Robert Wojciechowski
87c0cda9f6
Retain durable reservations until provider termination or explicit recovery, protect unresolved pools and credentials, and add scoped reconciliation. Consolidate lifecycle and secret policies and split prompt/media stages with regression coverage.
376 lines
20 KiB
Python
376 lines
20 KiB
Python
"""Master-owned asynchronous Runpod jobs, independent of worker callbacks."""
|
|
import asyncio
|
|
import time
|
|
import uuid
|
|
from pathlib import Path
|
|
|
|
from .runpod_client import RunpodClient, RunpodError, AmbiguousSubmission
|
|
from .serverless_ledger import JobLedger
|
|
from .serverless_state import (
|
|
PROVIDER_STATES, PROVIDER_TERMINAL, belongs_to_pool, is_unresolved,
|
|
occupies_capacity, terminal_provider_status,
|
|
)
|
|
|
|
|
|
async def decode_results(output, manifest):
|
|
from .serverless_media import collect_images
|
|
import torch
|
|
images = await collect_images(output, manifest)
|
|
return {job_id: [torch.from_numpy(image).unsqueeze(0) for image in batch] for job_id, batch in images.items()}
|
|
|
|
|
|
def state_directory():
|
|
import folder_paths
|
|
return Path(folder_paths.get_user_directory()) / "distributed" / "serverless"
|
|
|
|
|
|
def get_secret_store():
|
|
from ...utils.serverless_config import SecretStore
|
|
return SecretStore(state_directory() / "credentials.json")
|
|
|
|
|
|
def get_manager():
|
|
import server
|
|
from ...utils.network import get_client_session
|
|
ps = server.PromptServer.instance
|
|
manager = getattr(ps, "distributed_serverless", None)
|
|
if manager is None:
|
|
manager = ServerlessManager(ps, get_secret_store(), state_directory() / "jobs.json", get_client_session)
|
|
ps.distributed_serverless = manager
|
|
return manager
|
|
|
|
|
|
class ServerlessManager:
|
|
def __init__(self, prompt_server, secrets, ledger_path, session_factory):
|
|
self.server = prompt_server
|
|
self.secrets = secrets
|
|
self.ledger = JobLedger(ledger_path)
|
|
self.session_factory = session_factory
|
|
self.tasks = {}
|
|
self.runs = {}
|
|
self.pool_limits = {}
|
|
self.health_cache = {}
|
|
self.health_locks = {}
|
|
self.poll_seconds = 2.0
|
|
self.capacity_changed = asyncio.Condition()
|
|
self.recovery_lock = asyncio.Lock()
|
|
self._recover_local_records()
|
|
|
|
def _recover_local_records(self):
|
|
# WAITING is committed before admission; it cannot have reached Runpod.
|
|
# SUBMITTING is an intent record, so a lost ID must remain reserved.
|
|
for record in self.ledger.all():
|
|
if record.get("status") == "WAITING" and not record.get("job_id"):
|
|
self.ledger.update(record["record_id"], status="CANCELLED", updated_at=time.time())
|
|
elif record.get("status") == "SUBMITTING" and not record.get("job_id"):
|
|
self.ledger.update(record["record_id"], status="UNKNOWN", updated_at=time.time())
|
|
|
|
def _task_is_active(self, record_id):
|
|
task = self.tasks.get(record_id)
|
|
return task is not None and not task.done()
|
|
|
|
def pool_is_active(self, worker_id):
|
|
"""Protect both live local work and unresolved pre-restart provider work."""
|
|
for rid, run in self.runs.items():
|
|
pool_id = run["worker"].get("pool_id", run["worker"]["id"])
|
|
if str(pool_id) == str(worker_id) and self._task_is_active(rid):
|
|
return True
|
|
return any(belongs_to_pool(record, worker_id) and is_unresolved(record)
|
|
for record in self.ledger.all())
|
|
|
|
async def _reserve_provider_capacity(self, run):
|
|
pool_id = run["worker"].get("pool_id", run["worker"]["id"])
|
|
async with self.capacity_changed:
|
|
while True:
|
|
self._check_master(run)
|
|
outstanding = sum(belongs_to_pool(record, pool_id) and occupies_capacity(record)
|
|
for record in self.ledger.all())
|
|
if outstanding < run["worker"]["concurrency"]:
|
|
# Claim atomically before any await that could submit a billed job.
|
|
self.ledger.update(run["record_id"], status="SUBMITTING", updated_at=time.time())
|
|
return
|
|
await self.capacity_changed.wait()
|
|
|
|
async def _notify_capacity(self):
|
|
async with self.capacity_changed:
|
|
self.capacity_changed.notify_all()
|
|
|
|
async def _client(self, worker):
|
|
pool_id = worker.get("pool_id", worker["id"])
|
|
key = self.secrets.get(pool_id)
|
|
return RunpodClient(await self.session_factory(), worker["endpoint_id"], key)
|
|
|
|
async def health(self, worker, *, fresh=False):
|
|
pool_id = worker.get("pool_id", worker["id"])
|
|
cache_key = (pool_id, worker["endpoint_id"])
|
|
lock = self.health_locks.setdefault(cache_key, asyncio.Lock())
|
|
async with lock:
|
|
cached = self.health_cache.get(cache_key)
|
|
if cached and not fresh and time.monotonic() - cached[0] < 10:
|
|
return cached[1]
|
|
client = await self._client(worker)
|
|
data = await client.health()
|
|
jobs = data.get("jobs", {})
|
|
if not isinstance(jobs, dict) or not isinstance(data.get("workers"), dict):
|
|
raise RunpodError("Runpod returned an invalid endpoint health record")
|
|
try:
|
|
pending = max(0, int(jobs.get("inQueue", 0))) + max(0, int(jobs.get("inProgress", 0)))
|
|
except (ValueError, TypeError):
|
|
raise RunpodError("Runpod returned invalid queue counts") from None
|
|
result = {"ok": True, "queueRemaining": pending, "message": "Queued / processing" if pending else "Ready (scales from zero)"}
|
|
self.health_cache[cache_key] = (time.monotonic(), result)
|
|
return result
|
|
|
|
def _semaphore(self, worker):
|
|
pool_id = worker.get("pool_id", worker["id"])
|
|
limit = worker["concurrency"]
|
|
current = self.pool_limits.get(pool_id)
|
|
if current is None:
|
|
current = (limit, asyncio.Semaphore(limit))
|
|
self.pool_limits[pool_id] = current
|
|
elif current[0] != limit:
|
|
if self.pool_is_active(pool_id):
|
|
raise ValueError("Wait for this serverless pool's current jobs before changing concurrency")
|
|
current = (limit, asyncio.Semaphore(limit))
|
|
self.pool_limits[pool_id] = current
|
|
return current[1]
|
|
|
|
def ensure_capacity(self, count):
|
|
if sum(not task.done() for task in self.tasks.values()) + count > 64:
|
|
raise ValueError("Distributed already has too many pending serverless jobs (limit 64)")
|
|
|
|
async def start(self, execution_id, prompt_id, prepared_jobs):
|
|
self.ensure_capacity(len(prepared_jobs))
|
|
new_runs = []
|
|
# Journal every local slot before starting any billed submission.
|
|
for prepared in prepared_jobs:
|
|
worker = prepared["worker"]
|
|
semaphore = self._semaphore(worker)
|
|
record_id = uuid.uuid4().hex
|
|
now = time.time()
|
|
self.ledger.update(record_id, execution_id=execution_id, prompt_id=prompt_id,
|
|
worker_id=worker.get("pool_id", worker["id"]), participant_id=worker["id"],
|
|
endpoint_id=worker["endpoint_id"], job_id=None, status="WAITING", created_at=now, updated_at=now)
|
|
run = dict(prepared, execution_id=execution_id, prompt_id=prompt_id, record_id=record_id,
|
|
provider_job_id=None, client=None, deadline=time.monotonic() + worker["timeout_seconds"])
|
|
self.runs[record_id] = run
|
|
new_runs.append((record_id, run, semaphore))
|
|
for record_id, run, semaphore in new_runs:
|
|
self.tasks[record_id] = asyncio.create_task(self._run(run, semaphore), name=f"distributed-serverless-{record_id}")
|
|
# Bound in-memory completed tracking as well as the on-disk journal.
|
|
done = [rid for rid, task in self.tasks.items() if task.done()]
|
|
for rid in done[:-200]:
|
|
self.tasks.pop(rid, None)
|
|
self.runs.pop(rid, None)
|
|
|
|
def is_pending(self, collector_job_id, participant_id):
|
|
return any(not task.done() and run["worker"]["id"] == participant_id and
|
|
any(m["job_id"] == collector_job_id for m in run["manifest"])
|
|
for rid, task in self.tasks.items() for run in [self.runs[rid]])
|
|
|
|
def owns(self, collector_job_id, participant_id):
|
|
return any(run["worker"]["id"] == participant_id and
|
|
any(m["job_id"] == collector_job_id for m in run["manifest"]) for run in self.runs.values())
|
|
|
|
async def _deliver(self, run, images=None, error=None):
|
|
async with self.server.distributed_jobs_lock:
|
|
for binding in run["manifest"]:
|
|
queue = self.server.distributed_pending_jobs.get(binding["job_id"])
|
|
if queue is None:
|
|
continue # A cancelled/finished collector must never be resurrected.
|
|
if error:
|
|
queue.put_nowait({"worker_id": run["worker"]["id"], "serverless_error": error, "is_last": True})
|
|
else:
|
|
batch = images[binding["job_id"]]
|
|
for index, tensor in enumerate(batch):
|
|
queue.put_nowait({"worker_id": run["worker"]["id"], "image_index": index,
|
|
"tensor": tensor, "is_last": index == len(batch) - 1})
|
|
|
|
async def _cancel_provider(self, run):
|
|
if run["provider_job_id"] and run["client"]:
|
|
try:
|
|
result = await run["client"].cancel(run["provider_job_id"])
|
|
return terminal_provider_status(result)
|
|
except Exception:
|
|
return None
|
|
return None
|
|
|
|
def _check_master(self, run):
|
|
queue = getattr(self.server, "prompt_queue", None)
|
|
if queue is None:
|
|
return
|
|
history = queue.get_history(run["prompt_id"])
|
|
record = history.get(run["prompt_id"], {})
|
|
if record.get("status", {}).get("status_str") == "error":
|
|
raise RunpodError("Master workflow failed; serverless work stopped")
|
|
if not record and hasattr(queue, "get_current_queue"):
|
|
running, pending = queue.get_current_queue()
|
|
if not any(str(item[1]) == run["prompt_id"] for item in [*running, *pending]):
|
|
raise RunpodError("Master workflow was removed from the queue; serverless work stopped")
|
|
|
|
async def _run(self, run, semaphore):
|
|
rid = run["record_id"]
|
|
submitted = False
|
|
provider_terminal = False
|
|
try:
|
|
async with asyncio.timeout(max(0, run["deadline"] - time.monotonic())):
|
|
async with semaphore:
|
|
await self._reserve_provider_capacity(run)
|
|
client = await self._client(run["worker"])
|
|
run["client"] = client
|
|
submitted = True
|
|
job_id = await client.submit(run["workflow"], run["images"], run["worker"]["timeout_seconds"])
|
|
run["provider_job_id"] = job_id
|
|
# Persist immediately: all subsequent actions use this exact ID.
|
|
self.ledger.update(rid, job_id=job_id, status="IN_QUEUE", updated_at=time.time())
|
|
while True:
|
|
self._check_master(run)
|
|
response = await client.status(job_id)
|
|
status = response.get("status")
|
|
if not isinstance(status, str) or status not in PROVIDER_STATES:
|
|
raise RunpodError("Runpod returned an unknown job state")
|
|
provider_terminal = status in PROVIDER_TERMINAL
|
|
if status == "COMPLETED":
|
|
output = response.get("output")
|
|
if not isinstance(output, dict) or output.get("error") or output.get("errors"):
|
|
raise RunpodError("Serverless workflow reported an output error; inspect the provider job")
|
|
self.ledger.update(rid, status="DOWNLOADING", updated_at=time.time())
|
|
images = await decode_results(output, run["manifest"])
|
|
await self._deliver(run, images=images)
|
|
self.ledger.update(rid, status="COMPLETED", updated_at=time.time())
|
|
return
|
|
self.ledger.update(rid, status=status, updated_at=time.time())
|
|
if provider_terminal:
|
|
raise RunpodError(f"Serverless job ended with {status}")
|
|
await asyncio.sleep(self.poll_seconds)
|
|
except AmbiguousSubmission as exc:
|
|
self.ledger.update(rid, status="UNKNOWN", error=str(exc), updated_at=time.time())
|
|
await self._deliver(run, error=str(exc))
|
|
except (asyncio.TimeoutError, asyncio.CancelledError) as exc:
|
|
terminal = "FAILED" if provider_terminal else await self._cancel_provider(run)
|
|
if submitted and not run["provider_job_id"]:
|
|
status = "UNKNOWN"
|
|
error = "Submission interrupted before its job ID was received; check Runpod before retrying"
|
|
elif isinstance(exc, asyncio.TimeoutError):
|
|
status = terminal or ("CANCEL_REQUESTED" if run["provider_job_id"] else "TIMED_OUT")
|
|
error = "Serverless job exceeded its queue/execution deadline"
|
|
else:
|
|
status = terminal or ("CANCEL_REQUESTED" if run["provider_job_id"] else "CANCELLED")
|
|
error = "Serverless work stopped" if terminal or not submitted else "Serverless cancellation requested; provider confirmation is pending"
|
|
self.ledger.update(rid, status=status, error=error, updated_at=time.time())
|
|
await self._deliver(run, error=error)
|
|
except Exception as exc:
|
|
# If polling/decoding failed, stop only our own provider job, never the pool.
|
|
terminal = "FAILED" if provider_terminal else await self._cancel_provider(run)
|
|
error = str(exc) if isinstance(exc, RunpodError) else "Serverless processing failed; check the workflow and endpoint contract"
|
|
status = terminal or ("CANCEL_REQUESTED" if run["provider_job_id"] else "FAILED")
|
|
self.ledger.update(rid, status=status, error=error, updated_at=time.time())
|
|
await self._deliver(run, error=error)
|
|
finally:
|
|
# Completed tracking retains IDs/bindings only, not potentially large
|
|
# base64 inputs, workflow metadata, or a client holding credentials.
|
|
run.pop("workflow", None)
|
|
run.pop("images", None)
|
|
run["client"] = None
|
|
# A released local semaphore alone cannot release a durable reservation.
|
|
await self._notify_capacity()
|
|
|
|
async def wait_for_execution(self, execution_id):
|
|
tasks = [task for rid, task in self.tasks.items() if self.runs[rid]["execution_id"] == execution_id]
|
|
await asyncio.gather(*tasks, return_exceptions=False)
|
|
|
|
def _detached_records(self, worker_id=None, execution_id=None):
|
|
return [record for record in self.ledger.all()
|
|
if is_unresolved(record) and not self._task_is_active(record["record_id"])
|
|
and (worker_id is None or belongs_to_pool(record, worker_id))
|
|
and (execution_id is None or record.get("execution_id") == execution_id)]
|
|
|
|
async def _recover_provider_record(self, record, *, cancel=False):
|
|
"""Reconcile an exact saved ID, never submit or resume a workflow."""
|
|
if not record.get("job_id"):
|
|
return False
|
|
status = None
|
|
try:
|
|
client = await self._client({"id": record["worker_id"], "endpoint_id": record["endpoint_id"]})
|
|
response = await (client.cancel(record["job_id"]) if cancel else client.status(record["job_id"]))
|
|
status = terminal_provider_status(response)
|
|
except Exception:
|
|
pass # Unreachable/expired is not proof of provider termination.
|
|
if status is not None or cancel:
|
|
self.ledger.update(record["record_id"], status=status or "CANCEL_REQUESTED", updated_at=time.time())
|
|
return status is not None
|
|
|
|
async def reconcile(self, worker_id=None):
|
|
"""Read terminal states for detached jobs and release only proven capacity."""
|
|
resolved = 0
|
|
async with self.recovery_lock:
|
|
for record in self._detached_records(worker_id):
|
|
resolved += await self._recover_provider_record(record)
|
|
await self._notify_capacity()
|
|
return resolved
|
|
|
|
async def acknowledge_recovery(self, worker_id, record_id, *, provider_checked=False):
|
|
"""Explicit operator escape hatch for lost IDs/expired provider records.
|
|
|
|
This records a human's verification, not a cancellation or provider
|
|
status. It is never called automatically or allowed for live tasks.
|
|
"""
|
|
if provider_checked is not True:
|
|
raise ValueError("Check Runpod and confirm that this job is no longer running before resolving it")
|
|
async with self.recovery_lock:
|
|
record = next((r for r in self._detached_records(worker_id) if r["record_id"] == record_id), None)
|
|
if record is None:
|
|
raise ValueError("Unresolved inactive job not found for this pool")
|
|
self.ledger.update(record_id, status="RESOLVED", updated_at=time.time(),
|
|
error="Recovery explicitly acknowledged after operator checked Runpod; not provider-confirmed")
|
|
await self._notify_capacity()
|
|
|
|
async def cancel(self, worker_id=None, execution_id=None, collector_job_id=None):
|
|
targets = {}
|
|
for rid, task in self.tasks.items():
|
|
run = self.runs[rid]
|
|
if worker_id and run["worker"].get("pool_id", run["worker"]["id"]) != worker_id:
|
|
continue
|
|
if execution_id and run["execution_id"] != execution_id:
|
|
continue
|
|
if collector_job_id and not any(m["job_id"] == collector_job_id for m in run["manifest"]):
|
|
continue
|
|
if not task.done():
|
|
# Repeated requests must not interrupt the first cancellation's cleanup.
|
|
if not task.cancelling():
|
|
task.cancel()
|
|
targets[rid] = task
|
|
if targets:
|
|
await asyncio.gather(*targets.values(), return_exceptions=True)
|
|
recovered = 0
|
|
async with self.recovery_lock:
|
|
for record in self._detached_records(worker_id, execution_id):
|
|
rid = record["record_id"]
|
|
# A task cancelled before its first step never enters _run's handlers.
|
|
# A failed start() may also leave a journalled but unstarted local slot.
|
|
if record.get("status") == "WAITING" and not record.get("job_id"):
|
|
run = self.runs.get(rid)
|
|
if run is not None and (collector_job_id is None or any(m["job_id"] == collector_job_id for m in run["manifest"])):
|
|
self.ledger.update(rid, status="CANCELLED", updated_at=time.time())
|
|
run.pop("workflow", None)
|
|
run.pop("images", None)
|
|
await self._deliver(run, error="Serverless job cancelled before submission")
|
|
continue
|
|
# Do not send two cancellation requests for the same live job in
|
|
# one action. Its unresolved reservation survives for later recovery.
|
|
if rid in targets or collector_job_id or not record.get("job_id"):
|
|
continue
|
|
await self._recover_provider_record(record, cancel=True)
|
|
recovered += 1
|
|
await self._notify_capacity()
|
|
return len(targets) + recovered
|
|
|
|
def public_jobs(self, worker_id):
|
|
records = [record for record in reversed(self.ledger.all()) if belongs_to_pool(record, worker_id)]
|
|
# Never hide unresolved IDs behind the bounded completed history.
|
|
visible = [record for record in records if is_unresolved(record)]
|
|
visible.extend(record for record in records if not is_unresolved(record))
|
|
unresolved_count = sum(is_unresolved(record) for record in records)
|
|
return [{key: record.get(key) for key in ("record_id", "job_id", "status", "worker_id", "execution_id", "error")}
|
|
for record in visible[:max(20, unresolved_count)]]
|