def spawn_updater(self, node_id, init_commands, ray_start_commands,
node_resources, docker_config):
logger.info(f"Creating new (spawn_updater) updater thread for node"
f" {node_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")),
setup_commands=with_head_node_ip(init_commands),
ray_start_commands=with_head_node_ip(ray_start_commands),
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"],
process_runner=self.process_runner,
use_internal_ip=True,
docker_config=docker_config,
node_resources=node_resources)
updater.start()
self.updaters[node_id] = updater
def recover_if_needed(self, node_id, now):
if not self.can_update(node_id):
return
key = self.provider.internal_ip(node_id)
if key not in self.load_metrics.last_heartbeat_time_by_ip:
self.load_metrics.last_heartbeat_time_by_ip[key] = now
last_heartbeat_time = self.load_metrics.last_heartbeat_time_by_ip[key]
delta = now - last_heartbeat_time
if delta < AUTOSCALER_HEARTBEAT_TIMEOUT_S:
return
logger.warning("StandardAutoscaler: "
"{}: No heartbeat in {}s, "
"restarting Ray to recover...".format(node_id, delta))
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"]),
runtime_hash=self.runtime_hash,
process_runner=self.process_runner,
use_internal_ip=True,
docker_config=self.config.get("docker"))
updater.start()
self.updaters[node_id] = updater
def recover_if_needed(self, node_id, now):
if not self.can_update(node_id):
return
key = self.provider.internal_ip(node_id)
if key not in self.load_metrics.last_heartbeat_time_by_ip:
self.load_metrics.last_heartbeat_time_by_ip[key] = now
last_heartbeat_time = self.load_metrics.last_heartbeat_time_by_ip[key]
delta = now - last_heartbeat_time
if delta < AUTOSCALER_HEARTBEAT_TIMEOUT_S:
return
logger.warning("StandardAutoscaler: "
"{}: No heartbeat in {}s, "
"restarting Ray to recover...".format(node_id, delta))
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=with_head_node_ip(
self.config["worker_start_ray_commands"]),
runtime_hash=self.runtime_hash,
process_runner=self.process_runner,
use_internal_ip=True)
updater.start()
self.updaters[node_id] = updater
def spawn_updater(self, node_id, init_commands):
updater = NodeUpdaterThread(node_id,
self.config["provider"],
self.provider,
self.config["auth"],
self.config["cluster_name"],
self.config["file_mounts"],
with_head_node_ip(init_commands),
self.runtime_hash,
process_runner=self.process_runner,
use_internal_ip=True)
updater.start()
self.updaters[node_id] = updater
def rsync(config_file, source, target, override_cluster_name, down):
"""Rsyncs files.
Arguments:
config_file: path to the cluster yaml
source: source dir
target: target dir
override_cluster_name: set the name of the cluster
down: whether we're syncing remote -> local
"""
assert bool(source) == bool(target), (
"Must either provide both or neither source and target.")
config = yaml.safe_load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config)
head_node = _get_head_node(config,
config_file,
override_cluster_name,
create_if_needed=False)
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
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=[],
setup_commands=[],
ray_start_commands=[],
runtime_hash="",
)
if down:
rsync = updater.rsync_down
else:
rsync = updater.rsync_up
if source and target:
rsync(source, target)
else:
updater.sync_file_mounts(rsync)
finally:
provider.cleanup()
def kill_node(config_file, yes, override_cluster_name):
"""Kills a random Raylet worker."""
config = yaml.load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config)
confirm("This will kill a node in your cluster", yes)
provider = get_node_provider(config["provider"], config["cluster_name"])
nodes = provider.nodes({TAG_RAY_NODE_TYPE: "worker"})
node = random.choice(nodes)
logger.info("kill_node: Terminating worker {}".format(node))
updater = NodeUpdaterThread(node, config["provider"], provider,
config["auth"], config["cluster_name"],
config["file_mounts"], [], "")
_exec(updater, "ray stop", False, False)
time.sleep(5)
iip = provider.internal_ip(node)
if iip:
return iip
return provider.external_ip(node)
def spawn_updater(self, node_id, init_commands):
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.config["initialization_commands"]),
setup_commands=with_head_node_ip(init_commands),
runtime_hash=self.runtime_hash,
process_runner=self.process_runner,
use_internal_ip=True)
updater.start()
self.updaters[node_id] = updater
def run_docker_stop(node, container_name):
try:
updater = NodeUpdaterThread(
node_id=node,
provider_config=config["provider"],
provider=provider,
auth_config=config["auth"],
cluster_name=config["cluster_name"],
file_mounts=config["file_mounts"],
initialization_commands=[],
setup_commands=[],
ray_start_commands=[],
runtime_hash="",
file_mounts_contents_hash="",
is_head_node=False,
docker_config=config.get("docker"))
_exec(
updater,
f"docker stop {container_name}",
False,
False,
run_env="host")
except Exception:
cli_logger.warning(f"Docker stop failed on {node}")
cli_logger.old_warning(logger, f"Docker stop failed on {node}")
def spawn_updater(self, node_id, init_commands):
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.config["initialization_commands"]),
setup_commands=with_head_node_ip(init_commands),
runtime_hash=self.runtime_hash,
process_runner=self.process_runner,
use_internal_ip=True)
updater.start()
self.updaters[node_id] = updater
def spawn_updater(self, node_id, init_commands, ray_start_commands,
node_resources):
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.config["initialization_commands"]),
setup_commands=with_head_node_ip(init_commands),
ray_start_commands=with_head_node_ip(ray_start_commands),
runtime_hash=self.runtime_hash,
file_mounts_contents_hash=self.file_mounts_contents_hash,
cluster_synced_files=self.config["cluster_synced_files"],
process_runner=self.process_runner,
use_internal_ip=True,
docker_config=self.config.get("docker"),
node_resources=node_resources)
updater.start()
self.updaters[node_id] = updater
def kill_node(config_file, yes, hard, override_cluster_name):
"""Kills a random Raylet worker."""
config = yaml.safe_load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config)
cli_logger.confirm(yes, "A random node will be killed.")
cli_logger.old_confirm("This will kill a node in your cluster", yes)
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
nodes = provider.non_terminated_nodes({
TAG_RAY_NODE_KIND: NODE_KIND_WORKER
})
node = random.choice(nodes)
cli_logger.print("Shutdown " + cf.bold("{}"), node)
cli_logger.old_info(logger, "kill_node: Shutdown worker {}", node)
if hard:
provider.terminate_node(node)
else:
updater = NodeUpdaterThread(
node_id=node,
provider_config=config["provider"],
provider=provider,
auth_config=config["auth"],
cluster_name=config["cluster_name"],
file_mounts=config["file_mounts"],
initialization_commands=[],
setup_commands=[],
ray_start_commands=[],
runtime_hash="",
file_mounts_contents_hash="",
is_head_node=False,
docker_config=config.get("docker"))
_exec(updater, "ray stop", False, False)
time.sleep(5)
if config.get("provider", {}).get("use_internal_ips", False) is True:
node_ip = provider.internal_ip(node)
else:
node_ip = provider.external_ip(node)
finally:
provider.cleanup()
return node_ip
def kill_node(config_file, yes, hard, override_cluster_name):
"""Kills a random Raylet worker."""
config = yaml.safe_load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config)
confirm("This will kill a node in your cluster", yes)
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
nodes = provider.non_terminated_nodes({TAG_RAY_NODE_TYPE: "worker"})
node = random.choice(nodes)
logger.info("kill_node: Terminating worker {}".format(node))
if hard:
provider.terminate_node(node)
else:
updater = NodeUpdaterThread(node_id=node,
provider_config=config["provider"],
provider=provider,
auth_config=config["auth"],
cluster_name=config["cluster_name"],
file_mounts=config["file_mounts"],
initialization_commands=[],
setup_commands=[],
runtime_hash="")
_exec(updater, "ray stop", False, False)
time.sleep(5)
if config.get("provider", {}).get("use_internal_ips", False) is True:
node_ip = provider.internal_ip(node)
else:
node_ip = provider.external_ip(node)
finally:
provider.cleanup()
return node_ip
def rsync(config_file, source, target, override_cluster_name, down):
"""Rsyncs files.
Arguments:
config_file: path to the cluster yaml
source: source dir
target: target dir
override_cluster_name: set the name of the cluster
down: whether we're syncing remote -> local
"""
config = yaml.load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config)
head_node = _get_head_node(config,
config_file,
override_cluster_name,
create_if_needed=False)
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
updater = NodeUpdaterThread(
head_node,
config["provider"],
provider,
config["auth"],
config["cluster_name"],
config["file_mounts"],
[],
"",
)
if down:
rsync = updater.rsync_down
else:
rsync = updater.rsync_up
rsync(source, target, check_error=False)
finally:
provider.cleanup()
def rsync(config_file,
source,
target,
override_cluster_name,
down,
all_nodes=False):
"""Rsyncs files.
Arguments:
config_file: path to the cluster yaml
source: source dir
target: target dir
override_cluster_name: set the name of the cluster
down: whether we're syncing remote -> local
all_nodes: whether to sync worker nodes in addition to the head node
"""
assert bool(source) == bool(target), (
"Must either provide both or neither source and target.")
config = yaml.safe_load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config)
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
nodes = []
if all_nodes:
# technically we re-open the provider for no reason
# in get_worker_nodes but it's cleaner this way
# and _get_head_node does this too
nodes = _get_worker_nodes(config, override_cluster_name)
nodes += [
_get_head_node(
config,
config_file,
override_cluster_name,
create_if_needed=False)
]
for node_id in nodes:
updater = NodeUpdaterThread(
node_id=node_id,
provider_config=config["provider"],
provider=provider,
auth_config=config["auth"],
cluster_name=config["cluster_name"],
file_mounts=config["file_mounts"],
initialization_commands=[],
setup_commands=[],
ray_start_commands=[],
runtime_hash="",
)
if down:
rsync = updater.rsync_down
else:
rsync = updater.rsync_up
if source and target:
rsync(source, target)
else:
updater.sync_file_mounts(rsync)
finally:
provider.cleanup()
def exec_cluster(config_file,
cmd=None,
docker=False,
screen=False,
tmux=False,
stop=False,
start=False,
override_cluster_name=None,
port_forward=None,
with_output=False):
"""Runs a command on the specified cluster.
Arguments:
config_file: path to the cluster yaml
cmd: command to run
docker: whether to run command in docker container of config
screen: whether to run in a screen
tmux: whether to run in a tmux session
stop: whether to stop the cluster after command run
start: whether to start the cluster if it isn't up
override_cluster_name: set the name of the cluster
port_forward (int or list[int]): port(s) to forward
"""
assert not (screen and tmux), "Can specify only one of `screen` or `tmux`."
config = yaml.safe_load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config)
head_node = _get_head_node(
config, config_file, override_cluster_name, create_if_needed=start)
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
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=[],
setup_commands=[],
ray_start_commands=[],
runtime_hash="",
)
def wrap_docker(command):
container_name = config["docker"]["container_name"]
if not container_name:
raise ValueError("Docker container not specified in config.")
return with_docker_exec(
[command], container_name=container_name)[0]
if cmd:
cmd = wrap_docker(cmd) if docker else cmd
if stop:
shutdown_cmd = (
"ray stop; ray teardown ~/ray_bootstrap_config.yaml "
"--yes --workers-only")
if docker:
shutdown_cmd = wrap_docker(shutdown_cmd)
cmd += ("; {}; sudo shutdown -h now".format(shutdown_cmd))
result = _exec(
updater,
cmd,
screen,
tmux,
port_forward=port_forward,
with_output=with_output)
if tmux or screen:
attach_command_parts = ["ray attach", config_file]
if override_cluster_name is not None:
attach_command_parts.append(
"--cluster-name={}".format(override_cluster_name))
if tmux:
attach_command_parts.append("--tmux")
elif screen:
attach_command_parts.append("--screen")
attach_command = " ".join(attach_command_parts)
attach_info = "Use `{}` to check on command status.".format(
attach_command)
logger.info(attach_info)
return result
finally:
provider.cleanup()
def get_or_create_head_node(config, config_file, no_restart, restart_only, yes,
override_cluster_name):
"""Create the cluster head node, which in turn creates the workers."""
provider = get_node_provider(config["provider"], config["cluster_name"])
config_file = os.path.abspath(config_file)
try:
head_node_tags = {
TAG_RAY_NODE_TYPE: NODE_TYPE_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:
confirm("This will create a new cluster", yes)
elif not no_restart:
confirm("This will restart cluster services", yes)
launch_hash = hash_launch_conf(config["head_node"], config["auth"])
if head_node is None or provider.node_tags(head_node).get(
TAG_RAY_LAUNCH_CONFIG) != launch_hash:
if head_node is not None:
confirm("Head node config out-of-date. It will be terminated",
yes)
logger.info(
"get_or_create_head_node: "
"Shutting down outdated head node {}".format(head_node))
provider.terminate_node(head_node)
logger.info("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(config["head_node"], head_node_tags, 1)
start = time.time()
head_node = None
while True:
if time.time() - start > 5:
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(1)
# 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.
runtime_hash = hash_runtime_conf(config["file_mounts"], config)
logger.info("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)
if config["provider"]["type"] != "kubernetes":
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
# 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 config["provider"]["type"] != "kubernetes":
config["file_mounts"].update({
remote_key_path: config["auth"]["ssh_private_key"],
})
if restart_only:
init_commands = []
ray_start_commands = config["head_start_ray_commands"]
elif no_restart:
init_commands = config["head_setup_commands"]
ray_start_commands = []
else:
init_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=init_commands,
ray_start_commands=ray_start_commands,
runtime_hash=runtime_hash,
)
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:
logger.error("get_or_create_head_node: "
"Updating {} failed".format(head_node_ip))
sys.exit(1)
logger.info(
"get_or_create_head_node: "
"Head node up-to-date, IP address is: {}".format(head_node_ip))
monitor_str = "tail -n 100 -f /tmp/ray/session_*/logs/monitor*"
use_docker = "docker" in config and bool(
config["docker"]["container_name"])
if override_cluster_name:
modifiers = " --cluster-name={}".format(
quote(override_cluster_name))
else:
modifiers = ""
print("To monitor auto-scaling activity, you can run:\n\n"
" ray exec {} {}{}{}\n".format(
config_file, "--docker " if use_docker else "",
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()))
finally:
provider.cleanup()
def rsync(config_file: str,
source: Optional[str],
target: Optional[str],
override_cluster_name: Optional[str],
down: bool,
no_config_cache: bool = False,
all_nodes: bool = False):
"""Rsyncs files.
Arguments:
config_file: path to the cluster yaml
source: source dir
target: target dir
override_cluster_name: set the name of the cluster
down: whether we're syncing remote -> local
all_nodes: whether to sync worker nodes in addition to the head node
"""
if bool(source) != bool(target):
cli_logger.abort(
"Expected either both a source and a target, or neither.")
assert bool(source) == bool(target), (
"Must either provide both or neither source and target.")
config = yaml.safe_load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config, no_config_cache=no_config_cache)
is_file_mount = False
for remote_mount in config.get("file_mounts", {}).keys():
if remote_mount in (source if down else target):
is_file_mount = True
break
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
nodes = []
if all_nodes:
# technically we re-open the provider for no reason
# in get_worker_nodes but it's cleaner this way
# and _get_head_node does this too
nodes = _get_worker_nodes(config, override_cluster_name)
head_node = _get_head_node(config,
config_file,
override_cluster_name,
create_if_needed=False)
nodes += [head_node]
for node_id in nodes:
updater = NodeUpdaterThread(node_id=node_id,
provider_config=config["provider"],
provider=provider,
auth_config=config["auth"],
cluster_name=config["cluster_name"],
file_mounts=config["file_mounts"],
initialization_commands=[],
setup_commands=[],
ray_start_commands=[],
runtime_hash="",
file_mounts_contents_hash="",
is_head_node=(node_id == head_node),
docker_config=config.get("docker"))
if down:
rsync = updater.rsync_down
else:
rsync = updater.rsync_up
if source and target:
# print rsync progress for single file rsync
cmd_output_util.set_output_redirected(False)
set_rsync_silent(False)
rsync(source, target, is_file_mount)
else:
updater.sync_file_mounts(rsync)
finally:
provider.cleanup()
def exec_cluster(config_file: str,
*,
cmd: Any = None,
run_env: str = "auto",
screen: bool = False,
tmux: bool = False,
stop: bool = False,
start: bool = False,
override_cluster_name: Optional[str] = None,
no_config_cache: bool = False,
port_forward: Any = None,
with_output: bool = False):
"""Runs a command on the specified cluster.
Arguments:
config_file: path to the cluster yaml
cmd: command to run
run_env: whether to run the command on the host or in a container.
Select between "auto", "host" and "docker"
screen: whether to run in a screen
tmux: whether to run in a tmux session
stop: whether to stop the cluster after command run
start: whether to start the cluster if it isn't up
override_cluster_name: set the name of the cluster
port_forward (int or list[int]): port(s) to forward
"""
assert not (screen and tmux), "Can specify only one of `screen` or `tmux`."
assert run_env in RUN_ENV_TYPES, "--run_env must be in {}".format(
RUN_ENV_TYPES)
# TODO(rliaw): We default this to True to maintain backwards-compat.
# In the future we would want to support disabling login-shells
# and interactivity.
cmd_output_util.set_allow_interactive(True)
config = yaml.safe_load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config, no_config_cache=no_config_cache)
head_node = _get_head_node(config,
config_file,
override_cluster_name,
create_if_needed=start)
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
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=[],
setup_commands=[],
ray_start_commands=[],
runtime_hash="",
file_mounts_contents_hash="",
is_head_node=True,
docker_config=config.get("docker"))
is_docker = isinstance(updater.cmd_runner, DockerCommandRunner)
if cmd and stop:
cmd += "; ".join([
"ray stop",
"ray teardown ~/ray_bootstrap_config.yaml --yes --workers-only"
])
if is_docker and run_env == "docker":
updater.cmd_runner.shutdown_after_next_cmd()
else:
cmd += "; sudo shutdown -h now"
result = _exec(updater,
cmd,
screen,
tmux,
port_forward=port_forward,
with_output=with_output,
run_env=run_env)
if tmux or screen:
attach_command_parts = ["ray attach", config_file]
if override_cluster_name is not None:
attach_command_parts.append(
"--cluster-name={}".format(override_cluster_name))
if tmux:
attach_command_parts.append("--tmux")
elif screen:
attach_command_parts.append("--screen")
attach_command = " ".join(attach_command_parts)
cli_logger.print("Run `{}` to check command status.",
cf.bold(attach_command))
attach_info = "Use `{}` to check on command status.".format(
attach_command)
cli_logger.old_info(logger, attach_info)
return result
finally:
provider.cleanup()
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)
raw_config_file = config_file # used for printing to the user
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(1)
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
# 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_*/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 {}{} {}"), raw_config_file,
modifiers, quote(monitor_str))
cli_logger.print("Connect to a terminal on the cluster head")
cli_logger.print(cf.bold(" ray attach {}{}"), raw_config_file,
modifiers)
finally:
provider.cleanup()
def get_or_create_head_node(config, config_file, no_restart, restart_only, yes,
override_cluster_name):
"""Create the cluster head node, which in turn creates the workers."""
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
head_node_tags = {
TAG_RAY_NODE_TYPE: "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:
confirm("This will create a new cluster", yes)
elif not no_restart:
confirm("This will restart cluster services", yes)
launch_hash = hash_launch_conf(config["head_node"], config["auth"])
if head_node is None or provider.node_tags(head_node).get(
TAG_RAY_LAUNCH_CONFIG) != launch_hash:
if head_node is not None:
confirm("Head node config out-of-date. It will be terminated",
yes)
logger.info(
"get_or_create_head_node: "
"Terminating outdated head node {}".format(head_node))
provider.terminate_node(head_node)
logger.info("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(config["head_node"], head_node_tags, 1)
nodes = provider.non_terminated_nodes(head_node_tags)
assert len(nodes) == 1, "Failed to create head node."
head_node = nodes[0]
# 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.
runtime_hash = hash_runtime_conf(config["file_mounts"], config)
logger.info("get_or_create_head_node: Updating files on head node...")
# Rewrite the auth config so that the head node can update the workers
remote_key_path = "~/ray_bootstrap_key.pem"
remote_config = copy.deepcopy(config)
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
# 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({
remote_key_path: config["auth"]["ssh_private_key"],
"~/ray_bootstrap_config.yaml": remote_config_file.name
})
if restart_only:
init_commands = config["head_start_ray_commands"]
elif no_restart:
init_commands = (
config["setup_commands"] + config["head_setup_commands"])
else:
init_commands = (
config["setup_commands"] + config["head_setup_commands"] +
config["head_start_ray_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=init_commands,
runtime_hash=runtime_hash,
)
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:
logger.error("get_or_create_head_node: "
"Updating {} failed".format(head_node_ip))
sys.exit(1)
logger.info(
"get_or_create_head_node: "
"Head node up-to-date, IP address is: {}".format(head_node_ip))
monitor_str = "tail -n 100 -f /tmp/ray/session_*/logs/monitor*"
use_docker = bool(config["docker"]["container_name"])
if override_cluster_name:
modifiers = " --cluster-name={}".format(
quote(override_cluster_name))
else:
modifiers = ""
print("To monitor auto-scaling activity, you can run:\n\n"
" ray exec {} {}{}{}\n".format(
config_file, "--docker " if use_docker else " ",
quote(monitor_str), modifiers))
print("To open a console on the cluster:\n\n"
" ray attach {}{}\n".format(config_file, modifiers))
print("To ssh manually to the cluster, run:\n\n"
" ssh -i {} {}@{}\n".format(config["auth"]["ssh_private_key"],
config["auth"]["ssh_user"],
head_node_ip))
finally:
provider.cleanup()
def exec_cluster(config_file,
*,
cmd=None,
run_env="auto",
screen=False,
tmux=False,
stop=False,
start=False,
override_cluster_name=None,
port_forward=None,
with_output=False):
"""Runs a command on the specified cluster.
Arguments:
config_file: path to the cluster yaml
cmd: command to run
run_env: whether to run the command on the host or in a container.
Select between "auto", "host" and "docker"
screen: whether to run in a screen
tmux: whether to run in a tmux session
stop: whether to stop the cluster after command run
start: whether to start the cluster if it isn't up
override_cluster_name: set the name of the cluster
port_forward (int or list[int]): port(s) to forward
"""
assert not (screen and tmux), "Can specify only one of `screen` or `tmux`."
assert run_env in RUN_ENV_TYPES, "--run_env must be in {}".format(
RUN_ENV_TYPES)
config = yaml.safe_load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config)
head_node = _get_head_node(config,
config_file,
override_cluster_name,
create_if_needed=start)
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
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=[],
setup_commands=[],
ray_start_commands=[],
runtime_hash="",
docker_config=config.get("docker"))
is_docker = isinstance(updater.cmd_runner, DockerCommandRunner)
if cmd and stop:
cmd += "; ".join([
"ray stop",
"ray teardown ~/ray_bootstrap_config.yaml --yes --workers-only"
])
if is_docker and run_env == "docker":
updater.cmd_runner.shutdown_after_next_cmd()
else:
cmd += "; sudo shutdown -h now"
result = _exec(updater,
cmd,
screen,
tmux,
port_forward=port_forward,
with_output=with_output,
run_env=run_env)
if tmux or screen:
attach_command_parts = ["ray attach", config_file]
if override_cluster_name is not None:
attach_command_parts.append(
"--cluster-name={}".format(override_cluster_name))
if tmux:
attach_command_parts.append("--tmux")
elif screen:
attach_command_parts.append("--screen")
attach_command = " ".join(attach_command_parts)
attach_info = "Use `{}` to check on command status.".format(
attach_command)
logger.info(attach_info)
return result
finally:
provider.cleanup()
def exec_cluster(config_file, cmd, screen, tmux, stop, start,
override_cluster_name, port_forward):
"""Runs a command on the specified cluster.
Arguments:
config_file: path to the cluster yaml
cmd: command to run
screen: whether to run in a screen
tmux: whether to run in a tmux session
stop: whether to stop the cluster after command run
start: whether to start the cluster if it isn't up
override_cluster_name: set the name of the cluster
port_forward: port to forward
"""
assert not (screen and tmux), "Can specify only one of `screen` or `tmux`."
config = yaml.load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = _bootstrap_config(config)
head_node = _get_head_node(config,
config_file,
override_cluster_name,
create_if_needed=start)
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
updater = NodeUpdaterThread(
head_node,
config["provider"],
provider,
config["auth"],
config["cluster_name"],
config["file_mounts"],
[],
"",
)
if stop:
cmd += (
"; ray stop; ray teardown ~/ray_bootstrap_config.yaml --yes "
"--workers-only; sudo shutdown -h now")
_exec(updater,
cmd,
screen,
tmux,
expect_error=stop,
port_forward=port_forward)
if tmux or screen:
attach_command_parts = ["ray attach", config_file]
if override_cluster_name is not None:
attach_command_parts.append(
"--cluster-name={}".format(override_cluster_name))
if tmux:
attach_command_parts.append("--tmux")
elif screen:
attach_command_parts.append("--screen")
attach_command = " ".join(attach_command_parts)
attach_info = "Use `{}` to check on command status.".format(
attach_command)
logger.info(attach_info)
finally:
provider.cleanup()
def get_or_create_head_node(config, config_file, no_restart, restart_only, yes,
override_cluster_name):
"""Create the cluster head node, which in turn creates the workers."""
provider = get_node_provider(config["provider"], config["cluster_name"])
try:
head_node_tags = {
TAG_RAY_NODE_TYPE: "head",
}
nodes = provider.nodes(head_node_tags)
if len(nodes) > 0:
head_node = nodes[0]
else:
head_node = None
if not head_node:
confirm("This will create a new cluster", yes)
elif not no_restart:
confirm("This will restart cluster services", yes)
launch_hash = hash_launch_conf(config["head_node"], config["auth"])
if head_node is None or provider.node_tags(head_node).get(
TAG_RAY_LAUNCH_CONFIG) != launch_hash:
if head_node is not None:
confirm("Head node config out-of-date. It will be terminated",
yes)
logger.info(
"get_or_create_head_node: "
"Terminating outdated head node {}".format(head_node))
provider.terminate_node(head_node)
logger.info("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(config["head_node"], head_node_tags, 1)
nodes = provider.nodes(head_node_tags)
assert len(nodes) == 1, "Failed to create head node."
head_node = nodes[0]
# 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.
runtime_hash = hash_runtime_conf(config["file_mounts"], config)
logger.info("get_or_create_head_node: Updating files on head node...")
# Rewrite the auth config so that the head node can update the workers
remote_key_path = "~/ray_bootstrap_key.pem"
remote_config = copy.deepcopy(config)
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
# 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({
remote_key_path:
config["auth"]["ssh_private_key"],
"~/ray_bootstrap_config.yaml":
remote_config_file.name
})
if restart_only:
init_commands = config["head_start_ray_commands"]
elif no_restart:
init_commands = (config["setup_commands"] +
config["head_setup_commands"])
else:
init_commands = (config["setup_commands"] +
config["head_setup_commands"] +
config["head_start_ray_commands"])
updater = NodeUpdaterThread(
head_node,
config["provider"],
provider,
config["auth"],
config["cluster_name"],
config["file_mounts"],
init_commands,
runtime_hash,
)
updater.start()
updater.join()
# Refresh the node cache so we see the external ip if available
provider.nodes(head_node_tags)
if updater.exitcode != 0:
logger.error("get_or_create_head_node: "
"Updating {} failed".format(
provider.external_ip(head_node)))
sys.exit(1)
logger.info("get_or_create_head_node: "
"Head node up-to-date, IP address is: {}".format(
provider.external_ip(head_node)))
monitor_str = "tail -n 100 -f /tmp/ray/session_*/logs/monitor*"
for s in init_commands:
if ("ray start" in s and "docker exec" in s
and "--autoscaling-config" in s):
monitor_str = "docker exec {} /bin/sh -c {}".format(
config["docker"]["container_name"], quote(monitor_str))
if override_cluster_name:
modifiers = " --cluster-name={}".format(
quote(override_cluster_name))
else:
modifiers = ""
print("To monitor auto-scaling 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 ssh manually to the cluster, run:\n\n"
" ssh -i {} {}@{}\n".format(config["auth"]["ssh_private_key"],
config["auth"]["ssh_user"],
provider.external_ip(head_node)))
finally:
provider.cleanup()