def reset(self, errors_fatal=False):
sync_continuously = False
if hasattr(self, "config"):
sync_continuously = self.config.get(
"file_mounts_sync_continuously", False)
try:
with open(self.config_path) as f:
new_config = yaml.safe_load(f.read())
if new_config != getattr(self, "config", None):
try:
validate_config(new_config)
except Exception as e:
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)
(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"])
# Check whether we can enable the resource demand scheduler.
if "available_node_types" in self.config:
self.available_node_types = self.config["available_node_types"]
self.resource_demand_scheduler = ResourceDemandScheduler(
self.provider, self.available_node_types,
self.config["max_workers"])
else:
self.available_node_types = None
self.resource_demand_scheduler = None
except Exception as e:
if errors_fatal:
raise e
else:
logger.exception("StandardAutoscaler: "
"Error parsing config.")
def reset(self, errors_fatal=False):
sync_continuously = False
if hasattr(self, "config"):
sync_continuously = self.config.get(
"file_mounts_sync_continuously", False)
try:
with open(self.config_path) as f:
new_config = yaml.safe_load(f.read())
validate_config(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"])
# Check whether we can enable the resource demand scheduler.
if "available_node_types" in self.config:
self.available_node_types = self.config["available_node_types"]
self.resource_demand_scheduler = ResourceDemandScheduler(
self.provider, self.available_node_types,
self.config["max_workers"])
else:
self.available_node_types = None
self.resource_demand_scheduler = None
except Exception as e:
if errors_fatal:
raise e
else:
logger.exception("StandardAutoscaler: "
"Error parsing config.")
def reset(self, errors_fatal=False):
sync_continuously = False
if hasattr(self, "config"):
sync_continuously = self.config.get(
"file_mounts_sync_continuously", False)
try:
with open(self.config_path) as f:
new_config = yaml.safe_load(f.read())
if new_config != getattr(self, "config", None):
try:
validate_config(new_config)
except Exception as e:
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)
(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"])
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:
if errors_fatal:
raise e
else:
logger.exception("StandardAutoscaler: "
"Error parsing config.")
def get_or_create_head_node(config,
config_file,
no_restart,
restart_only,
yes,
override_cluster_name,
_provider=None,
_runner=subprocess):
"""Create the cluster head node, which in turn creates the workers."""
provider = (_provider or _get_node_provider(config["provider"],
config["cluster_name"]))
config = copy.deepcopy(config)
config_file = os.path.abspath(config_file)
try:
head_node_tags = {
TAG_RAY_NODE_KIND: NODE_KIND_HEAD,
}
nodes = provider.non_terminated_nodes(head_node_tags)
if len(nodes) > 0:
head_node = nodes[0]
else:
head_node = None
if not head_node:
cli_logger.confirm(yes, "No head node found. "
"Launching a new cluster.",
_abort=True)
cli_logger.old_confirm("This will create a new cluster", yes)
elif not no_restart:
cli_logger.old_confirm("This will restart cluster services", yes)
if head_node:
if restart_only:
cli_logger.confirm(
yes, "Updating cluster configuration and "
"restarting the cluster Ray runtime. "
"Setup commands will not be run due to `{}`.\n",
cf.bold("--restart-only"),
_abort=True)
elif no_restart:
cli_logger.print(
"Cluster Ray runtime will not be restarted due "
"to `{}`.", cf.bold("--no-restart"))
cli_logger.confirm(yes, "Updating cluster configuration and "
"running setup commands.",
_abort=True)
else:
cli_logger.print(
"Updating cluster configuration and running full setup.")
cli_logger.confirm(
yes,
cf.bold("Cluster Ray runtime will be restarted."),
_abort=True)
cli_logger.newline()
# TODO(ekl) this logic is duplicated in node_launcher.py (keep in sync)
head_node_config = copy.deepcopy(config["head_node"])
if "head_node_type" in config:
head_node_tags[TAG_RAY_USER_NODE_TYPE] = config["head_node_type"]
head_node_config.update(config["available_node_types"][
config["head_node_type"]]["node_config"])
launch_hash = hash_launch_conf(head_node_config, config["auth"])
if head_node is None or provider.node_tags(head_node).get(
TAG_RAY_LAUNCH_CONFIG) != launch_hash:
with cli_logger.group("Acquiring an up-to-date head node"):
if head_node is not None:
cli_logger.print(
"Currently running head node is out-of-date with "
"cluster configuration")
cli_logger.print(
"hash is {}, expected {}",
cf.bold(
provider.node_tags(head_node).get(
TAG_RAY_LAUNCH_CONFIG)), cf.bold(launch_hash))
cli_logger.confirm(yes, "Relaunching it.", _abort=True)
cli_logger.old_confirm(
"Head node config out-of-date. It will be terminated",
yes)
cli_logger.old_info(
logger, "get_or_create_head_node: "
"Shutting down outdated head node {}", head_node)
provider.terminate_node(head_node)
cli_logger.print("Terminated head node {}", head_node)
cli_logger.old_info(
logger,
"get_or_create_head_node: Launching new head node...")
head_node_tags[TAG_RAY_LAUNCH_CONFIG] = launch_hash
head_node_tags[TAG_RAY_NODE_NAME] = "ray-{}-head".format(
config["cluster_name"])
provider.create_node(head_node_config, head_node_tags, 1)
cli_logger.print("Launched a new head node")
start = time.time()
head_node = None
with cli_logger.timed("Fetching the new head node"):
while True:
if time.time() - start > 50:
cli_logger.abort(
"Head node fetch timed out.") # todo: msg
raise RuntimeError("Failed to create head node.")
nodes = provider.non_terminated_nodes(head_node_tags)
if len(nodes) == 1:
head_node = nodes[0]
break
time.sleep(POLL_INTERVAL)
cli_logger.newline()
with cli_logger.group(
"Setting up head node",
_numbered=("<>", 1, 1),
# cf.bold(provider.node_tags(head_node)[TAG_RAY_NODE_NAME]),
_tags=dict()): # add id, ARN to tags?
# TODO(ekl) right now we always update the head node even if the
# hash matches.
# We could prompt the user for what they want to do here.
# No need to pass in cluster_sync_files because we use this
# hash to set up the head node
(runtime_hash, file_mounts_contents_hash) = hash_runtime_conf(
config["file_mounts"], None, config)
cli_logger.old_info(
logger,
"get_or_create_head_node: Updating files on head node...")
# Rewrite the auth config so that the head
# node can update the workers
remote_config = copy.deepcopy(config)
# drop proxy options if they exist, otherwise
# head node won't be able to connect to workers
remote_config["auth"].pop("ssh_proxy_command", None)
if "ssh_private_key" in config["auth"]:
remote_key_path = "~/ray_bootstrap_key.pem"
remote_config["auth"]["ssh_private_key"] = remote_key_path
# Adjust for new file locations
new_mounts = {}
for remote_path in config["file_mounts"]:
new_mounts[remote_path] = remote_path
remote_config["file_mounts"] = new_mounts
remote_config["no_restart"] = no_restart
remote_config = provider.prepare_for_head_node(remote_config)
# Now inject the rewritten config and SSH key into the head node
remote_config_file = tempfile.NamedTemporaryFile(
"w", prefix="ray-bootstrap-")
remote_config_file.write(json.dumps(remote_config))
remote_config_file.flush()
config["file_mounts"].update(
{"~/ray_bootstrap_config.yaml": remote_config_file.name})
if "ssh_private_key" in config["auth"]:
config["file_mounts"].update({
remote_key_path:
config["auth"]["ssh_private_key"],
})
cli_logger.print("Prepared bootstrap config")
if restart_only:
setup_commands = []
ray_start_commands = config["head_start_ray_commands"]
elif no_restart:
setup_commands = config["head_setup_commands"]
ray_start_commands = []
else:
setup_commands = config["head_setup_commands"]
ray_start_commands = config["head_start_ray_commands"]
if not no_restart:
warn_about_bad_start_command(ray_start_commands)
updater = NodeUpdaterThread(
node_id=head_node,
provider_config=config["provider"],
provider=provider,
auth_config=config["auth"],
cluster_name=config["cluster_name"],
file_mounts=config["file_mounts"],
initialization_commands=config["initialization_commands"],
setup_commands=setup_commands,
ray_start_commands=ray_start_commands,
process_runner=_runner,
runtime_hash=runtime_hash,
file_mounts_contents_hash=file_mounts_contents_hash,
is_head_node=True,
docker_config=config.get("docker"))
updater.start()
updater.join()
# Refresh the node cache so we see the external ip if available
provider.non_terminated_nodes(head_node_tags)
if config.get("provider", {}).get("use_internal_ips",
False) is True:
head_node_ip = provider.internal_ip(head_node)
else:
head_node_ip = provider.external_ip(head_node)
if updater.exitcode != 0:
# todo: this does not follow the mockup and is not good enough
cli_logger.abort("Failed to setup head node.")
cli_logger.old_error(
logger, "get_or_create_head_node: "
"Updating {} failed", head_node_ip)
sys.exit(1)
cli_logger.old_info(
logger, "get_or_create_head_node: "
"Head node up-to-date, IP address is: {}", head_node_ip)
monitor_str = "tail -n 100 -f /tmp/ray/session_latest/logs/monitor*"
if override_cluster_name:
modifiers = " --cluster-name={}".format(
quote(override_cluster_name))
else:
modifiers = ""
if cli_logger.old_style:
print("To monitor autoscaling activity, you can run:\n\n"
" ray exec {} {}{}\n".format(config_file,
quote(monitor_str), modifiers))
print("To open a console on the cluster:\n\n"
" ray attach {}{}\n".format(config_file, modifiers))
print("To get a remote shell to the cluster manually, run:\n\n"
" {}\n".format(
updater.cmd_runner.remote_shell_command_str()))
cli_logger.newline()
with cli_logger.group("Useful commands"):
cli_logger.print("Monitor autoscaling with")
cli_logger.print(cf.bold(" ray exec {}{} {}"), config_file,
modifiers, quote(monitor_str))
cli_logger.print("Connect to a terminal on the cluster head:")
cli_logger.print(cf.bold(" ray attach {}{}"), config_file,
modifiers)
remote_shell_str = updater.cmd_runner.remote_shell_command_str()
cli_logger.print("Get a remote shell to the cluster manually:")
cli_logger.print(" {}", remote_shell_str.strip())
finally:
provider.cleanup()
def get_or_create_head_node(config: Dict[str, Any],
printable_config_file: str,
no_restart: bool,
restart_only: bool,
yes: bool,
override_cluster_name: Optional[str],
no_monitor_on_head: bool = False,
_provider: Optional[NodeProvider] = None,
_runner: ModuleType = subprocess) -> None:
"""Create the cluster head node, which in turn creates the workers."""
global_event_system.execute_callback(
CreateClusterEvent.cluster_booting_started)
provider = (_provider or _get_node_provider(config["provider"],
config["cluster_name"]))
config = copy.deepcopy(config)
head_node_tags = {
TAG_RAY_NODE_KIND: NODE_KIND_HEAD,
}
nodes = provider.non_terminated_nodes(head_node_tags)
if len(nodes) > 0:
head_node = nodes[0]
else:
head_node = None
if not head_node:
cli_logger.confirm(
yes,
"No head node found. "
"Launching a new cluster.",
_abort=True)
if head_node:
if restart_only:
cli_logger.confirm(
yes,
"Updating cluster configuration and "
"restarting the cluster Ray runtime. "
"Setup commands will not be run due to `{}`.\n",
cf.bold("--restart-only"),
_abort=True)
elif no_restart:
cli_logger.print(
"Cluster Ray runtime will not be restarted due "
"to `{}`.", cf.bold("--no-restart"))
cli_logger.confirm(
yes,
"Updating cluster configuration and "
"running setup commands.",
_abort=True)
else:
cli_logger.print(
"Updating cluster configuration and running full setup.")
cli_logger.confirm(
yes,
cf.bold("Cluster Ray runtime will be restarted."),
_abort=True)
cli_logger.newline()
# TODO(ekl) this logic is duplicated in node_launcher.py (keep in sync)
head_node_config = copy.deepcopy(config["head_node"])
head_node_resources = None
if "head_node_type" in config:
head_node_type = config["head_node_type"]
head_node_tags[TAG_RAY_USER_NODE_TYPE] = head_node_type
head_config = config["available_node_types"][head_node_type]
head_node_config.update(head_config["node_config"])
# Not necessary to keep in sync with node_launcher.py
# Keep in sync with autoscaler.py _node_resources
head_node_resources = head_config.get("resources")
launch_hash = hash_launch_conf(head_node_config, config["auth"])
if head_node is None or provider.node_tags(head_node).get(
TAG_RAY_LAUNCH_CONFIG) != launch_hash:
with cli_logger.group("Acquiring an up-to-date head node"):
global_event_system.execute_callback(
CreateClusterEvent.acquiring_new_head_node)
if head_node is not None:
cli_logger.print(
"Currently running head node is out-of-date with "
"cluster configuration")
cli_logger.print(
"hash is {}, expected {}",
cf.bold(
provider.node_tags(head_node)
.get(TAG_RAY_LAUNCH_CONFIG)), cf.bold(launch_hash))
cli_logger.confirm(yes, "Relaunching it.", _abort=True)
provider.terminate_node(head_node)
cli_logger.print("Terminated head node {}", head_node)
head_node_tags[TAG_RAY_LAUNCH_CONFIG] = launch_hash
head_node_tags[TAG_RAY_NODE_NAME] = "ray-{}-head".format(
config["cluster_name"])
head_node_tags[TAG_RAY_NODE_STATUS] = STATUS_UNINITIALIZED
provider.create_node(head_node_config, head_node_tags, 1)
cli_logger.print("Launched a new head node")
start = time.time()
head_node = None
with cli_logger.group("Fetching the new head node"):
while True:
if time.time() - start > 50:
cli_logger.abort(
"Head node fetch timed out.") # todo: msg
raise RuntimeError("Failed to create head node.")
nodes = provider.non_terminated_nodes(head_node_tags)
if len(nodes) == 1:
head_node = nodes[0]
break
time.sleep(POLL_INTERVAL)
cli_logger.newline()
global_event_system.execute_callback(CreateClusterEvent.head_node_acquired)
with cli_logger.group(
"Setting up head node",
_numbered=("<>", 1, 1),
# cf.bold(provider.node_tags(head_node)[TAG_RAY_NODE_NAME]),
_tags=dict()): # add id, ARN to tags?
# TODO(ekl) right now we always update the head node even if the
# hash matches.
# We could prompt the user for what they want to do here.
# No need to pass in cluster_sync_files because we use this
# hash to set up the head node
(runtime_hash, file_mounts_contents_hash) = hash_runtime_conf(
config["file_mounts"], None, config)
if not no_monitor_on_head:
# Return remote_config_file to avoid prematurely closing it.
config, remote_config_file = _set_up_config_for_head_node(
config, provider, no_restart)
cli_logger.print("Prepared bootstrap config")
if restart_only:
# Docker may re-launch nodes, requiring setup
# commands to be rerun.
if config.get("docker", {}).get("container_name"):
setup_commands = config["head_setup_commands"]
else:
setup_commands = []
ray_start_commands = config["head_start_ray_commands"]
elif no_restart:
setup_commands = config["head_setup_commands"]
ray_start_commands = []
else:
setup_commands = config["head_setup_commands"]
ray_start_commands = config["head_start_ray_commands"]
if not no_restart:
warn_about_bad_start_command(ray_start_commands,
no_monitor_on_head)
updater = NodeUpdaterThread(
node_id=head_node,
provider_config=config["provider"],
provider=provider,
auth_config=config["auth"],
cluster_name=config["cluster_name"],
file_mounts=config["file_mounts"],
initialization_commands=config["initialization_commands"],
setup_commands=setup_commands,
ray_start_commands=ray_start_commands,
process_runner=_runner,
runtime_hash=runtime_hash,
file_mounts_contents_hash=file_mounts_contents_hash,
is_head_node=True,
node_resources=head_node_resources,
rsync_options={
"rsync_exclude": config.get("rsync_exclude"),
"rsync_filter": config.get("rsync_filter")
},
docker_config=config.get("docker"),
restart_only=restart_only)
updater.start()
updater.join()
# Refresh the node cache so we see the external ip if available
provider.non_terminated_nodes(head_node_tags)
if updater.exitcode != 0:
# todo: this does not follow the mockup and is not good enough
cli_logger.abort("Failed to setup head node.")
sys.exit(1)
global_event_system.execute_callback(
CreateClusterEvent.cluster_booting_completed, {
"head_node_id": head_node,
})
monitor_str = "tail -n 100 -f /tmp/ray/session_latest/logs/monitor*"
if override_cluster_name:
modifiers = " --cluster-name={}".format(quote(override_cluster_name))
else:
modifiers = ""
cli_logger.newline()
with cli_logger.group("Useful commands"):
printable_config_file = os.path.abspath(printable_config_file)
cli_logger.print("Monitor autoscaling with")
cli_logger.print(
cf.bold(" ray exec {}{} {}"), printable_config_file, modifiers,
quote(monitor_str))
cli_logger.print("Connect to a terminal on the cluster head:")
cli_logger.print(
cf.bold(" ray attach {}{}"), printable_config_file, modifiers)
remote_shell_str = updater.cmd_runner.remote_shell_command_str()
cli_logger.print("Get a remote shell to the cluster manually:")
cli_logger.print(" {}", remote_shell_str.strip())