stroblmeandClaude Opus 5
058f16ec1d
Each was a loose end recorded under `### SDK` in the notepad. `serve` takes its own pidfile down on SIGTERM. uvicorn restores the handler it found and re-raises the signal it stopped on, so the default handler ended the process without unwinding and the `finally` never ran — which is what a stop sends, and what left `serve.pid` behind. `serve.log` is cut back past 5 MB by the engine rather than by the screen that started it, so an adopted engine is bounded too. Gated on its own stdout being an appended regular file, which is what makes the cut safe: the kernel then puts the next write at the new end. Cards are counted from `/dev/nvidia[0-9]*`, so `FLOW_GPUS`/`--gpus` of 0 means "work it out" the way `FLOW_CPUS` always has. The engine counts, not the accountant — a remote worker builds one of those from its own inventory, and detecting there would hand it the engine host's cards. The worker counts last: what a batch job says it was granted still wins. `GET /runs/metrics/names` is the distinct over a selection that `--list` and the terminal's metric picker were approximating by reading the newest run that had measured anything, which missed a name only an older run ever wrote. `MetricSink` announces each batch it has written (`run_metric`, carrying the names). Not a per-point event: one covers up to 500 points or two seconds of them, and the rows stay the record. The terminal comparison fills in as the first readings land instead of staying blank until reopened, and the browser refetches the run and any comparison rather than the list behind them. `retry --group` pages the list route by `before` instead of stopping at 500. The terminal dashboard takes the terminal's colours (`ansi-dark`), and the web UI can re-pair from Settings without disconnecting first. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PRQ9bmTvCbqCwXo9mxZzzV
490 lines
20 KiB
Python
490 lines
20 KiB
Python
import asyncio
|
|
import contextlib
|
|
import hashlib
|
|
import inspect
|
|
import logging
|
|
import tempfile
|
|
from collections.abc import AsyncIterator, Callable, Coroutine
|
|
from contextlib import AbstractAsyncContextManager, asynccontextmanager
|
|
from pathlib import Path
|
|
from typing import Any
|
|
|
|
from fastapi import FastAPI, Request
|
|
from fastapi.concurrency import run_in_threadpool
|
|
from fastapi.responses import JSONResponse
|
|
from fastapi.routing import APIRoute
|
|
from fluksio_worker.worker_main import ARTIFACT_DIR_ENV, ARTIFACT_VOLATILE_DIR_ENV
|
|
from starlette.middleware.cors import CORSMiddleware
|
|
|
|
from fluksio import __version__
|
|
from fluksio.api.main import api_router
|
|
from fluksio.api.routes.alerts import read_config as read_alerts_config
|
|
from fluksio.cloud import config as cloud_config
|
|
from fluksio.core import security
|
|
from fluksio.core.config import settings
|
|
from fluksio.core.db import engine as db_engine
|
|
from fluksio.core.db import prepare
|
|
from fluksio.flow import logs, modules
|
|
from fluksio.flow.alerts import AlertManager
|
|
from fluksio.flow.artifacts import ArtifactStore, VolatileStore
|
|
from fluksio.flow.controller import FlowController, RebuildBusy
|
|
from fluksio.flow.dashboards import DashboardStore
|
|
from fluksio.flow.events import event_bus
|
|
from fluksio.flow.executor import ExecutionService
|
|
from fluksio.flow.metrics import MetricsCollector
|
|
from fluksio.flow.nodes.http import close_shared_client
|
|
from fluksio.flow.pipeline import ValueSource
|
|
from fluksio.flow.placement import Placer
|
|
from fluksio.flow.plugins import load_plugins
|
|
from fluksio.flow.provision import load_provisioners
|
|
from fluksio.flow.queue import MemoryWorkQueue, RedisWorkQueue, WorkQueue
|
|
from fluksio.flow.remote import RemoteWorkerHub
|
|
from fluksio.flow.resources import (
|
|
ResourceAccountant,
|
|
fair_share_env,
|
|
machine_gpus,
|
|
)
|
|
from fluksio.flow.runs import RUN_STATE_TTL, RunService, sweep_artifacts
|
|
from fluksio.flow.secrets import init_secrets
|
|
from fluksio.flow.state import MemoryState, RedisState, StateBackend
|
|
from fluksio.flow.store import FlowStore
|
|
from fluksio.flow.watchdog import LoopWatchdog
|
|
from fluksio.flow.workers import PythonWorkerPool
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
def custom_generate_unique_id(route: APIRoute) -> str:
|
|
return f"{route.tags[0]}-{route.name}"
|
|
|
|
|
|
if settings.SENTRY_DSN and settings.ENVIRONMENT != "local":
|
|
# Imported here rather than at the top: it is a `fluksio[server]` extra, so
|
|
# a pip install without one has no sentry to import — and no DSN either.
|
|
import sentry_sdk
|
|
|
|
# `enable_tracing` was removed in sentry-sdk 2.x; this is what it meant.
|
|
sentry_sdk.init(dsn=str(settings.SENTRY_DSN), traces_sample_rate=1.0)
|
|
|
|
|
|
def _state_backend() -> StateBackend:
|
|
if settings.REDIS_HOST:
|
|
return RedisState(host=settings.REDIS_HOST, port=settings.REDIS_PORT)
|
|
return MemoryState()
|
|
|
|
|
|
def _work_queue(namespace: str = "queue") -> WorkQueue:
|
|
"""Redis makes queued work survive the process; memory does not pretend to."""
|
|
if settings.REDIS_HOST:
|
|
return RedisWorkQueue(
|
|
host=settings.REDIS_HOST, port=settings.REDIS_PORT, namespace=namespace
|
|
)
|
|
return MemoryWorkQueue()
|
|
|
|
|
|
def _run_state(namespace: str) -> StateBackend:
|
|
"""A state backend of a run's own, which is what isolates it.
|
|
|
|
It expires: a finished run's messages are read out into its result, and
|
|
what is left is only worth keeping while someone might look at it.
|
|
"""
|
|
if settings.REDIS_HOST:
|
|
return RedisState(
|
|
host=settings.REDIS_HOST,
|
|
port=settings.REDIS_PORT,
|
|
namespace=namespace,
|
|
ttl=RUN_STATE_TTL,
|
|
)
|
|
return MemoryState()
|
|
|
|
|
|
async def _sweep_artifacts(store: ArtifactStore, controller: FlowController) -> None:
|
|
"""Take unreferenced artifact bytes off the disk, on a slow loop.
|
|
|
|
A flow streaming media writes one artifact per frame, so a store nothing
|
|
prunes only grows. Runs in a thread: it walks a directory and reads state.
|
|
"""
|
|
interval = settings.ARTIFACT_GC_INTERVAL_S
|
|
if interval <= 0:
|
|
return
|
|
while True:
|
|
await asyncio.sleep(interval)
|
|
try:
|
|
await run_in_threadpool(
|
|
sweep_artifacts,
|
|
store,
|
|
controller.state,
|
|
settings.ARTIFACT_GC_GRACE_S,
|
|
)
|
|
except Exception:
|
|
logger.exception("Artifact sweep failed")
|
|
|
|
|
|
#: How often the ring is measured. Seconds rather than the sweep's hour: it is
|
|
#: bounded by size and the frames arriving are what push the old ones out.
|
|
VOLATILE_TRIM_S = 5.0
|
|
|
|
|
|
async def _trim_volatile(ring: VolatileStore) -> None:
|
|
"""Hold the volatile ring inside its bound.
|
|
|
|
``put`` trims what it wrote, but a worker in this container writes to the
|
|
ring itself and the engine never sees that one — so the bound needs
|
|
something of its own watching it.
|
|
"""
|
|
if ring.limit_bytes <= 0:
|
|
return
|
|
while True:
|
|
await asyncio.sleep(VOLATILE_TRIM_S)
|
|
try:
|
|
await run_in_threadpool(ring.trim)
|
|
except Exception:
|
|
logger.exception("Could not trim the volatile artifact ring")
|
|
|
|
|
|
def _volatile_root() -> Path:
|
|
"""Where the ring goes: memory if this machine has some to lend.
|
|
|
|
Named for the data directory rather than fixed, so two engines on one host
|
|
have a ring each instead of quietly evicting each other's frames.
|
|
"""
|
|
configured = settings.ARTIFACT_VOLATILE_DIR
|
|
if configured is not None:
|
|
return configured
|
|
tag = hashlib.sha256(str(settings.DATA_DIR.resolve()).encode()).hexdigest()[:8]
|
|
shm = Path("/dev/shm")
|
|
parent = shm if shm.is_dir() else Path(tempfile.gettempdir())
|
|
return parent / f"fluksio-volatile-{tag}"
|
|
|
|
|
|
def _mcp_sessions() -> AbstractAsyncContextManager[None]:
|
|
"""The MCP session manager's run scope, or nothing when MCP is off."""
|
|
if not settings.MCP_ENABLED:
|
|
return contextlib.nullcontext()
|
|
from fluksio.mcp.server import mcp as mcp_server
|
|
|
|
return mcp_server.session_manager.run()
|
|
|
|
|
|
@asynccontextmanager
|
|
async def lifespan(app: FastAPI) -> AsyncIterator[None]:
|
|
"""Start the flow engine alongside the API."""
|
|
# The schema and the first superuser, before anything reads either. It is
|
|
# idempotent, so a deployment that ran this from its own prestart step
|
|
# pays a version check for it and nothing else.
|
|
await run_in_threadpool(prepare, db_engine)
|
|
event_bus.bind(asyncio.get_running_loop())
|
|
|
|
# Everything acquired below registers how to release it, and the `finally`
|
|
# at the end walks the list backwards. It used to be a fixed block of
|
|
# shutdown steps *after* every acquisition — so a failure part-way through
|
|
# startup reached none of them and left the worker pool's subprocesses and
|
|
# every background task behind. Under `--reload` that is once per bad edit.
|
|
started: list[Callable[[], Any]] = []
|
|
|
|
async def _release() -> None:
|
|
for close in reversed(started):
|
|
try:
|
|
result = close()
|
|
if inspect.isawaitable(result):
|
|
await result
|
|
except Exception:
|
|
logger.exception("Error while shutting down")
|
|
|
|
def _background(
|
|
coro: Coroutine[Any, Any, None], name: str, once: bool = False
|
|
) -> asyncio.Task[None]:
|
|
"""Start a background task that says something if it ever stops.
|
|
|
|
The loops catch their own exceptions *inside* the loop, so one raised
|
|
anywhere else simply ended the task — an engine that went on serving
|
|
with no metrics, no alerts or no artifact sweep and nothing anywhere
|
|
saying so. ``once`` is for a task that is meant to finish.
|
|
"""
|
|
task = asyncio.create_task(coro, name=name)
|
|
|
|
def _finished(done: asyncio.Task[None]) -> None:
|
|
if done.cancelled():
|
|
return
|
|
exc = done.exception()
|
|
if exc is None:
|
|
if not once:
|
|
logger.warning("Background task '%s' stopped on its own", name)
|
|
return
|
|
logger.error("Background task '%s' died: %s", name, exc, exc_info=exc)
|
|
event_bus.publish(
|
|
{"type": "engine_degraded", "detail": f"{name} stopped: {exc}"}
|
|
)
|
|
|
|
task.add_done_callback(_finished)
|
|
started.append(task.cancel)
|
|
return task
|
|
|
|
# The startup itself is inside the try, so a step that raises part-way
|
|
# through still reaches the release below.
|
|
try:
|
|
# Node code is user code, and `print` is how it says things.
|
|
logs.install()
|
|
init_secrets(settings.SECRETS_FILE, settings.SECRET_KEY)
|
|
# Connectors register their node types before any flow is built with them.
|
|
load_plugins()
|
|
|
|
alerts = AlertManager(event_bus, config=read_alerts_config())
|
|
execution = ExecutionService(
|
|
queue=_work_queue(),
|
|
max_workers=settings.FLOW_MAX_WORKERS,
|
|
events=event_bus,
|
|
max_cascades=settings.FLOW_MAX_CASCADES,
|
|
)
|
|
store = FlowStore(settings.FLOWS_DIR)
|
|
# The packages node code imports. In the background, because a first
|
|
# install of anything substantial takes minutes and `uv` is given five
|
|
# of them twice over — held here, the container never answers its
|
|
# healthcheck, autoheal restarts it, and it never finishes installing at
|
|
# all. A node whose import is not there yet fails and is retried, which
|
|
# is recoverable; a restart loop is not.
|
|
_background(
|
|
run_in_threadpool(modules.reconcile, store), "module-reconcile", once=True
|
|
)
|
|
# Beside the flows rather than in them: an artifact is what a run produced,
|
|
# not something anyone wrote, so it has no business in the git repository.
|
|
artifacts = ArtifactStore(settings.FLOWS_DIR.parent / "artifacts")
|
|
# Frames a flow only shows live, held in memory and bounded by size.
|
|
# The store falls through to it, so a volatile reference is an
|
|
# ordinary one everywhere but in how long its bytes last.
|
|
volatile = VolatileStore(_volatile_root(), settings.ARTIFACT_VOLATILE_BYTES)
|
|
artifacts.volatile = volatile
|
|
app.state.artifact_store = artifacts
|
|
# Zero means "work it out" for both, as it always has for cores.
|
|
accountant = ResourceAccountant(
|
|
cpus=settings.FLOW_CPUS, gpus=settings.FLOW_GPUS or machine_gpus()
|
|
)
|
|
app.state.resources = accountant
|
|
# Every machine a node could run on: this one, and whatever attaches.
|
|
placer = Placer(local=accountant, events=event_bus)
|
|
# Where more machines can be asked for when nothing attached will do.
|
|
placer.provisioners = load_provisioners(settings.PROVISIONERS_FILE, event_bus)
|
|
app.state.placer = placer
|
|
pool = PythonWorkerPool(
|
|
python=modules.venv_python(),
|
|
size=settings.FLOW_MAX_WORKERS,
|
|
events=event_bus,
|
|
# A worker in this container writes to the store directly; a remote one
|
|
# is given a URL instead. Node code calls the same two functions.
|
|
env={
|
|
ARTIFACT_DIR_ENV: str(artifacts.root),
|
|
ARTIFACT_VOLATILE_DIR_ENV: str(volatile.root),
|
|
# Every slot can be busy at once, so a worker left to size its own
|
|
# thread pool to the machine means as many processes as there are
|
|
# slots, each believing it has the whole of it. A node that says
|
|
# what it needs overrides this; one that says nothing gets a share.
|
|
**fair_share_env(accountant.cpus, settings.FLOW_MAX_WORKERS),
|
|
},
|
|
)
|
|
pool.start()
|
|
# On a thread: stopping a worker waits up to five seconds on each child,
|
|
# and with a pool per declared environment that is a shutdown the event
|
|
# loop should not be holding.
|
|
started.append(lambda: run_in_threadpool(pool.stop))
|
|
app.state.worker_pool = pool
|
|
# Assigned rather than passed both ways: the hub tells the placer when a
|
|
# machine comes or goes, and the placer needs the hub to know what is there.
|
|
worker_hub = RemoteWorkerHub(on_change=placer.wake)
|
|
placer.hub = worker_hub
|
|
app.state.worker_hub = worker_hub
|
|
controller = FlowController(
|
|
store=store,
|
|
state=_state_backend(),
|
|
events=event_bus,
|
|
max_workers=settings.FLOW_MAX_WORKERS,
|
|
fastapi_app=app,
|
|
execution=execution,
|
|
alerts=alerts,
|
|
workers=pool,
|
|
remote=worker_hub,
|
|
resources=accountant,
|
|
placer=placer,
|
|
)
|
|
app.state.flow_controller = controller
|
|
# A "dashboard" alert channel puts its alert into the graph. Bound here
|
|
# rather than passed in: the manager is built before the controller is.
|
|
alerts.publish = lambda name, value: controller.publish_message(
|
|
name, value, ValueSource(kind="api", id="alerts", label="Alerts")
|
|
)
|
|
dashboards = DashboardStore(controller.store)
|
|
app.state.dashboard_store = dashboards
|
|
controller.dashboards = dashboards
|
|
# Charts need a deeper series than the default; tell the engine
|
|
# before it starts recording.
|
|
controller.set_history_limits(dashboards.history_requirements())
|
|
# Runs read from a stream of their own: a burst of sweep runs must not
|
|
# stand between the automations and their work, and a run that takes an
|
|
# hour must not be judged by the cascade reaper's timings.
|
|
run_service = RunService(
|
|
controller=controller,
|
|
queue=_work_queue("run"),
|
|
state_factory=_run_state,
|
|
artifacts=artifacts,
|
|
parallel=settings.FLOW_MAX_RUNS,
|
|
)
|
|
app.state.run_service = run_service
|
|
# Said out loud because it is the only way to tell that a settings file was
|
|
# read at all — the numbers are what someone raising them is looking for.
|
|
logger.info(
|
|
"Engine limits: workers=%s cascades=%s runs=%s",
|
|
settings.FLOW_MAX_WORKERS,
|
|
execution.max_cascades,
|
|
run_service.parallel,
|
|
)
|
|
watchdog = LoopWatchdog(event_bus)
|
|
app.state.watchdog = watchdog
|
|
|
|
_background(watchdog.run(), "loop-watchdog")
|
|
_background(alerts.run(), "alert-manager")
|
|
_background(MetricsCollector(event_bus).run(), "metrics-collector")
|
|
_background(_sweep_artifacts(artifacts, controller), "artifact-gc")
|
|
_background(_trim_volatile(volatile), "artifact-ring")
|
|
await controller.start()
|
|
started.append(controller.stop)
|
|
run_service.start()
|
|
started.append(lambda: run_in_threadpool(run_service.stop))
|
|
# A machine asked for and not yet arrived would hold an allocation nobody
|
|
# is going to use.
|
|
started.append(
|
|
lambda: asyncio.gather(
|
|
*(
|
|
run_in_threadpool(provisioner.shutdown)
|
|
for provisioner in placer.provisioners
|
|
)
|
|
)
|
|
)
|
|
started.append(lambda: run_in_threadpool(close_shared_client))
|
|
# Optional, and off unless someone enrolled this instance: the
|
|
# connector dials the portal, nothing dials in.
|
|
cloud_task: asyncio.Task[None] | None = None
|
|
app.state.cloud_connector = None
|
|
app.state.cloud_task = None
|
|
from fluksio.cloud import connector as cloud_connector
|
|
|
|
if cloud_config.exists():
|
|
cloud_connector.start(app)
|
|
cloud_task = app.state.cloud_task
|
|
|
|
def _stop_cloud() -> None:
|
|
# Re-read from app.state: enrolling at runtime replaces this.
|
|
running = getattr(app.state, "cloud_task", None) or cloud_task
|
|
if running is not None:
|
|
running.cancel()
|
|
|
|
started.append(_stop_cloud)
|
|
# Watched whether or not one exists now: enrolling from the CLI writes the
|
|
# config from another process entirely, and an engine already serving
|
|
# should pick it up rather than need restarting.
|
|
_background(cloud_connector.watch_enrolment(app), "cloud-enrolment-watch")
|
|
# A mounted sub-app gets no lifespan of its own, so the MCP session
|
|
# manager is entered here; without it every /mcp request fails.
|
|
async with _mcp_sessions():
|
|
yield
|
|
except Exception:
|
|
# A startup that fails part-way lands here too, because everything
|
|
# since `event_bus.bind` registered its own release.
|
|
logger.exception("The engine could not start")
|
|
raise
|
|
finally:
|
|
await _release()
|
|
if settings.MCP_ENABLED:
|
|
from fluksio.mcp.http import aclose
|
|
|
|
await aclose()
|
|
|
|
|
|
# The schema enumerates every endpoint this instance serves, including the
|
|
# paths trigger nodes mount at runtime. That is exactly what a developer wants
|
|
# and exactly what an internet-facing deployment should not hand out, so it
|
|
# follows the environment — the same rule the portal's backend uses. The
|
|
# generated client is built from a local run, not from the deployed host.
|
|
_docs_enabled = settings.ENVIRONMENT != "production"
|
|
|
|
app = FastAPI(
|
|
title=settings.PROJECT_NAME,
|
|
version=__version__,
|
|
openapi_url=f"{settings.API_V1_STR}/openapi.json" if _docs_enabled else None,
|
|
docs_url="/docs" if _docs_enabled else None,
|
|
redoc_url="/redoc" if _docs_enabled else None,
|
|
generate_unique_id_function=custom_generate_unique_id,
|
|
lifespan=lifespan,
|
|
)
|
|
|
|
# Set all CORS enabled origins
|
|
if settings.all_cors_origins:
|
|
app.add_middleware(
|
|
CORSMiddleware,
|
|
allow_origins=settings.all_cors_origins,
|
|
allow_credentials=True,
|
|
allow_methods=["*"],
|
|
allow_headers=["*"],
|
|
)
|
|
|
|
app.include_router(api_router, prefix=settings.API_V1_STR)
|
|
|
|
|
|
@app.exception_handler(RebuildBusy)
|
|
async def rebuild_busy(request: Request, exc: RebuildBusy) -> JSONResponse: # noqa: ARG001
|
|
"""Every route that deploys something answers a wedged rebuild the same way.
|
|
|
|
The request was fine and retrying it may well work, so this is the engine
|
|
saying it is busy rather than the request having gone wrong.
|
|
"""
|
|
return JSONResponse(status_code=503, content={"detail": str(exc)})
|
|
|
|
|
|
# Tagged because the operation-id builder reads the first tag; the route
|
|
# itself stays out of the schema.
|
|
@app.get(
|
|
"/.well-known/oauth-authorization-server",
|
|
include_in_schema=False,
|
|
tags=["oauth"],
|
|
)
|
|
def oauth_authorization_server() -> JSONResponse:
|
|
"""RFC 8414 metadata, so an agent can find its way in unaided.
|
|
|
|
The authorization endpoint is the dashboard rather than the API: approving
|
|
a client needs a signed-in human, and the browser session lives there.
|
|
"""
|
|
issuer = settings.oauth_issuer
|
|
return JSONResponse(
|
|
content={
|
|
"issuer": issuer,
|
|
"authorization_endpoint": (
|
|
f"{settings.FRONTEND_HOST.rstrip('/')}/oauth/authorize"
|
|
),
|
|
"token_endpoint": f"{issuer}{settings.API_V1_STR}/oauth/token",
|
|
"registration_endpoint": f"{issuer}{settings.API_V1_STR}/oauth/register",
|
|
"jwks_uri": f"{issuer}/.well-known/jwks.json",
|
|
"response_types_supported": ["code"],
|
|
"grant_types_supported": ["authorization_code", "refresh_token"],
|
|
"code_challenge_methods_supported": ["S256"],
|
|
"token_endpoint_auth_methods_supported": ["none"],
|
|
"scopes_supported": [security.MCP_SCOPE],
|
|
},
|
|
headers={"Cache-Control": "public, max-age=3600"},
|
|
)
|
|
|
|
|
|
@app.get("/.well-known/jwks.json", include_in_schema=False, tags=["oauth"])
|
|
def jwks() -> JSONResponse:
|
|
"""The public half of the MCP signing key."""
|
|
return JSONResponse(
|
|
content=security.public_jwks(),
|
|
headers={"Cache-Control": "public, max-age=3600"},
|
|
)
|
|
|
|
|
|
# Mounted last, and at the root: the SDK serves both /mcp and the protected
|
|
# resource metadata that has to sit beside it, so mounting under /mcp would put
|
|
# that metadata somewhere no client looks for it.
|
|
if settings.MCP_ENABLED:
|
|
from fluksio.mcp.http import build_http_app
|
|
|
|
app.mount("/", build_http_app(app))
|