def run(self):
update_start_time = time.time()
if (
cmd_output_util.does_allow_interactive()
and cmd_output_util.is_output_redirected()
):
# this is most probably a bug since the user has no control
# over these settings
msg = (
"Output was redirected for an interactive command. "
"Either do not pass `--redirect-command-output` "
"or also pass in `--use-normal-shells`."
)
cli_logger.abort(msg)
try:
with LogTimer(
self.log_prefix + "Applied config {}".format(self.runtime_hash)
):
self.do_update()
except Exception as e:
self.provider.set_node_tags(
self.node_id, {TAG_RAY_NODE_STATUS: STATUS_UPDATE_FAILED}
)
cli_logger.error("New status: {}", cf.bold(STATUS_UPDATE_FAILED))
cli_logger.error("!!!")
if hasattr(e, "cmd"):
cli_logger.error(
"Setup command `{}` failed with exit code {}. stderr:",
cf.bold(e.cmd),
e.returncode,
)
else:
cli_logger.verbose_error("{}", str(vars(e)))
# todo: handle this better somehow?
cli_logger.error("{}", str(e))
# todo: print stderr here
cli_logger.error("!!!")
cli_logger.newline()
if isinstance(e, click.ClickException):
# todo: why do we ignore this here
return
raise
tags_to_set = {
TAG_RAY_NODE_STATUS: STATUS_UP_TO_DATE,
TAG_RAY_RUNTIME_CONFIG: self.runtime_hash,
}
if self.file_mounts_contents_hash is not None:
tags_to_set[TAG_RAY_FILE_MOUNTS_CONTENTS] = self.file_mounts_contents_hash
self.provider.set_node_tags(self.node_id, tags_to_set)
cli_logger.labeled_value("New status", STATUS_UP_TO_DATE)
self.update_time = time.time() - update_start_time
self.exitcode = 0
def teardown_cluster(config_file: str, yes: bool, workers_only: bool,
override_cluster_name: Optional[str],
keep_min_workers: bool):
"""Destroys all nodes of a Ray cluster described by a config json."""
config = yaml.safe_load(open(config_file).read())
if override_cluster_name is not None:
config["cluster_name"] = override_cluster_name
config = prepare_config(config)
validate_config(config)
cli_logger.confirm(yes, "Destroying cluster.", _abort=True)
cli_logger.old_confirm("This will destroy your cluster", yes)
if not workers_only:
try:
exec_cluster(config_file,
cmd="ray stop",
run_env="auto",
screen=False,
tmux=False,
stop=False,
start=False,
override_cluster_name=override_cluster_name,
port_forward=None,
with_output=False)
except Exception as e:
# todo: add better exception info
cli_logger.verbose_error("{}", str(e))
cli_logger.warning(
"Exception occured when stopping the cluster Ray runtime "
"(use -v to dump teardown exceptions).")
cli_logger.warning(
"Ignoring the exception and "
"attempting to shut down the cluster nodes anyway.")
cli_logger.old_exception(
logger, "Ignoring error attempting a clean shutdown.")
provider = _get_node_provider(config["provider"], config["cluster_name"])
try:
def remaining_nodes():
workers = provider.non_terminated_nodes(
{TAG_RAY_NODE_KIND: NODE_KIND_WORKER})
if keep_min_workers:
min_workers = config.get("min_workers", 0)
cli_logger.print(
"{} random worker nodes will not be shut down. " +
cf.dimmed("(due to {})"), cf.bold(min_workers),
cf.bold("--keep-min-workers"))
cli_logger.old_info(logger,
"teardown_cluster: Keeping {} nodes...",
min_workers)
workers = random.sample(workers, len(workers) - min_workers)
# todo: it's weird to kill the head node but not all workers
if workers_only:
cli_logger.print(
"The head node will not be shut down. " +
cf.dimmed("(due to {})"), cf.bold("--workers-only"))
return workers
head = provider.non_terminated_nodes(
{TAG_RAY_NODE_KIND: NODE_KIND_HEAD})
return head + workers
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}")
# Loop here to check that both the head and worker nodes are actually
# really gone
A = remaining_nodes()
container_name = config.get("docker", {}).get("container_name")
if container_name:
for node in A:
run_docker_stop(node, container_name)
with LogTimer("teardown_cluster: done."):
while A:
cli_logger.old_info(
logger, "teardown_cluster: "
"Shutting down {} nodes...", len(A))
provider.terminate_nodes(A)
cli_logger.print("Requested {} nodes to shut down.",
cf.bold(len(A)),
_tags=dict(interval="1s"))
time.sleep(
POLL_INTERVAL) # todo: interval should be a variable
A = remaining_nodes()
cli_logger.print("{} nodes remaining after {} second(s).",
cf.bold(len(A)), POLL_INTERVAL)
cli_logger.success("No nodes remaining.")
finally:
provider.cleanup()
from ray.autoscaler._private.cli_logger import cli_logger
import colorful as cf
cli_logger.old_style = False
cli_logger.verbosity = 999
cli_logger.detect_colors()
cli_logger.print(
cf.bold("Bold ") + cf.italic("Italic ") + cf.underlined("Underlined"))
cli_logger.labeled_value("Label", "value")
cli_logger.print("List: {}", cli_logger.render_list([1, 2, 3]))
cli_logger.newline()
cli_logger.very_verbose("Very verbose")
cli_logger.verbose("Verbose")
cli_logger.verbose_warning("Verbose warning")
cli_logger.verbose_error("Verbose error")
cli_logger.print("Info")
cli_logger.success("Success")
cli_logger.warning("Warning")
cli_logger.error("Error")
cli_logger.newline()
try:
cli_logger.abort("Abort")
except Exception:
pass
try:
cli_logger.doassert(False, "Assert")
except Exception:
pass
cli_logger.newline()
cli_logger.confirm(True, "example")
def handle_boto_error(exc, msg, *args, **kwargs):
if cli_logger.old_style:
# old-style logging doesn't do anything here
# so we exit early
return
error_code = None
error_info = None
# todo: not sure if these exceptions always have response
if hasattr(exc, "response"):
error_info = exc.response.get("Error", None)
if error_info is not None:
error_code = error_info.get("Code", None)
generic_message_args = [
"{}\n"
"Error code: {}",
msg.format(*args, **kwargs),
cf.bold(error_code)
]
# apparently
# ExpiredTokenException
# ExpiredToken
# RequestExpired
# are all the same pretty much
credentials_expiration_codes = [
"ExpiredTokenException", "ExpiredToken", "RequestExpired"
]
if error_code in credentials_expiration_codes:
# "An error occurred (ExpiredToken) when calling the
# GetInstanceProfile operation: The security token
# included in the request is expired"
# "An error occurred (RequestExpired) when calling the
# DescribeKeyPairs operation: Request has expired."
token_command = (
"aws sts get-session-token "
"--serial-number arn:aws:iam::" + cf.underlined("ROOT_ACCOUNT_ID")
+ ":mfa/" + cf.underlined("AWS_USERNAME") + " --token-code " +
cf.underlined("TWO_FACTOR_AUTH_CODE"))
secret_key_var = (
"export AWS_SECRET_ACCESS_KEY = " + cf.underlined("REPLACE_ME") +
" # found at Credentials.SecretAccessKey")
session_token_var = (
"export AWS_SESSION_TOKEN = " + cf.underlined("REPLACE_ME") +
" # found at Credentials.SessionToken")
access_key_id_var = (
"export AWS_ACCESS_KEY_ID = " + cf.underlined("REPLACE_ME") +
" # found at Credentials.AccessKeyId")
# fixme: replace with a Github URL that points
# to our repo
aws_session_script_url = ("https://gist.github.com/maximsmol/"
"a0284e1d97b25d417bd9ae02e5f450cf")
cli_logger.verbose_error(*generic_message_args)
cli_logger.verbose(vars(exc))
cli_logger.panic("Your AWS session has expired.")
cli_logger.newline()
cli_logger.panic("You can request a new one using")
cli_logger.panic(cf.bold(token_command))
cli_logger.panic("then expose it to Ray by setting")
cli_logger.panic(cf.bold(secret_key_var))
cli_logger.panic(cf.bold(session_token_var))
cli_logger.panic(cf.bold(access_key_id_var))
cli_logger.newline()
cli_logger.panic("You can find a script that automates this at:")
cli_logger.panic(cf.underlined(aws_session_script_url))
# Do not re-raise the exception here because it looks awful
# and we already print all the info in verbose
cli_logger.abort()
# todo: any other errors that we should catch separately?
cli_logger.panic(*generic_message_args)
cli_logger.newline()
with cli_logger.verbatim_error_ctx("Boto3 error:"):
cli_logger.verbose("{}", str(vars(exc)))
cli_logger.panic("{}", str(exc))
cli_logger.abort()
