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)
updater = NodeUpdaterProcess(head_node,
config["provider"],
config["auth"],
config["cluster_name"],
config["file_mounts"], [],
"",
redirect_output=False)
if down:
rsync = updater.rsync_down
else:
rsync = updater.rsync_up
rsync(source, target, check_error=False)
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("Terminating worker {}".format(node))
updater = NodeUpdaterProcess(node,
config["provider"],
config["auth"],
config["cluster_name"],
config["file_mounts"], [],
"",
redirect_output=False)
_exec(updater, "ray stop", False, False)
time.sleep(5)
return provider.external_ip(node)
def exec_cluster(config_file, cmd, screen, 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
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
"""
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)
updater = NodeUpdaterProcess(
head_node,
config["provider"],
config["auth"],
config["cluster_name"],
config["file_mounts"], [],
"",
redirect_output=False)
if stop:
cmd += ("; ray stop; ray teardown ~/ray_bootstrap_config.yaml --yes "
"--workers-only; sudo shutdown -h now")
_exec(updater, cmd, screen, expect_error=stop, port_forward=port_forward)
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)
updater = NodeUpdaterProcess(head_node,
config["provider"],
config["auth"],
config["cluster_name"],
config["file_mounts"], [],
"",
redirect_output=False)
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)
def get_head_updater(config):
provider = get_provider(config)
head_node_tags = {
TAG_RAY_NODE_TYPE: "Head",
}
nodes = provider.nodes(head_node_tags)
if len(nodes) > 0:
head_node = nodes[0]
else:
print("Head node of cluster ({}) not found!".format(
config["cluster_name"]))
sys.exit(1)
runtime_hash = hash_runtime_conf(config["file_mounts"], config)
return NodeUpdaterProcess(head_node,
config["provider"],
config["auth"],
config["cluster_name"],
config["file_mounts"], [],
runtime_hash,
redirect_output=False)
def file_sync(config_file):
"""Returns head node IP for given configuration file if exists."""
config = yaml.load(open(config_file).read())
validate_config(config)
config = fillout_defaults(config)
importer = NODE_PROVIDERS.get(config["provider"]["type"])
if not importer:
raise NotImplementedError("Unsupported provider {}".format(
config["provider"]))
bootstrap_config, provider_cls = importer()
config = bootstrap_config(config)
provider = provider_cls(config["provider"], config["cluster_name"])
head_node_tags = {
TAG_RAY_NODE_TYPE: "Head",
}
nodes = provider.nodes(head_node_tags)
if len(nodes) > 0:
head_node = nodes[0]
else:
print("Head node of cluster ({}) not found!".format(
config["cluster_name"]))
sys.exit(1)
runtime_hash = hash_runtime_conf(config["file_mounts"], config)
updater = NodeUpdaterProcess(
head_node,
config["provider"],
config["auth"],
config["cluster_name"],
config["file_mounts"],
[],
runtime_hash,
redirect_output=False)
updater.sync_files(config["file_mounts"])
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"])
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("Terminating outdated head node {}".format(head_node))
provider.terminate_node(head_node)
logger.info("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 in this case.
runtime_hash = hash_runtime_conf(config["file_mounts"], config)
logger.info("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 = NodeUpdaterProcess(head_node,
config["provider"],
config["auth"],
config["cluster_name"],
config["file_mounts"],
init_commands,
runtime_hash,
redirect_output=False)
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("Updating {} failed".format(
provider.external_ip(head_node)))
sys.exit(1)
logger.info("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)))
def get_or_create_head_node(config, no_restart):
"""Create the cluster head node, which in turn creates the workers."""
provider = get_node_provider(config["provider"], config["cluster_name"])
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")
elif not no_restart:
confirm("This will restart your cluster")
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:
print("Terminating outdated head node {}".format(head_node))
provider.terminate_node(head_node)
print("Launching new head node...")
head_node_tags[TAG_RAY_LAUNCH_CONFIG] = launch_hash
head_node_tags[TAG_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 in this case.
runtime_hash = hash_runtime_conf(config["file_mounts"], config)
print("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 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 = NodeUpdaterProcess(head_node,
config["provider"],
config["auth"],
config["cluster_name"],
config["file_mounts"],
init_commands,
runtime_hash,
redirect_output=False)
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:
print("Error: updating {} failed".format(
provider.external_ip(head_node)))
sys.exit(1)
print("Head node up-to-date, IP address is: {}".format(
provider.external_ip(head_node)))
print("To monitor auto-scaling activity, you can run:\n\n"
" ssh -i {} {}@{} 'tail -f /tmp/raylogs/monitor-*'\n".format(
config["auth"]["ssh_private_key"], config["auth"]["ssh_user"],
provider.external_ip(head_node)))
print("To login to the cluster, run:\n\n"
" ssh -i {} {}@{}\n".format(config["auth"]["ssh_private_key"],
config["auth"]["ssh_user"],
provider.external_ip(head_node)))
def get_or_create_head_node(config):
"""Create the cluster head node, which in turn creates the workers."""
provider = get_node_provider(config["provider"], config["cluster_name"])
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
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:
print("Terminating outdated head node {}".format(head_node))
provider.terminate_node(head_node)
print("Launching new head node...")
head_node_tags[TAG_RAY_LAUNCH_CONFIG] = launch_hash
head_node_tags[TAG_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]
runtime_hash = hash_runtime_conf(config["file_mounts"], config)
if provider.node_tags(head_node).get(
TAG_RAY_RUNTIME_CONFIG) != runtime_hash:
print("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
# 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
})
updater = NodeUpdaterProcess(
head_node,
config["provider"],
config["auth"],
config["cluster_name"],
config["file_mounts"],
config["head_init_commands"],
runtime_hash,
redirect_output=False)
updater.start()
updater.join()
if updater.exitcode != 0:
print("Error: updating {} failed".format(
provider.external_ip(head_node)))
sys.exit(1)
print(
"Head node up-to-date, IP address is: {}".format(
provider.external_ip(head_node)))
print(
"To monitor auto-scaling activity, you can run:\n\n"
" ssh -i {} {}@{} 'tail -f /tmp/raylogs/monitor-*'\n".format(
config["auth"]["ssh_private_key"],
config["auth"]["ssh_user"],
provider.external_ip(head_node)))