class APIHead(dashboard_utils.DashboardHeadModule):
def __init__(self, dashboard_head):
super().__init__(dashboard_head)
self._gcs_job_info_stub = None
self._gcs_actor_info_stub = None
self._dashboard_head = dashboard_head
assert _internal_kv_initialized()
self._job_info_client = JobInfoStorageClient()
# For offloading CPU intensive work.
self._thread_pool = concurrent.futures.ThreadPoolExecutor(
max_workers=2, thread_name_prefix="api_head")
@routes.get("/api/actors/kill")
async def kill_actor_gcs(self, req) -> aiohttp.web.Response:
actor_id = req.query.get("actor_id")
force_kill = req.query.get("force_kill", False) in ("true", "True")
no_restart = req.query.get("no_restart", False) in ("true", "True")
if not actor_id:
return dashboard_optional_utils.rest_response(
success=False, message="actor_id is required.")
request = gcs_service_pb2.KillActorViaGcsRequest()
request.actor_id = bytes.fromhex(actor_id)
request.force_kill = force_kill
request.no_restart = no_restart
await self._gcs_actor_info_stub.KillActorViaGcs(request, timeout=5)
message = (f"Force killed actor with id {actor_id}" if force_kill else
f"Requested actor with id {actor_id} to terminate. " +
"It will exit once running tasks complete")
return dashboard_optional_utils.rest_response(success=True,
message=message)
@routes.get("/api/snapshot")
async def snapshot(self, req):
(
job_info,
job_submission_data,
actor_data,
serve_data,
session_name,
) = await asyncio.gather(
self.get_job_info(),
self.get_job_submission_info(),
self.get_actor_info(),
self.get_serve_info(),
self.get_session_name(),
)
snapshot = {
"jobs": job_info,
"job_submission": job_submission_data,
"actors": actor_data,
"deployments": serve_data,
"session_name": session_name,
"ray_version": ray.__version__,
"ray_commit": ray.__commit__,
}
return dashboard_optional_utils.rest_response(success=True,
message="hello",
snapshot=snapshot)
@routes.get("/api/component_activities")
async def get_component_activities(self, req) -> aiohttp.web.Response:
# Get activity information for driver
timeout = req.query.get("timeout", None)
if timeout and timeout.isdigit():
timeout = int(timeout)
else:
timeout = 5
driver_activity_info = await self._get_job_activity_info(
timeout=timeout)
resp = {"driver": dataclasses.asdict(driver_activity_info)}
return aiohttp.web.Response(
text=json.dumps(resp),
content_type="application/json",
status=aiohttp.web.HTTPOk.status_code,
)
async def _get_job_activity_info(self,
timeout: int) -> RayActivityResponse:
# Returns if there is Ray activity from drivers (job).
# Drivers in namespaces that start with _ray_internal_job_info_ are not
# considered activity.
request = gcs_service_pb2.GetAllJobInfoRequest()
reply = await self._gcs_job_info_stub.GetAllJobInfo(request,
timeout=timeout)
num_active_drivers = 0
for job_table_entry in reply.job_info_list:
is_dead = bool(job_table_entry.is_dead)
in_internal_namespace = job_table_entry.config.ray_namespace.startswith(
JobInfoStorageClient.JOB_DATA_KEY_PREFIX)
if not is_dead and not in_internal_namespace:
num_active_drivers += 1
return RayActivityResponse(
is_active=num_active_drivers > 0,
reason=f"Number of active drivers: {num_active_drivers}"
if num_active_drivers else None,
timestamp=datetime.now().timestamp(),
)
def _get_job_info(self, metadata: Dict[str, str]) -> Optional[JobInfo]:
# If a job submission ID has been added to a job, the status is
# guaranteed to be returned.
job_submission_id = metadata.get(JOB_ID_METADATA_KEY)
return self._job_info_client.get_info(job_submission_id)
async def get_job_info(self):
"""Return info for each job. Here a job is a Ray driver."""
request = gcs_service_pb2.GetAllJobInfoRequest()
reply = await self._gcs_job_info_stub.GetAllJobInfo(request, timeout=5)
jobs = {}
for job_table_entry in reply.job_info_list:
job_id = job_table_entry.job_id.hex()
metadata = dict(job_table_entry.config.metadata)
config = {
"namespace":
job_table_entry.config.ray_namespace,
"metadata":
metadata,
"runtime_env":
RuntimeEnv.deserialize(job_table_entry.config.runtime_env_info.
serialized_runtime_env),
}
info = self._get_job_info(metadata)
entry = {
"status": None if info is None else info.status,
"status_message": None if info is None else info.message,
"is_dead": job_table_entry.is_dead,
"start_time": job_table_entry.start_time,
"end_time": job_table_entry.end_time,
"config": config,
}
jobs[job_id] = entry
return jobs
async def get_job_submission_info(self):
"""Info for Ray job submission. Here a job can have 0 or many drivers."""
jobs = {}
for job_submission_id, job_info in self._job_info_client.get_all_jobs(
).items():
if job_info is not None:
entry = {
"job_submission_id": job_submission_id,
"status": job_info.status,
"message": job_info.message,
"error_type": job_info.error_type,
"start_time": job_info.start_time,
"end_time": job_info.end_time,
"metadata": job_info.metadata,
"runtime_env": job_info.runtime_env,
"entrypoint": job_info.entrypoint,
}
jobs[job_submission_id] = entry
return jobs
async def get_actor_info(self):
# TODO (Alex): GCS still needs to return actors from dead jobs.
request = gcs_service_pb2.GetAllActorInfoRequest()
request.show_dead_jobs = True
reply = await self._gcs_actor_info_stub.GetAllActorInfo(request,
timeout=5)
actors = {}
for actor_table_entry in reply.actor_table_data:
actor_id = actor_table_entry.actor_id.hex()
runtime_env = json.loads(actor_table_entry.serialized_runtime_env)
entry = {
"job_id":
actor_table_entry.job_id.hex(),
"state":
gcs_pb2.ActorTableData.ActorState.Name(
actor_table_entry.state),
"name":
actor_table_entry.name,
"namespace":
actor_table_entry.ray_namespace,
"runtime_env":
runtime_env,
"start_time":
actor_table_entry.start_time,
"end_time":
actor_table_entry.end_time,
"is_detached":
actor_table_entry.is_detached,
"resources":
dict(actor_table_entry.required_resources),
"actor_class":
actor_table_entry.class_name,
"current_worker_id":
actor_table_entry.address.worker_id.hex(),
"current_raylet_id":
actor_table_entry.address.raylet_id.hex(),
"ip_address":
actor_table_entry.address.ip_address,
"port":
actor_table_entry.address.port,
"metadata":
dict(),
}
actors[actor_id] = entry
deployments = await self.get_serve_info()
for _, deployment_info in deployments.items():
for replica_actor_id, actor_info in deployment_info[
"actors"].items():
if replica_actor_id in actors:
serve_metadata = dict()
serve_metadata["replica_tag"] = actor_info[
"replica_tag"]
serve_metadata["deployment_name"] = deployment_info[
"name"]
serve_metadata["version"] = actor_info["version"]
actors[replica_actor_id]["metadata"][
"serve"] = serve_metadata
return actors
async def get_serve_info(self) -> Dict[str, Any]:
# Conditionally import serve to prevent ModuleNotFoundError from serve
# dependencies when only ray[default] is installed (#17712)
try:
from ray.serve.constants import SERVE_CONTROLLER_NAME
from ray.serve.controller import SNAPSHOT_KEY as SERVE_SNAPSHOT_KEY
except Exception:
return {}
# Serve wraps Ray's internal KV store and specially formats the keys.
# These are the keys we are interested in:
# SERVE_CONTROLLER_NAME(+ optional random letters):SERVE_SNAPSHOT_KEY
# TODO: Convert to async GRPC, if CPU usage is not a concern.
def get_deployments():
serve_keys = _internal_kv_list(
SERVE_CONTROLLER_NAME,
namespace=ray_constants.KV_NAMESPACE_SERVE)
serve_snapshot_keys = filter(
lambda k: SERVE_SNAPSHOT_KEY in str(k), serve_keys)
deployments_per_controller: List[Dict[str, Any]] = []
for key in serve_snapshot_keys:
val_bytes = _internal_kv_get(
key, namespace=ray_constants.KV_NAMESPACE_SERVE
) or "{}".encode("utf-8")
deployments_per_controller.append(
json.loads(val_bytes.decode("utf-8")))
# Merge the deployments dicts of all controllers.
deployments: Dict[str, Any] = {
k: v
for d in deployments_per_controller for k, v in d.items()
}
# Replace the keys (deployment names) with their hashes to prevent
# collisions caused by the automatic conversion to camelcase by the
# dashboard agent.
return {
hashlib.sha1(name.encode()).hexdigest(): info
for name, info in deployments.items()
}
return await asyncio.get_event_loop().run_in_executor(
executor=self._thread_pool, func=get_deployments)
async def get_session_name(self):
# TODO(yic): Convert to async GRPC.
def get_session():
return ray.experimental.internal_kv._internal_kv_get(
"session_name",
namespace=ray_constants.KV_NAMESPACE_SESSION).decode()
return await asyncio.get_event_loop().run_in_executor(
executor=self._thread_pool, func=get_session)
async def run(self, server):
self._gcs_job_info_stub = gcs_service_pb2_grpc.JobInfoGcsServiceStub(
self._dashboard_head.aiogrpc_gcs_channel)
self._gcs_actor_info_stub = gcs_service_pb2_grpc.ActorInfoGcsServiceStub(
self._dashboard_head.aiogrpc_gcs_channel)
@staticmethod
def is_minimal_module():
return False
class JobManager:
"""Provide python APIs for job submission and management.
It does not provide persistence, all info will be lost if the cluster
goes down.
"""
JOB_ACTOR_NAME = "_ray_internal_job_actor_{job_id}"
# Time that we will sleep while tailing logs if no new log line is
# available.
LOG_TAIL_SLEEP_S = 1
JOB_MONITOR_LOOP_PERIOD_S = 1
def __init__(self):
self._job_info_client = JobInfoStorageClient()
self._log_client = JobLogStorageClient()
self._supervisor_actor_cls = ray.remote(JobSupervisor)
self._recover_running_jobs()
def _recover_running_jobs(self):
"""Recovers all running jobs from the status client.
For each job, we will spawn a coroutine to monitor it.
Each will be added to self._running_jobs and reconciled.
"""
all_jobs = self._job_info_client.get_all_jobs()
for job_id, job_info in all_jobs.items():
if not job_info.status.is_terminal():
create_task(self._monitor_job(job_id))
def _get_actor_for_job(self, job_id: str) -> Optional[ActorHandle]:
try:
return ray.get_actor(self.JOB_ACTOR_NAME.format(job_id=job_id))
except ValueError: # Ray returns ValueError for nonexistent actor.
return None
async def _monitor_job(self,
job_id: str,
job_supervisor: Optional[ActorHandle] = None):
"""Monitors the specified job until it enters a terminal state.
This is necessary because we need to handle the case where the
JobSupervisor dies unexpectedly.
"""
is_alive = True
if job_supervisor is None:
job_supervisor = self._get_actor_for_job(job_id)
if job_supervisor is None:
logger.error(f"Failed to get job supervisor for job {job_id}.")
self._job_info_client.put_status(
job_id,
JobStatus.FAILED,
message=
"Unexpected error occurred: Failed to get job supervisor.",
)
is_alive = False
while is_alive:
try:
await job_supervisor.ping.remote()
await asyncio.sleep(self.JOB_MONITOR_LOOP_PERIOD_S)
except Exception as e:
is_alive = False
if self._job_info_client.get_status(job_id).is_terminal():
# If the job is already in a terminal state, then the actor
# exiting is expected.
pass
elif isinstance(e, RuntimeEnvSetupError):
logger.info(
f"Failed to set up runtime_env for job {job_id}.")
self._job_info_client.put_status(
job_id,
JobStatus.FAILED,
message=f"runtime_env setup failed: {e}",
)
else:
logger.warning(
f"Job supervisor for job {job_id} failed unexpectedly: {e}."
)
self._job_info_client.put_status(
job_id,
JobStatus.FAILED,
message=f"Unexpected error occurred: {e}",
)
# Kill the actor defensively to avoid leaking actors in unexpected error cases.
if job_supervisor is not None:
ray.kill(job_supervisor, no_restart=True)
def _get_current_node_resource_key(self) -> str:
"""Get the Ray resource key for current node.
It can be used for actor placement.
"""
current_node_id = ray.get_runtime_context().node_id.hex()
for node in ray.nodes():
if node["NodeID"] == current_node_id:
# Found the node.
for key in node["Resources"].keys():
if key.startswith("node:"):
return key
else:
raise ValueError(
"Cannot find the node dictionary for current node.")
def _handle_supervisor_startup(self, job_id: str,
result: Optional[Exception]):
"""Handle the result of starting a job supervisor actor.
If started successfully, result should be None. Otherwise it should be
an Exception.
On failure, the job will be marked failed with a relevant error
message.
"""
if result is None:
return
def submit_job(
self,
*,
entrypoint: str,
job_id: Optional[str] = None,
runtime_env: Optional[Dict[str, Any]] = None,
metadata: Optional[Dict[str, str]] = None,
_start_signal_actor: Optional[ActorHandle] = None,
) -> str:
"""
Job execution happens asynchronously.
1) Generate a new unique id for this job submission, each call of this
method assumes they're independent submission with its own new
ID, job supervisor actor, and child process.
2) Create new detached actor with same runtime_env as job spec
Actual setting up runtime_env, subprocess group, driver command
execution, subprocess cleaning up and running status update to GCS
is all handled by job supervisor actor.
Args:
entrypoint: Driver command to execute in subprocess shell.
Represents the entrypoint to start user application.
runtime_env: Runtime environment used to execute driver command,
which could contain its own ray.init() to configure runtime
env at ray cluster, task and actor level.
metadata: Support passing arbitrary data to driver command in
case needed.
_start_signal_actor: Used in testing only to capture state
transitions between PENDING -> RUNNING. Regular user shouldn't
need this.
Returns:
job_id: Generated uuid for further job management. Only valid
within the same ray cluster.
"""
if job_id is None:
job_id = generate_job_id()
elif self._job_info_client.get_status(job_id) is not None:
raise RuntimeError(f"Job {job_id} already exists.")
logger.info(f"Starting job with job_id: {job_id}")
job_info = JobInfo(
entrypoint=entrypoint,
status=JobStatus.PENDING,
start_time=int(time.time() * 1000),
metadata=metadata,
runtime_env=runtime_env,
)
self._job_info_client.put_info(job_id, job_info)
# Wait for the actor to start up asynchronously so this call always
# returns immediately and we can catch errors with the actor starting
# up.
try:
supervisor = self._supervisor_actor_cls.options(
lifetime="detached",
name=self.JOB_ACTOR_NAME.format(job_id=job_id),
num_cpus=0,
# Currently we assume JobManager is created by dashboard server
# running on headnode, same for job supervisor actors scheduled
resources={
self._get_current_node_resource_key(): 0.001,
},
runtime_env=runtime_env,
).remote(job_id, entrypoint, metadata or {})
supervisor.run.remote(_start_signal_actor=_start_signal_actor)
# Monitor the job in the background so we can detect errors without
# requiring a client to poll.
create_task(self._monitor_job(job_id, job_supervisor=supervisor))
except Exception as e:
self._job_info_client.put_status(
job_id,
JobStatus.FAILED,
message=f"Failed to start job supervisor: {e}.",
)
return job_id
def stop_job(self, job_id) -> bool:
"""Request a job to exit, fire and forget.
Returns whether or not the job was running.
"""
job_supervisor_actor = self._get_actor_for_job(job_id)
if job_supervisor_actor is not None:
# Actor is still alive, signal it to stop the driver, fire and
# forget
job_supervisor_actor.stop.remote()
return True
else:
return False
def get_job_status(self, job_id: str) -> Optional[JobStatus]:
"""Get latest status of a job."""
return self._job_info_client.get_status(job_id)
def get_job_info(self, job_id: str) -> Optional[JobInfo]:
"""Get latest info of a job."""
return self._job_info_client.get_info(job_id)
def list_jobs(self) -> Dict[str, JobInfo]:
"""Get info for all jobs."""
return self._job_info_client.get_all_jobs()
def get_job_logs(self, job_id: str) -> str:
"""Get all logs produced by a job."""
return self._log_client.get_logs(job_id)
async def tail_job_logs(self, job_id: str) -> Iterator[str]:
"""Return an iterator following the logs of a job."""
if self.get_job_status(job_id) is None:
raise RuntimeError(f"Job '{job_id}' does not exist.")
for line in self._log_client.tail_logs(job_id):
if line is None:
# Return if the job has exited and there are no new log lines.
status = self.get_job_status(job_id)
if status not in {JobStatus.PENDING, JobStatus.RUNNING}:
return
await asyncio.sleep(self.LOG_TAIL_SLEEP_S)
else:
yield line