return None logger.debug("workers: {}".format(workers)) for _, worker in workers.items(): # if the worker has only been requested, it will not have a host_id. if 'host_id' not in worker: # @todo- we will skip for now, but we need something more robust in case the worker is never claimed. continue # ignore workers on different hosts if not Config.get('spawner', 'host_id') == worker['host_id']: continue # first check if worker is responsive; if not, will need to manually kill logger.info("Checking health for worker: {}".format(worker)) ch = WorkerChannel(worker_id=worker['id']) worker_id = worker.get('id') try: logger.debug("Issuing status check to channel: {}".format( worker['ch_name'])) result = ch.put_sync('status', timeout=5) except channelpy.exceptions.ChannelTimeoutException: logger.info( "Worker did not respond, removing container and deleting worker." ) try: rm_container(worker['cid']) except DockerError: pass try: Worker.delete_worker(actor_id, worker_id) except Exception as e: logger.error( "Got exception trying to delete worker: {}".format(e)) finally: try: ch.close() except Exception as e: logger.error( "Got an error trying to close the worker channel for dead worker. Exception: {}" .format(e)) if not result == 'ok': logger.error( "Worker responded unexpectedly: {}, deleting worker.".format( result)) try: rm_container(worker['cid']) Worker.delete_worker(actor_id, worker_id) except Exception as e: logger.error( "Got error removing/deleting worker: {}".format(e)) else: # worker is healthy so update last health check: Worker.update_worker_health_time(actor_id, worker_id) logger.info("Worker ok.") # now check if the worker has been idle beyond the ttl: if ttl < 0: # ttl < 0 means infinite life logger.info("Infinite ttl configured; leaving worker") return # we don't shut down workers that are currently running: if not worker['status'] == codes.BUSY: last_execution = int(float(worker.get('last_execution_time', 0))) if last_execution + ttl < time.time(): # shutdown worker logger.info("Shutting down worker beyond ttl.") shutdown_worker(worker['id']) else: logger.info("Still time left for this worker.") elif worker['status'] == codes.ERROR: # shutdown worker logger.info("Shutting down worker in error status.") shutdown_worker(worker['id']) else: logger.debug("Worker not in READY status, will postpone.")