def testImportingCorrectClass(self):
"""Check correct import when coordinator_address is in config yaml."""
provider_config = {"coordinator_address": "fake_address:1234"}
coordinator_node_provider = _NODE_PROVIDERS.get("local")(provider_config)
assert coordinator_node_provider is CoordinatorSenderNodeProvider
local_node_provider = _NODE_PROVIDERS.get("local")({})
assert local_node_provider is LocalNodeProvider
def _fillout_available_node_types_resources(
cluster_config: Dict[str, Any]) -> Dict[str, Any]:
"""Fills out missing "resources" field for available_node_types."""
if "available_node_types" in cluster_config:
importer = _NODE_PROVIDERS.get(cluster_config["provider"]["type"])
if importer is not None:
provider_cls = importer(cluster_config["provider"])
return provider_cls.fillout_available_node_types_resources(
cluster_config)
return cluster_config
def _bootstrap_config(config: Dict[str, Any],
no_config_cache: bool = False) -> Dict[str, Any]:
config = prepare_config(config)
hasher = hashlib.sha1()
hasher.update(json.dumps([config], sort_keys=True).encode("utf-8"))
cache_key = os.path.join(tempfile.gettempdir(),
"ray-config-{}".format(hasher.hexdigest()))
if os.path.exists(cache_key) and not no_config_cache:
cli_logger.old_info(logger, "Using cached config at {}", cache_key)
config_cache = json.loads(open(cache_key).read())
if config_cache.get("_version", -1) == CONFIG_CACHE_VERSION:
# todo: is it fine to re-resolve? afaik it should be.
# we can have migrations otherwise or something
# but this seems overcomplicated given that resolving is
# relatively cheap
try_reload_log_state(config_cache["config"]["provider"],
config_cache.get("provider_log_info"))
if log_once("_printed_cached_config_warning"):
cli_logger.verbose_warning(
"Loaded cached provider configuration "
"from " + cf.bold("{}"), cache_key)
if cli_logger.verbosity == 0:
cli_logger.warning("Loaded cached provider configuration")
cli_logger.warning(
"If you experience issues with "
"the cloud provider, try re-running "
"the command with {}.", cf.bold("--no-config-cache"))
return config_cache["config"]
else:
cli_logger.warning(
"Found cached cluster config "
"but the version " + cf.bold("{}") + " "
"(expected " + cf.bold("{}") + ") does not match.\n"
"This is normal if cluster launcher was updated.\n"
"Config will be re-resolved.",
config_cache.get("_version", "none"), CONFIG_CACHE_VERSION)
validate_config(config)
importer = _NODE_PROVIDERS.get(config["provider"]["type"])
if not importer:
raise NotImplementedError("Unsupported provider {}".format(
config["provider"]))
provider_cls = importer(config["provider"])
def testValidateDefaultConfigAWSMultiNodeTypes(self):
aws_config_path = os.path.join(
RAY_PATH, "autoscaler/aws/example-multi-node-type.yaml")
with open(aws_config_path) as f:
config = yaml.safe_load(f)
new_config = copy.deepcopy(config)
# modify it here
new_config["available_node_types"] = {
"cpu_4_ondemand":
new_config["available_node_types"]["cpu_4_ondemand"],
"cpu_16_spot": new_config["available_node_types"]["cpu_16_spot"],
"gpu_8_ondemand":
new_config["available_node_types"]["gpu_8_ondemand"]
}
orig_new_config = copy.deepcopy(new_config)
expected_available_node_types = orig_new_config["available_node_types"]
expected_available_node_types["cpu_4_ondemand"]["resources"] = {
"CPU": 4
}
expected_available_node_types["cpu_16_spot"]["resources"] = {
"CPU": 16,
"memory": 41231686041,
"Custom1": 1,
"is_spot": 1
}
expected_available_node_types["gpu_8_ondemand"]["resources"] = {
"CPU": 32,
"memory": 157195803033,
"GPU": 4,
"accelerator_type:V100": 1
}
boto3_dict = {
"InstanceTypes": [{
"InstanceType": "m4.xlarge",
"VCpuInfo": {
"DefaultVCpus": 4
},
"MemoryInfo": {
"SizeInMiB": 16384
}
}, {
"InstanceType": "m4.4xlarge",
"VCpuInfo": {
"DefaultVCpus": 16
},
"MemoryInfo": {
"SizeInMiB": 65536
}
}, {
"InstanceType": "p3.8xlarge",
"VCpuInfo": {
"DefaultVCpus": 32
},
"MemoryInfo": {
"SizeInMiB": 249856
},
"GpuInfo": {
"Gpus": [{
"Name": "V100",
"Count": 4
}]
}
}]
}
describe_instance_types_mock = Mock()
describe_instance_types_mock.describe_instance_types = MagicMock(
return_value=boto3_dict)
client_cache_mock = MagicMock(
return_value=describe_instance_types_mock)
with patch.multiple(
"ray.autoscaler._private.aws.node_provider",
client_cache=client_cache_mock,
):
new_config = prepare_config(new_config)
importer = _NODE_PROVIDERS.get(new_config["provider"]["type"])
provider_cls = importer(new_config["provider"])
try:
new_config = \
provider_cls.fillout_available_node_types_resources(
new_config)
validate_config(new_config)
expected_available_node_types == new_config[
"available_node_types"]
except Exception:
self.fail(
"Config did not pass multi node types auto fill test!")
def create_or_update_cluster(config_file: str,
override_min_workers: Optional[int],
override_max_workers: Optional[int],
no_restart: bool,
restart_only: bool,
yes: bool,
override_cluster_name: Optional[str] = None,
no_config_cache: bool = False,
redirect_command_output: Optional[bool] = False,
use_login_shells: bool = True) -> None:
"""Create or updates an autoscaling Ray cluster from a config json."""
set_using_login_shells(use_login_shells)
if not use_login_shells:
cmd_output_util.set_allow_interactive(False)
if redirect_command_output is None:
# Do not redirect by default.
cmd_output_util.set_output_redirected(False)
else:
cmd_output_util.set_output_redirected(redirect_command_output)
def handle_yaml_error(e):
cli_logger.error("Cluster config invalid")
cli_logger.newline()
cli_logger.error("Failed to load YAML file " + cf.bold("{}"),
config_file)
cli_logger.newline()
with cli_logger.verbatim_error_ctx("PyYAML error:"):
cli_logger.error(e)
cli_logger.abort()
try:
config = yaml.safe_load(open(config_file).read())
except FileNotFoundError:
cli_logger.abort(
"Provided cluster configuration file ({}) does not exist",
cf.bold(config_file))
raise
except yaml.parser.ParserError as e:
handle_yaml_error(e)
raise
except yaml.scanner.ScannerError as e:
handle_yaml_error(e)
raise
# todo: validate file_mounts, ssh keys, etc.
importer = _NODE_PROVIDERS.get(config["provider"]["type"])
if not importer:
cli_logger.abort(
"Unknown provider type " + cf.bold("{}") + "\n"
"Available providers are: {}", config["provider"]["type"],
cli_logger.render_list([
k for k in _NODE_PROVIDERS.keys()
if _NODE_PROVIDERS[k] is not None
]))
raise NotImplementedError("Unsupported provider {}".format(
config["provider"]))
printed_overrides = False
def handle_cli_override(key, override):
if override is not None:
if key in config:
nonlocal printed_overrides
printed_overrides = True
cli_logger.warning(
"`{}` override provided on the command line.\n"
" Using " + cf.bold("{}") + cf.dimmed(
" [configuration file has " + cf.bold("{}") + "]"),
key, override, config[key])
config[key] = override
handle_cli_override("min_workers", override_min_workers)
handle_cli_override("max_workers", override_max_workers)
handle_cli_override("cluster_name", override_cluster_name)
if printed_overrides:
cli_logger.newline()
cli_logger.labeled_value("Cluster", config["cluster_name"])
cli_logger.newline()
config = _bootstrap_config(config, no_config_cache=no_config_cache)
try_logging_config(config)
get_or_create_head_node(config, config_file, no_restart, restart_only, yes,
override_cluster_name)
def _bootstrap_config(config: Dict[str, Any],
no_config_cache: bool = False) -> Dict[str, Any]:
config = prepare_config(config)
hasher = hashlib.sha1()
hasher.update(json.dumps([config], sort_keys=True).encode("utf-8"))
cache_key = os.path.join(tempfile.gettempdir(),
"ray-config-{}".format(hasher.hexdigest()))
if os.path.exists(cache_key) and not no_config_cache:
config_cache = json.loads(open(cache_key).read())
if config_cache.get("_version", -1) == CONFIG_CACHE_VERSION:
# todo: is it fine to re-resolve? afaik it should be.
# we can have migrations otherwise or something
# but this seems overcomplicated given that resolving is
# relatively cheap
try_reload_log_state(config_cache["config"]["provider"],
config_cache.get("provider_log_info"))
if log_once("_printed_cached_config_warning"):
cli_logger.verbose_warning(
"Loaded cached provider configuration "
"from " + cf.bold("{}"), cache_key)
if cli_logger.verbosity == 0:
cli_logger.warning("Loaded cached provider configuration")
cli_logger.warning(
"If you experience issues with "
"the cloud provider, try re-running "
"the command with {}.", cf.bold("--no-config-cache"))
return config_cache["config"]
else:
cli_logger.warning(
"Found cached cluster config "
"but the version " + cf.bold("{}") + " "
"(expected " + cf.bold("{}") + ") does not match.\n"
"This is normal if cluster launcher was updated.\n"
"Config will be re-resolved.",
config_cache.get("_version", "none"), CONFIG_CACHE_VERSION)
importer = _NODE_PROVIDERS.get(config["provider"]["type"])
if not importer:
raise NotImplementedError("Unsupported provider {}".format(
config["provider"]))
provider_cls = importer(config["provider"])
cli_logger.print("Checking {} environment settings",
_PROVIDER_PRETTY_NAMES.get(config["provider"]["type"]))
try:
config = provider_cls.fillout_available_node_types_resources(config)
except Exception as exc:
if cli_logger.verbosity > 2:
logger.exception("Failed to autodetect node resources.")
else:
cli_logger.warning(
f"Failed to autodetect node resources: {str(exc)}. "
"You can see full stack trace with higher verbosity.")
# NOTE: if `resources` field is missing, validate_config for providers
# other than AWS and Kubernetes will fail (the schema error will ask the
# user to manually fill the resources) as we currently support autofilling
# resources for AWS and Kubernetes only.
validate_config(config)
resolved_config = provider_cls.bootstrap_config(config)
if not no_config_cache:
with open(cache_key, "w") as f:
config_cache = {
"_version": CONFIG_CACHE_VERSION,
"provider_log_info": try_get_log_state(config["provider"]),
"config": resolved_config
}
f.write(json.dumps(config_cache))
return resolved_config
def create_or_update_cluster(
config_file: str,
override_min_workers: Optional[int],
override_max_workers: Optional[int],
no_restart: bool,
restart_only: bool,
yes: bool,
override_cluster_name: Optional[str] = None,
no_config_cache: bool = False,
redirect_command_output: Optional[bool] = False,
use_login_shells: bool = True,
no_monitor_on_head: bool = False) -> Dict[str, Any]:
"""Creates or updates an autoscaling Ray cluster from a config json."""
# no_monitor_on_head is an internal flag used by the Ray K8s operator.
# If True, prevents autoscaling config sync to the Ray head during cluster
# creation. See https://github.com/ray-project/ray/pull/13720.
set_using_login_shells(use_login_shells)
if not use_login_shells:
cmd_output_util.set_allow_interactive(False)
if redirect_command_output is None:
# Do not redirect by default.
cmd_output_util.set_output_redirected(False)
else:
cmd_output_util.set_output_redirected(redirect_command_output)
def handle_yaml_error(e):
cli_logger.error("Cluster config invalid")
cli_logger.newline()
cli_logger.error("Failed to load YAML file " + cf.bold("{}"),
config_file)
cli_logger.newline()
with cli_logger.verbatim_error_ctx("PyYAML error:"):
cli_logger.error(e)
cli_logger.abort()
try:
config = yaml.safe_load(open(config_file).read())
except FileNotFoundError:
cli_logger.abort(
"Provided cluster configuration file ({}) does not exist",
cf.bold(config_file))
except yaml.parser.ParserError as e:
handle_yaml_error(e)
raise
except yaml.scanner.ScannerError as e:
handle_yaml_error(e)
raise
global_event_system.execute_callback(CreateClusterEvent.up_started,
{"cluster_config": config})
# todo: validate file_mounts, ssh keys, etc.
importer = _NODE_PROVIDERS.get(config["provider"]["type"])
if not importer:
cli_logger.abort(
"Unknown provider type " + cf.bold("{}") + "\n"
"Available providers are: {}", config["provider"]["type"],
cli_logger.render_list([
k for k in _NODE_PROVIDERS.keys()
if _NODE_PROVIDERS[k] is not None
]))
printed_overrides = False
def handle_cli_override(key, override):
if override is not None:
if key in config:
nonlocal printed_overrides
printed_overrides = True
cli_logger.warning(
"`{}` override provided on the command line.\n"
" Using " + cf.bold("{}") + cf.dimmed(
" [configuration file has " + cf.bold("{}") + "]"),
key, override, config[key])
config[key] = override
handle_cli_override("min_workers", override_min_workers)
handle_cli_override("max_workers", override_max_workers)
handle_cli_override("cluster_name", override_cluster_name)
if printed_overrides:
cli_logger.newline()
cli_logger.labeled_value("Cluster", config["cluster_name"])
cli_logger.newline()
config = _bootstrap_config(config, no_config_cache=no_config_cache)
try_logging_config(config)
get_or_create_head_node(config, config_file, no_restart, restart_only, yes,
override_cluster_name, no_monitor_on_head)
return config