import copy import logging import math import operator import os import queue import subprocess import threading import time from collections import Counter, defaultdict, namedtuple from dataclasses import dataclass, field from enum import Enum from typing import Any, Callable, Dict, FrozenSet, List, Optional, Set, Tuple, Union import yaml import ray from ray._common.utils import PLACEMENT_GROUP_BUNDLE_RESOURCE_NAME from ray.autoscaler._private.constants import ( AUTOSCALER_HEARTBEAT_TIMEOUT_S, AUTOSCALER_MAX_CONCURRENT_LAUNCHES, AUTOSCALER_MAX_LAUNCH_BATCH, AUTOSCALER_MAX_NUM_FAILURES, AUTOSCALER_STATUS_LOG, AUTOSCALER_UPDATE_INTERVAL_S, DISABLE_LAUNCH_CONFIG_CHECK_KEY, DISABLE_NODE_UPDATERS_KEY, FOREGROUND_NODE_LAUNCH_KEY, WORKER_LIVENESS_CHECK_KEY, ) from ray.autoscaler._private.event_summarizer import EventSummarizer from ray.autoscaler._private.legacy_info_string import legacy_log_info_string from ray.autoscaler._private.load_metrics import LoadMetrics from ray.autoscaler._private.local.node_provider import ( LocalNodeProvider, record_local_head_state_if_needed, ) from ray.autoscaler._private.node_launcher import BaseNodeLauncher, NodeLauncher from ray.autoscaler._private.node_provider_availability_tracker import ( NodeAvailabilitySummary, NodeProviderAvailabilityTracker, ) from ray.autoscaler._private.node_tracker import NodeTracker from ray.autoscaler._private.prom_metrics import AutoscalerPrometheusMetrics from ray.autoscaler._private.providers import _get_node_provider from ray.autoscaler._private.resource_demand_scheduler import ( ResourceDemandScheduler, ResourceDict, get_bin_pack_residual, placement_groups_to_resource_demands, ) from ray.autoscaler._private.updater import NodeUpdaterThread from ray.autoscaler._private.util import ( ConcurrentCounter, NodeCount, NodeID, NodeIP, NodeType, NodeTypeConfigDict, format_info_string, hash_launch_conf, hash_runtime_conf, validate_config, with_head_node_ip, ) from ray.autoscaler.node_provider import NodeProvider from ray.autoscaler.tags import ( NODE_KIND_HEAD, NODE_KIND_UNMANAGED, NODE_KIND_WORKER, STATUS_UP_TO_DATE, STATUS_UPDATE_FAILED, TAG_RAY_FILE_MOUNTS_CONTENTS, TAG_RAY_LAUNCH_CONFIG, TAG_RAY_NODE_KIND, TAG_RAY_NODE_STATUS, TAG_RAY_RUNTIME_CONFIG, TAG_RAY_USER_NODE_TYPE, ) from ray.exceptions import RpcError logger = logging.getLogger(__name__) # Status of a node e.g. "up-to-date", see ray/autoscaler/tags.py NodeStatus = str # Tuple of modified fields for the given node_id returned by should_update # that will be passed into a NodeUpdaterThread. UpdateInstructions = namedtuple( "UpdateInstructions", ["node_id", "setup_commands", "ray_start_commands", "docker_config"], ) NodeLaunchData = Tuple[NodeTypeConfigDict, NodeCount, Optional[NodeType]] @dataclass class AutoscalerSummary: active_nodes: Dict[NodeType, int] idle_nodes: Optional[Dict[NodeType, int]] pending_nodes: List[Tuple[NodeIP, NodeType, NodeStatus]] pending_launches: Dict[NodeType, int] failed_nodes: List[Tuple[NodeIP, NodeType]] node_availability_summary: NodeAvailabilitySummary = field( default_factory=lambda: NodeAvailabilitySummary({}) ) # A dictionary of node IP to a list of reasons the node is not idle. node_activities: Optional[Dict[str, Tuple[NodeIP, List[str]]]] = None pending_resources: Dict[str, int] = field(default_factory=lambda: {}) # A mapping from node name (the same key as `usage_by_node`) to node type. # Optional for deployment modes which have the concept of node types and # backwards compatibility. node_type_mapping: Optional[Dict[str, str]] = None # Whether the autoscaler summary is v1 or v2. legacy: bool = False class NonTerminatedNodes: """Class to extract and organize information on non-terminated nodes.""" def __init__(self, provider: NodeProvider): start_time = time.time() # All non-terminated nodes self.all_node_ids = provider.non_terminated_nodes({}) # Managed worker nodes (node kind "worker"): self.worker_ids: List[NodeID] = [] # The head node (node kind "head") self.head_id: Optional[NodeID] = None for node in self.all_node_ids: node_kind = provider.node_tags(node)[TAG_RAY_NODE_KIND] if node_kind == NODE_KIND_WORKER: self.worker_ids.append(node) elif node_kind == NODE_KIND_HEAD: self.head_id = node # Note: For typical use-cases, self.all_node_ids == self.worker_ids + # [self.head_id]. The difference being in the case of unmanaged nodes. # Record the time of the non_terminated nodes call. This typically # translates to a "describe" or "list" call on most cluster managers # which can be quite expensive. Note that we include the processing # time because on some clients, there may be pagination and the # underlying api calls may be done lazily. self.non_terminated_nodes_time = time.time() - start_time logger.info( f"The autoscaler took {round(self.non_terminated_nodes_time, 3)}" " seconds to fetch the list of non-terminated nodes." ) def remove_terminating_nodes(self, terminating_nodes: List[NodeID]) -> None: """Remove nodes we're in the process of terminating from internal state.""" def not_terminating(node): return node not in terminating_nodes self.worker_ids = list(filter(not_terminating, self.worker_ids)) self.all_node_ids = list(filter(not_terminating, self.all_node_ids)) # Whether a worker should be kept based on the min_workers and # max_workers constraints. # # keep: should keep the worker # terminate: should terminate the worker # decide_later: the worker can be terminated if needed KeepOrTerminate = Enum("KeepOrTerminate", "keep terminate decide_later") class StandardAutoscaler: """The autoscaling control loop for a Ray cluster. There are two ways to start an autoscaling cluster: manually by running `ray start --head --autoscaling-config=/path/to/config.yaml` on a instance that has permission to launch other instances, or you can also use `ray up /path/to/config.yaml` from your laptop, which will configure the right AWS/Cloud roles automatically. See the Ray documentation (https://docs.ray.io/en/latest/) for a full definition of autoscaling behavior. StandardAutoscaler's `update` method is periodically called in `monitor.py`'s monitoring loop. StandardAutoscaler is also used to bootstrap clusters (by adding workers until the cluster size that can handle the resource demand is met). """ def __init__( self, # TODO(ekl): require config reader to be a callable always. config_reader: Union[str, Callable[[], dict]], load_metrics: LoadMetrics, gcs_client: "ray._raylet.GcsClient", session_name: Optional[str] = None, max_launch_batch: int = AUTOSCALER_MAX_LAUNCH_BATCH, max_concurrent_launches: int = AUTOSCALER_MAX_CONCURRENT_LAUNCHES, max_failures: int = AUTOSCALER_MAX_NUM_FAILURES, process_runner: Any = subprocess, update_interval_s: int = AUTOSCALER_UPDATE_INTERVAL_S, prefix_cluster_info: bool = False, event_summarizer: Optional[EventSummarizer] = None, prom_metrics: Optional[AutoscalerPrometheusMetrics] = None, ): """Create a StandardAutoscaler. Args: config_reader: Path to a Ray Autoscaler YAML, or a function to read and return the latest config. load_metrics: Provides metrics for the Ray cluster. session_name: The current Ray session name when this autoscaler is deployed. max_launch_batch: Max number of nodes to launch in one request. max_concurrent_launches: Max number of nodes that can be concurrently launched. This value and `max_launch_batch` determine the number of batches that are used to launch nodes. max_failures: Number of failures that the autoscaler will tolerate before exiting. process_runner: Subproc-like interface used by the CommandRunner. update_interval_s: Seconds between running the autoscaling loop. prefix_cluster_info: Whether to add the cluster name to info strs. event_summarizer: Utility to consolidate duplicated messages. prom_metrics: Prometheus metrics for autoscaler-related operations. gcs_client: client for interactions with the GCS. Used to drain nodes before termination. """ if isinstance(config_reader, str): # Auto wrap with file reader. def read_fn(): with open(config_reader) as f: new_config = yaml.safe_load(f.read()) return new_config self.config_reader = read_fn else: self.config_reader = config_reader self.node_provider_availability_tracker = NodeProviderAvailabilityTracker() # Prefix each line of info string with cluster name if True self.prefix_cluster_info = prefix_cluster_info # Keep this before self.reset (self.provider needs to be created # exactly once). self.provider = None # Keep this before self.reset (if an exception occurs in reset # then prom_metrics must be instantitiated to increment the # exception counter) self.prom_metrics = prom_metrics or AutoscalerPrometheusMetrics( session_name=session_name ) # noqa self.resource_demand_scheduler = None self.reset(errors_fatal=True) self.load_metrics = load_metrics self.max_failures = max_failures self.max_launch_batch = max_launch_batch self.max_concurrent_launches = max_concurrent_launches self.process_runner = process_runner self.event_summarizer = event_summarizer or EventSummarizer() # Map from node_id to NodeUpdater threads self.updaters: Dict[NodeID, NodeUpdaterThread] = {} self.num_failed_updates: Dict[NodeID, int] = defaultdict(int) self.num_successful_updates: Dict[NodeID, int] = defaultdict(int) self.num_failures = 0 self.last_update_time = 0.0 self.update_interval_s = update_interval_s # Keeps track of pending and running nodes self.non_terminated_nodes: Optional[NonTerminatedNodes] = None # Tracks nodes scheduled for termination self.nodes_to_terminate: List[NodeID] = [] # Disable NodeUpdater threads if true. # Should be set to true in situations where another component, such as # a Kubernetes operator, is responsible for Ray setup on nodes. self.disable_node_updaters = self.config["provider"].get( DISABLE_NODE_UPDATERS_KEY, False ) logger.info(f"{DISABLE_NODE_UPDATERS_KEY}:{self.disable_node_updaters}") # Disable launch configuration checking if set to true. # This setting is used in scenarios where there is no meaningful node type # to enforce, such as in fake multinode situations. When this option is enabled, # outdated nodes will not be terminated. self.disable_launch_config_check = self.config["provider"].get( DISABLE_LAUNCH_CONFIG_CHECK_KEY, False ) logger.info( f"{DISABLE_LAUNCH_CONFIG_CHECK_KEY}:{self.disable_launch_config_check}" ) # By default, the autoscaler launches nodes in batches asynchronously in # background threads. # When the following flag is set, that behavior is disabled, so that nodes # are launched in the main thread, all in one batch, blocking until all # NodeProvider.create_node calls have returned. self.foreground_node_launch = self.config["provider"].get( FOREGROUND_NODE_LAUNCH_KEY, False ) logger.info(f"{FOREGROUND_NODE_LAUNCH_KEY}:{self.foreground_node_launch}") # By default, the autoscaler kills and/or tries to recover # a worker node if it hasn't produced a resource heartbeat in the last 30 # seconds. The worker_liveness_check flag allows disabling this behavior in # settings where another component, such as a Kubernetes operator, is # responsible for healthchecks. self.worker_liveness_check = self.config["provider"].get( WORKER_LIVENESS_CHECK_KEY, True ) logger.info(f"{WORKER_LIVENESS_CHECK_KEY}:{self.worker_liveness_check}") # Node launchers self.foreground_node_launcher: Optional[BaseNodeLauncher] = None self.launch_queue: Optional[queue.Queue[NodeLaunchData]] = None self.pending_launches = ConcurrentCounter() if self.foreground_node_launch: self.foreground_node_launcher = BaseNodeLauncher( provider=self.provider, pending=self.pending_launches, event_summarizer=self.event_summarizer, node_provider_availability_tracker=self.node_provider_availability_tracker, # noqa: E501 Flake and black disagree how to format this. session_name=session_name, node_types=self.available_node_types, prom_metrics=self.prom_metrics, ) else: self.launch_queue = queue.Queue() max_batches = math.ceil(max_concurrent_launches / float(max_launch_batch)) for i in range(int(max_batches)): node_launcher = NodeLauncher( provider=self.provider, queue=self.launch_queue, index=i, pending=self.pending_launches, event_summarizer=self.event_summarizer, node_provider_availability_tracker=self.node_provider_availability_tracker, # noqa: E501 Flake and black disagreee how to format this. session_name=session_name, node_types=self.available_node_types, prom_metrics=self.prom_metrics, ) node_launcher.daemon = True node_launcher.start() # NodeTracker maintains soft state to track the number of recently # failed nodes. It is best effort only. self.node_tracker = NodeTracker() # Expand local file_mounts to allow ~ in the paths. This can't be done # earlier when the config is written since we might be on different # platform and the expansion would result in wrong path. self.config["file_mounts"] = { remote: os.path.expanduser(local) for remote, local in self.config["file_mounts"].items() } self.gcs_client = gcs_client for local_path in self.config["file_mounts"].values(): assert os.path.exists(local_path) logger.info("StandardAutoscaler: {}".format(self.config)) @property def all_node_types(self) -> Set[str]: return self.config["available_node_types"].keys() def update(self): try: self.reset(errors_fatal=False) self._update() except Exception as e: self.prom_metrics.update_loop_exceptions.inc() logger.exception("StandardAutoscaler: Error during autoscaling.") self.num_failures += 1 if self.num_failures > self.max_failures: logger.critical("StandardAutoscaler: Too many errors, abort.") raise e def _update(self): # For type checking, assert that these objects have been instantitiated. assert self.provider assert self.resource_demand_scheduler now = time.time() # Throttle autoscaling updates to this interval to avoid exceeding # rate limits on API calls. if now - self.last_update_time < self.update_interval_s: return self.last_update_time = now # Query the provider to update the list of non-terminated nodes self.non_terminated_nodes = NonTerminatedNodes(self.provider) # Back off the update if the provider says it's not safe to proceed. if not self.provider.safe_to_scale(): logger.info( "Backing off of autoscaler update." f" Will try again in {self.update_interval_s} seconds." ) return # This will accumulate the nodes we need to terminate. self.nodes_to_terminate = [] # Update status strings if AUTOSCALER_STATUS_LOG: logger.info(self.info_string()) legacy_log_info_string(self, self.non_terminated_nodes.worker_ids) if not self.provider.is_readonly(): self.terminate_nodes_to_enforce_config_constraints(now) if self.disable_node_updaters: # Don't handle unhealthy nodes if the liveness check is disabled. # self.worker_liveness_check is True by default. if self.worker_liveness_check: self.terminate_unhealthy_nodes(now) else: self.process_completed_updates() self.update_nodes() # Don't handle unhealthy nodes if the liveness check is disabled. # self.worker_liveness_check is True by default. if self.worker_liveness_check: self.attempt_to_recover_unhealthy_nodes(now) self.set_prometheus_updater_data() # Update running nodes gauge num_workers = len(self.non_terminated_nodes.worker_ids) self.prom_metrics.running_workers.set(num_workers) # Remove IPs from LoadMetrics that are not known to the NodeProvider. self.load_metrics.prune_active_ips( active_ips=[ self.provider.internal_ip(node_id) for node_id in self.non_terminated_nodes.all_node_ids ] ) # Dict[NodeType, int], List[ResourceDict] to_launch, unfulfilled = self.resource_demand_scheduler.get_nodes_to_launch( self.non_terminated_nodes.all_node_ids, self.pending_launches.breakdown(), self.load_metrics.get_resource_demand_vector(), self.load_metrics.get_resource_utilization(), self.load_metrics.get_pending_placement_groups(), self.load_metrics.get_static_node_resources_by_ip(), ensure_min_cluster_size=self.load_metrics.get_resource_requests(), node_availability_summary=self.node_provider_availability_tracker.summary(), ) self._report_pending_infeasible(unfulfilled) if not self.provider.is_readonly(): self.launch_required_nodes(to_launch) # Execute optional end-of-update logic. # Keep this method call at the end of autoscaler._update(). self.provider.post_process() # Record the amount of time the autoscaler took for # this _update() iteration. update_time = time.time() - self.last_update_time logger.info( f"The autoscaler took {round(update_time, 3)}" " seconds to complete the update iteration." ) self.prom_metrics.update_time.observe(update_time) def terminate_nodes_to_enforce_config_constraints(self, now: float): """Terminates nodes to enforce constraints defined by the autoscaling config. (1) Terminates nodes in excess of `max_workers`. (2) Terminates nodes idle for longer than `idle_timeout_minutes`. (3) Terminates outdated nodes, namely nodes whose configs don't match `node_config` for the relevant node type. Avoids terminating non-outdated nodes required by autoscaler.sdk.request_resources(). """ # For type checking, assert that these objects have been instantitiated. assert self.non_terminated_nodes assert self.provider last_used = self.load_metrics.ray_nodes_last_used_time_by_ip idle_timeout_s = 60 * self.config["idle_timeout_minutes"] last_used_cutoff = now - idle_timeout_s # Sort based on last used to make sure to keep min_workers that # were most recently used. Otherwise, _keep_min_workers_of_node_type # might keep a node that should be terminated. sorted_node_ids = self._sort_based_on_last_used( self.non_terminated_nodes.worker_ids, last_used ) # Don't terminate nodes needed by request_resources() nodes_not_allowed_to_terminate = self._get_nodes_needed_for_request_resources( sorted_node_ids ) # Tracks counts of nodes we intend to keep for each node type. node_type_counts = defaultdict(int) def keep_node(node_id: NodeID) -> None: assert self.provider # Update per-type counts. tags = self.provider.node_tags(node_id) if TAG_RAY_USER_NODE_TYPE in tags: node_type = tags[TAG_RAY_USER_NODE_TYPE] node_type_counts[node_type] += 1 # Nodes that we could terminate, if needed. nodes_we_could_terminate: List[NodeID] = [] for node_id in sorted_node_ids: # Make sure to not kill idle node types if the number of workers # of that type is lower/equal to the min_workers of that type # or it is needed for request_resources(). should_keep_or_terminate, reason = self._keep_worker_of_node_type( node_id, node_type_counts ) if should_keep_or_terminate == KeepOrTerminate.terminate: self.schedule_node_termination(node_id, reason, logger.info) continue if ( should_keep_or_terminate == KeepOrTerminate.keep or node_id in nodes_not_allowed_to_terminate ) and self.launch_config_ok(node_id): keep_node(node_id) continue node_ip = self.provider.internal_ip(node_id) if node_ip in last_used and last_used[node_ip] < last_used_cutoff: self.schedule_node_termination(node_id, "idle", logger.info) # Get the local time of the node's last use as a string. formatted_last_used_time = time.asctime( time.localtime(last_used[node_ip]) ) logger.info(f"Node last used: {formatted_last_used_time}.") # Note that the current time will appear in the log prefix. elif not self.launch_config_ok(node_id): self.schedule_node_termination(node_id, "outdated", logger.info) else: keep_node(node_id) nodes_we_could_terminate.append(node_id) # Terminate nodes if there are too many num_workers = len(self.non_terminated_nodes.worker_ids) num_extra_nodes_to_terminate = ( num_workers - len(self.nodes_to_terminate) - self.config["max_workers"] ) if num_extra_nodes_to_terminate > len(nodes_we_could_terminate): logger.warning( "StandardAutoscaler: trying to terminate " f"{num_extra_nodes_to_terminate} nodes, while only " f"{len(nodes_we_could_terminate)} are safe to terminate." " Inconsistent config is likely." ) num_extra_nodes_to_terminate = len(nodes_we_could_terminate) # If num_extra_nodes_to_terminate is negative or zero, # we would have less than max_workers nodes after terminating # nodes_to_terminate and we do not need to terminate anything else. if num_extra_nodes_to_terminate > 0: extra_nodes_to_terminate = nodes_we_could_terminate[ -num_extra_nodes_to_terminate: ] for node_id in extra_nodes_to_terminate: self.schedule_node_termination(node_id, "max workers", logger.info) self.terminate_scheduled_nodes() def schedule_node_termination( self, node_id: NodeID, reason_opt: Optional[str], logger_method: Callable ) -> None: # For type checking, assert that this object has been instantitiated. assert self.provider if reason_opt is None: raise Exception("reason should be not None.") reason: str = reason_opt node_ip = self.provider.internal_ip(node_id) # Log, record an event, and add node_id to nodes_to_terminate. logger_method( "StandardAutoscaler: " f"Terminating the node with id {node_id}" f" and ip {node_ip}." f" ({reason})" ) self.event_summarizer.add( "Removing {} nodes of type " + self._get_node_type(node_id) + " ({}).".format(reason), quantity=1, aggregate=operator.add, ) self.nodes_to_terminate.append(node_id) def terminate_scheduled_nodes(self): """Terminate scheduled nodes and clean associated autoscaler state.""" # For type checking, assert that these objects have been instantitiated. assert self.provider assert self.non_terminated_nodes if not self.nodes_to_terminate: return # Drain the nodes self.drain_nodes_via_gcs(self.nodes_to_terminate) # Terminate the nodes self.provider.terminate_nodes(self.nodes_to_terminate) for node in self.nodes_to_terminate: self.node_tracker.untrack(node) self.prom_metrics.stopped_nodes.inc() # Update internal node lists self.non_terminated_nodes.remove_terminating_nodes(self.nodes_to_terminate) self.nodes_to_terminate = [] def drain_nodes_via_gcs(self, provider_node_ids_to_drain: List[NodeID]): """Send an RPC request to the GCS to drain (prepare for termination) the nodes with the given node provider ids. note: The current implementation of DrainNode on the GCS side is to de-register and gracefully shut down the Raylets. In the future, the behavior may change to better reflect the name "Drain." See https://github.com/ray-project/ray/pull/19350. """ # For type checking, assert that this object has been instantitiated. assert self.provider # The GCS expects Node ids in the request, rather than NodeProvider # ids. To get the Node ids of the nodes to we're draining, we make # the following translations of identifiers: # node provider node id -> ip -> node id # Convert node provider node ids to ips. node_ips = set() failed_ip_fetch = False for provider_node_id in provider_node_ids_to_drain: # If the provider's call to fetch ip fails, the exception is not # fatal. Log the exception and proceed. try: ip = self.provider.internal_ip(provider_node_id) node_ips.add(ip) except Exception: logger.exception( "Failed to get ip of node with id" f" {provider_node_id} during scale-down." ) failed_ip_fetch = True if failed_ip_fetch: self.prom_metrics.drain_node_exceptions.inc() # Only attempt to drain connected nodes, i.e. nodes with ips in # LoadMetrics. connected_node_ips = node_ips & self.load_metrics.node_id_by_ip.keys() # Convert ips to Node ids. # (The assignment ip->node_id is well-defined under current # assumptions. See "use_node_id_as_ip" in monitor.py) node_ids_to_drain = { self.load_metrics.node_id_by_ip[ip] for ip in connected_node_ips } if not node_ids_to_drain: return logger.info(f"Draining {len(node_ids_to_drain)} raylet(s).") try: # A successful response indicates that the GCS has marked the # desired nodes as "drained." The cloud provider can then terminate # the nodes without the GCS printing an error. # Check if we succeeded in draining all of the intended nodes by # looking at the RPC response. drained_node_ids = set( self.gcs_client.drain_nodes(node_ids_to_drain, timeout=5) ) failed_to_drain = node_ids_to_drain - drained_node_ids if failed_to_drain: self.prom_metrics.drain_node_exceptions.inc() logger.error(f"Failed to drain {len(failed_to_drain)} raylet(s).") # If we get a gRPC error with an UNIMPLEMENTED code, fail silently. # This error indicates that the GCS is using Ray version < 1.8.0, # for which DrainNode is not implemented. except RpcError as e: # If the code is UNIMPLEMENTED, pass. if e.rpc_code == ray._raylet.GRPC_STATUS_CODE_UNIMPLEMENTED: pass # Otherwise, it's a plain old gRPC error and we should log it. else: self.prom_metrics.drain_node_exceptions.inc() logger.exception("Failed to drain Ray nodes. Traceback follows.") except Exception: # We don't need to interrupt the autoscaler update with an # exception, but we should log what went wrong and record the # failure in Prometheus. self.prom_metrics.drain_node_exceptions.inc() logger.exception("Failed to drain Ray nodes. Traceback follows.") def launch_required_nodes(self, to_launch: Dict[NodeType, int]) -> None: if to_launch: for node_type, count in to_launch.items(): self.launch_new_node(count, node_type=node_type) def update_nodes(self): """Run NodeUpdaterThreads to run setup commands, sync files, and/or start Ray. """ # Update nodes with out-of-date files. # TODO(edoakes): Spawning these threads directly seems to cause # problems. They should at a minimum be spawned as daemon threads. # See https://github.com/ray-project/ray/pull/5903 for more info. T = [] for node_id, setup_commands, ray_start_commands, docker_config in ( self.should_update(node_id) for node_id in self.non_terminated_nodes.worker_ids ): if node_id is not None: resources = self._node_resources(node_id) labels = self._node_labels(node_id) logger.debug(f"{node_id}: Starting new thread runner.") T.append( threading.Thread( target=self.spawn_updater, args=( node_id, setup_commands, ray_start_commands, resources, labels, docker_config, ), ) ) for t in T: t.start() for t in T: t.join() def process_completed_updates(self): """Clean up completed NodeUpdaterThreads.""" completed_nodes = [] for node_id, updater in self.updaters.items(): if not updater.is_alive(): completed_nodes.append(node_id) if completed_nodes: failed_nodes = [] for node_id in completed_nodes: updater = self.updaters[node_id] if updater.exitcode == 0: self.num_successful_updates[node_id] += 1 self.prom_metrics.successful_updates.inc() if updater.for_recovery: self.prom_metrics.successful_recoveries.inc() if updater.update_time: self.prom_metrics.worker_update_time.observe( updater.update_time ) # Mark the node as active to prevent the node recovery # logic immediately trying to restart Ray on the new node. self.load_metrics.mark_active(self.provider.internal_ip(node_id)) else: failed_nodes.append(node_id) self.num_failed_updates[node_id] += 1 self.prom_metrics.failed_updates.inc() if updater.for_recovery: self.prom_metrics.failed_recoveries.inc() self.node_tracker.untrack(node_id) del self.updaters[node_id] if failed_nodes: # Some nodes in failed_nodes may already have been terminated # during an update (for being idle after missing a heartbeat). # Update the list of non-terminated workers. for node_id in failed_nodes: # Check if the node has already been terminated. if node_id in self.non_terminated_nodes.worker_ids: self.schedule_node_termination( node_id, "launch failed", logger.error ) else: logger.warning( f"StandardAutoscaler: {node_id}:" " Failed to update node." " Node has already been terminated." ) self.terminate_scheduled_nodes() def set_prometheus_updater_data(self): """Record total number of active NodeUpdaterThreads and how many of these are being run to recover nodes. """ self.prom_metrics.updating_nodes.set(len(self.updaters)) num_recovering = 0 for updater in self.updaters.values(): if updater.for_recovery: num_recovering += 1 self.prom_metrics.recovering_nodes.set(num_recovering) def _report_pending_infeasible(self, unfulfilled: List[ResourceDict]): """Emit event messages for infeasible or unschedulable tasks. This adds messages to the event summarizer for warning on infeasible or "cluster full" resource requests. Args: unfulfilled: List of resource demands that would be unfulfilled even after full scale-up. """ # For type checking, assert that this object has been instantitiated. assert self.resource_demand_scheduler infeasible = [] for bundle in unfulfilled: placement_group = any( "_group_" in k or k == PLACEMENT_GROUP_BUNDLE_RESOURCE_NAME for k in bundle ) if placement_group: continue if not self.resource_demand_scheduler.is_feasible(bundle): infeasible.append(bundle) if infeasible: for request in infeasible: self.event_summarizer.add_once_per_interval( "Error: No available node types can fulfill resource " "request {}. Add suitable node types to this cluster to " "resolve this issue.".format(request), key="infeasible_{}".format(sorted(request.items())), interval_s=30, ) def _sort_based_on_last_used( self, nodes: List[NodeID], last_used: Dict[str, float] ) -> List[NodeID]: """Sort the nodes based on the last time they were used. The first item in the return list is the most recently used. """ last_used_copy = copy.deepcopy(last_used) # Add the unconnected nodes as the least recently used (the end of # list). This prioritizes connected nodes. least_recently_used = -1 def last_time_used(node_id: NodeID): assert self.provider node_ip = self.provider.internal_ip(node_id) if node_ip not in last_used_copy: return least_recently_used else: return last_used_copy[node_ip] return sorted(nodes, key=last_time_used, reverse=True) def _get_nodes_needed_for_request_resources( self, sorted_node_ids: List[NodeID] ) -> FrozenSet[NodeID]: # TODO(ameer): try merging this with resource_demand_scheduler # code responsible for adding nodes for get_resource_requests() and get_pending_placement_groups(). """Returns the nodes NOT allowed to terminate due to get_resource_requests() and get_pending_placement_groups(). Args: sorted_node_ids: the node ids sorted based on last used (LRU last). Returns: FrozenSet[NodeID]: a set of nodes (node ids) that we should NOT terminate. """ # For type checking, assert that this object has been instantitiated. assert self.provider nodes_not_allowed_to_terminate: Set[NodeID] = set() resource_demands, strict_spreads = placement_groups_to_resource_demands( self.load_metrics.get_pending_placement_groups() ) resource_demands.extend(self.load_metrics.get_resource_requests()) if not resource_demands and not strict_spreads: return frozenset(nodes_not_allowed_to_terminate) static_node_resources: Dict[ NodeIP, ResourceDict ] = self.load_metrics.get_static_node_resources_by_ip() head_node_resources: ResourceDict = copy.deepcopy( self.available_node_types[self.config["head_node_type"]]["resources"] ) # TODO(ameer): this is somewhat duplicated in # resource_demand_scheduler.py. if not head_node_resources: # Legacy yaml might include {} in the resources field. head_node_ip = self.provider.internal_ip(self.non_terminated_nodes.head_id) head_node_resources = static_node_resources.get(head_node_ip, {}) node_total_resources: List[ResourceDict] = [head_node_resources] resource_demand_vector_worker_node_ids = [] # Get max resources on all the non terminated nodes. for node_id in sorted_node_ids: tags = self.provider.node_tags(node_id) if TAG_RAY_USER_NODE_TYPE in tags: node_type = tags[TAG_RAY_USER_NODE_TYPE] node_resources: ResourceDict = copy.deepcopy( self.available_node_types[node_type]["resources"] ) if not node_resources: # Legacy yaml might include {} in the resources field. node_ip = self.provider.internal_ip(node_id) node_resources = static_node_resources.get(node_ip, {}) node_total_resources.append(node_resources) resource_demand_vector_worker_node_ids.append(node_id) # Since it is sorted based on last used, we "keep" nodes that are # most recently used when we binpack. We assume get_bin_pack_residual # is following the given order here. node_remaining_resources = copy.deepcopy(node_total_resources) for strict_spread in strict_spreads: unfulfilled, updated_node_remaining_resources = get_bin_pack_residual( node_remaining_resources, strict_spread, strict_spread=True ) if unfulfilled: continue node_remaining_resources = updated_node_remaining_resources _, node_remaining_resources = get_bin_pack_residual( node_remaining_resources, resource_demands ) # Remove the first entry (the head node). node_total_resources.pop(0) # Remove the first entry (the head node). node_remaining_resources.pop(0) for i, node_id in enumerate(resource_demand_vector_worker_node_ids): if ( node_remaining_resources[i] == node_total_resources[i] and node_total_resources[i] ): # No resources of the node were needed for request_resources(). # node_total_resources[i] is an empty dict for legacy yamls # before the node is connected. pass else: nodes_not_allowed_to_terminate.add(node_id) return frozenset(nodes_not_allowed_to_terminate) def _keep_worker_of_node_type( self, node_id: NodeID, node_type_counts: Dict[NodeType, int] ) -> Tuple[KeepOrTerminate, Optional[str]]: """Determines if a worker should be kept based on the min_workers and max_workers constraint of the worker's node_type. Returns KeepOrTerminate.keep when both of the following hold: (a) The worker's node_type is present among the keys of the current config's available_node_types dict. (b) Deleting the node would violate the min_workers constraint for that worker's node_type. Returns KeepOrTerminate.terminate when both the following hold: (a) The worker's node_type is not present among the keys of the current config's available_node_types dict. (b) Keeping the node would violate the max_workers constraint for that worker's node_type. Return KeepOrTerminate.decide_later otherwise. Args: node_type_counts(Dict[NodeType, int]): The non_terminated node types counted so far. Returns: KeepOrTerminate: keep if the node should be kept, terminate if the node should be terminated, decide_later if we are allowed to terminate it, but do not have to. Optional[str]: reason for termination. Not None on KeepOrTerminate.terminate, None otherwise. """ # For type checking, assert that this object has been instantitiated. assert self.provider tags = self.provider.node_tags(node_id) if TAG_RAY_USER_NODE_TYPE in tags: node_type = tags[TAG_RAY_USER_NODE_TYPE] min_workers = self.available_node_types.get(node_type, {}).get( "min_workers", 0 ) max_workers = self.available_node_types.get(node_type, {}).get( "max_workers", 0 ) if node_type not in self.available_node_types: # The node type has been deleted from the cluster config. # Allow terminating it if needed. available_node_types = list(self.available_node_types.keys()) return ( KeepOrTerminate.terminate, f"not in available_node_types: {available_node_types}", ) new_count = node_type_counts[node_type] + 1 if new_count <= min(min_workers, max_workers): return KeepOrTerminate.keep, None if new_count > max_workers: return KeepOrTerminate.terminate, "max_workers_per_type" return KeepOrTerminate.decide_later, None def _node_resources(self, node_id): node_type = self.provider.node_tags(node_id).get(TAG_RAY_USER_NODE_TYPE) if self.available_node_types: return self.available_node_types.get(node_type, {}).get("resources", {}) else: return {} def _node_labels(self, node_id): node_type = self.provider.node_tags(node_id).get(TAG_RAY_USER_NODE_TYPE) if self.available_node_types: return self.available_node_types.get(node_type, {}).get("labels", {}) else: return {} def reset(self, errors_fatal=False): sync_continuously = False if hasattr(self, "config"): sync_continuously = self.config.get("file_mounts_sync_continuously", False) try: new_config = self.config_reader() if new_config != getattr(self, "config", None): try: validate_config(new_config) except Exception as e: self.prom_metrics.config_validation_exceptions.inc() logger.debug( "Cluster config validation failed. The version of " "the ray CLI you launched this cluster with may " "be higher than the version of ray being run on " "the cluster. Some new features may not be " "available until you upgrade ray on your cluster.", exc_info=e, ) logger.debug( f"New config after validation: {new_config}," f" of type: {type(new_config)}" ) (new_runtime_hash, new_file_mounts_contents_hash) = hash_runtime_conf( new_config["file_mounts"], new_config["cluster_synced_files"], [ new_config["worker_setup_commands"], new_config["worker_start_ray_commands"], ], generate_file_mounts_contents_hash=sync_continuously, ) self.config = new_config self.runtime_hash = new_runtime_hash self.file_mounts_contents_hash = new_file_mounts_contents_hash if not self.provider: self.provider = _get_node_provider( self.config["provider"], self.config["cluster_name"] ) # If using the LocalNodeProvider, make sure the head node is marked # non-terminated. if isinstance(self.provider, LocalNodeProvider): record_local_head_state_if_needed(self.provider) self.available_node_types = self.config["available_node_types"] upscaling_speed = self.config.get("upscaling_speed") aggressive = self.config.get("autoscaling_mode") == "aggressive" target_utilization_fraction = self.config.get("target_utilization_fraction") if upscaling_speed: upscaling_speed = float(upscaling_speed) # TODO(ameer): consider adding (if users ask) an option of # initial_upscaling_num_workers. elif aggressive: upscaling_speed = 99999 logger.warning( "Legacy aggressive autoscaling mode " "detected. Replacing it by setting upscaling_speed to " "99999." ) elif target_utilization_fraction: upscaling_speed = 1 / max(target_utilization_fraction, 0.001) - 1 logger.warning( "Legacy target_utilization_fraction config " "detected. Replacing it by setting upscaling_speed to " + "1 / target_utilization_fraction - 1." ) else: upscaling_speed = 1.0 if self.resource_demand_scheduler: # The node types are autofilled internally for legacy yamls, # overwriting the class will remove the inferred node resources # for legacy yamls. self.resource_demand_scheduler.reset_config( self.provider, self.available_node_types, self.config["max_workers"], self.config["head_node_type"], upscaling_speed, ) else: self.resource_demand_scheduler = ResourceDemandScheduler( self.provider, self.available_node_types, self.config["max_workers"], self.config["head_node_type"], upscaling_speed, ) except Exception as e: self.prom_metrics.reset_exceptions.inc() if errors_fatal: raise e else: logger.exception("StandardAutoscaler: Error parsing config.") def launch_config_ok(self, node_id): if self.disable_launch_config_check: return True node_tags = self.provider.node_tags(node_id) tag_launch_conf = node_tags.get(TAG_RAY_LAUNCH_CONFIG) node_type = node_tags.get(TAG_RAY_USER_NODE_TYPE) if node_type not in self.available_node_types: # The node type has been deleted from the cluster config. # Don't keep the node. return False # The `worker_nodes` field is deprecated in favor of per-node-type # node_configs. We allow it for backwards-compatibility. launch_config = copy.deepcopy(self.config.get("worker_nodes", {})) if node_type: launch_config.update( self.config["available_node_types"][node_type]["node_config"] ) calculated_launch_hash = hash_launch_conf(launch_config, self.config["auth"]) if calculated_launch_hash != tag_launch_conf: return False return True def files_up_to_date(self, node_id): node_tags = self.provider.node_tags(node_id) applied_config_hash = node_tags.get(TAG_RAY_RUNTIME_CONFIG) applied_file_mounts_contents_hash = node_tags.get(TAG_RAY_FILE_MOUNTS_CONTENTS) if applied_config_hash != self.runtime_hash or ( self.file_mounts_contents_hash is not None and self.file_mounts_contents_hash != applied_file_mounts_contents_hash ): logger.info( "StandardAutoscaler: " "{}: Runtime state is ({},{}), want ({},{})".format( node_id, applied_config_hash, applied_file_mounts_contents_hash, self.runtime_hash, self.file_mounts_contents_hash, ) ) return False return True def heartbeat_on_time(self, node_id: NodeID, now: float) -> bool: """Determine whether we've received a heartbeat from a node within the last AUTOSCALER_HEARTBEAT_TIMEOUT_S seconds. """ # For type checking, assert that this object has been instantitiated. assert self.provider key = self.provider.internal_ip(node_id) if key in self.load_metrics.last_heartbeat_time_by_ip: last_heartbeat_time = self.load_metrics.last_heartbeat_time_by_ip[key] delta = now - last_heartbeat_time if delta < AUTOSCALER_HEARTBEAT_TIMEOUT_S: return True return False def terminate_unhealthy_nodes(self, now: float): """Terminated nodes for which we haven't received a heartbeat on time. These nodes are subsequently terminated. """ # For type checking, assert that these objects have been instantitiated. assert self.provider assert self.non_terminated_nodes for node_id in self.non_terminated_nodes.worker_ids: node_status = self.provider.node_tags(node_id)[TAG_RAY_NODE_STATUS] # We're not responsible for taking down # nodes with pending or failed status: if not node_status == STATUS_UP_TO_DATE: continue # This node is up-to-date. If it hasn't had the chance to produce # a heartbeat, fake the heartbeat now (see logic for completed node # updaters). ip = self.provider.internal_ip(node_id) if ip not in self.load_metrics.last_heartbeat_time_by_ip: self.load_metrics.mark_active(ip) # Heartbeat indicates node is healthy: if self.heartbeat_on_time(node_id, now): continue self.schedule_node_termination( node_id, "lost contact with raylet", logger.warning ) self.terminate_scheduled_nodes() def attempt_to_recover_unhealthy_nodes(self, now): for node_id in self.non_terminated_nodes.worker_ids: self.recover_if_needed(node_id, now) def recover_if_needed(self, node_id, now): if not self.can_update(node_id): return if self.heartbeat_on_time(node_id, now): return logger.warning( "StandardAutoscaler: " "{}: No recent heartbeat, " "restarting Ray to recover...".format(node_id) ) self.event_summarizer.add( "Restarting {} nodes of type " + self._get_node_type(node_id) + " (lost contact with raylet).", quantity=1, aggregate=operator.add, ) head_node_ip = self.provider.internal_ip(self.non_terminated_nodes.head_id) updater = NodeUpdaterThread( node_id=node_id, provider_config=self.config["provider"], provider=self.provider, auth_config=self.config["auth"], cluster_name=self.config["cluster_name"], file_mounts={}, initialization_commands=[], setup_commands=[], ray_start_commands=with_head_node_ip( self.config["worker_start_ray_commands"], head_node_ip ), runtime_hash=self.runtime_hash, file_mounts_contents_hash=self.file_mounts_contents_hash, process_runner=self.process_runner, use_internal_ip=True, is_head_node=False, docker_config=self._get_node_specific_docker_config(node_id), node_resources=self._node_resources(node_id), node_labels=self._node_labels(node_id), for_recovery=True, ) updater.start() self.updaters[node_id] = updater def _get_node_type(self, node_id: str) -> str: # For type checking, assert that this object has been instantitiated. assert self.provider node_tags = self.provider.node_tags(node_id) if TAG_RAY_USER_NODE_TYPE in node_tags: return node_tags[TAG_RAY_USER_NODE_TYPE] else: return "unknown_node_type" def _get_node_type_specific_fields(self, node_id: str, fields_key: str) -> Any: # For type checking, assert that this object has been instantitiated. assert self.provider fields = self.config[fields_key] node_tags = self.provider.node_tags(node_id) if TAG_RAY_USER_NODE_TYPE in node_tags: node_type = node_tags[TAG_RAY_USER_NODE_TYPE] if node_type not in self.available_node_types: raise ValueError(f"Unknown node type tag: {node_type}.") node_specific_config = self.available_node_types[node_type] if fields_key in node_specific_config: fields = node_specific_config[fields_key] return fields def _get_node_specific_docker_config(self, node_id): if "docker" not in self.config: return {} docker_config = copy.deepcopy(self.config.get("docker", {})) node_specific_docker = self._get_node_type_specific_fields(node_id, "docker") docker_config.update(node_specific_docker) return docker_config def should_update(self, node_id): if not self.can_update(node_id): return UpdateInstructions(None, None, None, None) # no update status = self.provider.node_tags(node_id).get(TAG_RAY_NODE_STATUS) if status == STATUS_UP_TO_DATE and self.files_up_to_date(node_id): return UpdateInstructions(None, None, None, None) # no update successful_updated = self.num_successful_updates.get(node_id, 0) > 0 if successful_updated and self.config.get("restart_only", False): setup_commands = [] ray_start_commands = self.config["worker_start_ray_commands"] elif successful_updated and self.config.get("no_restart", False): setup_commands = self._get_node_type_specific_fields( node_id, "worker_setup_commands" ) ray_start_commands = [] else: setup_commands = self._get_node_type_specific_fields( node_id, "worker_setup_commands" ) ray_start_commands = self.config["worker_start_ray_commands"] docker_config = self._get_node_specific_docker_config(node_id) return UpdateInstructions( node_id=node_id, setup_commands=setup_commands, ray_start_commands=ray_start_commands, docker_config=docker_config, ) def spawn_updater( self, node_id, setup_commands, ray_start_commands, node_resources, node_labels, docker_config, ): logger.info( f"Creating new (spawn_updater) updater thread for node" f" {node_id}." ) ip = self.provider.internal_ip(node_id) node_type = self._get_node_type(node_id) self.node_tracker.track(node_id, ip, node_type) head_node_ip = self.provider.internal_ip(self.non_terminated_nodes.head_id) updater = NodeUpdaterThread( node_id=node_id, provider_config=self.config["provider"], provider=self.provider, auth_config=self.config["auth"], cluster_name=self.config["cluster_name"], file_mounts=self.config["file_mounts"], initialization_commands=with_head_node_ip( self._get_node_type_specific_fields(node_id, "initialization_commands"), head_node_ip, ), setup_commands=with_head_node_ip(setup_commands, head_node_ip), ray_start_commands=with_head_node_ip(ray_start_commands, head_node_ip), runtime_hash=self.runtime_hash, file_mounts_contents_hash=self.file_mounts_contents_hash, is_head_node=False, cluster_synced_files=self.config["cluster_synced_files"], rsync_options={ "rsync_exclude": self.config.get("rsync_exclude"), "rsync_filter": self.config.get("rsync_filter"), }, process_runner=self.process_runner, use_internal_ip=True, docker_config=docker_config, node_resources=node_resources, node_labels=node_labels, ) updater.start() self.updaters[node_id] = updater def can_update(self, node_id): if self.disable_node_updaters: return False if node_id in self.updaters: return False if not self.launch_config_ok(node_id): return False if self.num_failed_updates.get(node_id, 0) > 0: # TODO(ekl) retry? return False logger.debug( f"{node_id} is not being updated and " "passes config check (can_update=True)." ) return True def launch_new_node(self, count: int, node_type: str) -> None: logger.info("StandardAutoscaler: Queue {} new nodes for launch".format(count)) self.pending_launches.inc(node_type, count) config = copy.deepcopy(self.config) if self.foreground_node_launch: assert self.foreground_node_launcher is not None # Launch in the main thread and block. self.foreground_node_launcher.launch_node(config, count, node_type) else: assert self.launch_queue is not None # Split into individual launch requests of the max batch size. while count > 0: # Enqueue launch data for the background NodeUpdater threads. self.launch_queue.put( (config, min(count, self.max_launch_batch), node_type) ) count -= self.max_launch_batch def kill_workers(self): logger.error("StandardAutoscaler: kill_workers triggered") nodes = self.workers() if nodes: self.provider.terminate_nodes(nodes) for node in nodes: self.node_tracker.untrack(node) self.prom_metrics.stopped_nodes.inc() logger.error("StandardAutoscaler: terminated {} node(s)".format(len(nodes))) def summary(self) -> Optional[AutoscalerSummary]: """Summarizes the active, pending, and failed node launches. An active node is a node whose raylet is actively reporting heartbeats. A pending node is non-active node whose node tag is uninitialized, waiting for ssh, syncing files, or setting up. If a node is not pending or active, it is failed. Returns: AutoscalerSummary: The summary. """ # For type checking, assert that this object has been instantitiated. assert self.provider if not self.non_terminated_nodes: return None active_nodes: Dict[NodeType, int] = Counter() pending_nodes = [] failed_nodes = [] non_failed = set() node_type_mapping = {} now = time.time() for node_id in self.non_terminated_nodes.all_node_ids: ip = self.provider.internal_ip(node_id) node_tags = self.provider.node_tags(node_id) if not all( tag in node_tags for tag in ( TAG_RAY_NODE_KIND, TAG_RAY_USER_NODE_TYPE, TAG_RAY_NODE_STATUS, ) ): # In some node providers, creation of a node and tags is not # atomic, so just skip it. continue if node_tags[TAG_RAY_NODE_KIND] == NODE_KIND_UNMANAGED: continue node_type = node_tags[TAG_RAY_USER_NODE_TYPE] node_type_mapping[ip] = node_type is_active = self.heartbeat_on_time(node_id, now) if is_active: active_nodes[node_type] += 1 non_failed.add(node_id) else: status = node_tags[TAG_RAY_NODE_STATUS] completed_states = [STATUS_UP_TO_DATE, STATUS_UPDATE_FAILED] is_pending = status not in completed_states if is_pending: pending_nodes.append((node_id, ip, node_type, status)) non_failed.add(node_id) failed_nodes = self.node_tracker.get_all_failed_node_info(non_failed) # The concurrent counter leaves some 0 counts in, so we need to # manually filter those out. pending_launches = {} for node_type, count in self.pending_launches.breakdown().items(): if count: pending_launches[node_type] = count pending_resources = {} for node_resources in self.resource_demand_scheduler.calculate_node_resources( nodes=[node_id for node_id, _, _, _ in pending_nodes], pending_nodes=pending_launches, # We don't fill this field out because we're intentionally only # passing pending nodes (which aren't tracked by load metrics # anyways). unused_resources_by_ip={}, )[0]: for key, value in node_resources.items(): pending_resources[key] = value + pending_resources.get(key, 0) return AutoscalerSummary( # Convert active_nodes from counter to dict for later serialization active_nodes=dict(active_nodes), idle_nodes=None, pending_nodes=[ (ip, node_type, status) for _, ip, node_type, status in pending_nodes ], pending_launches=pending_launches, failed_nodes=failed_nodes, node_availability_summary=self.node_provider_availability_tracker.summary(), pending_resources=pending_resources, node_type_mapping=node_type_mapping, legacy=True, ) def info_string(self): lm_summary = self.load_metrics.summary() autoscaler_summary = self.summary() assert autoscaler_summary return "\n" + format_info_string(lm_summary, autoscaler_summary)