feat(scada): add calibrated monitoring and incident recovery
This commit is contained in:
@@ -0,0 +1,267 @@
|
||||
"""Keep the listening socket alive while replacement ASGI workers become ready."""
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import logging
|
||||
import multiprocessing
|
||||
import signal
|
||||
import threading
|
||||
import time
|
||||
from dataclasses import dataclass
|
||||
from functools import partial
|
||||
from multiprocessing.connection import Connection
|
||||
from multiprocessing.process import BaseProcess
|
||||
from pathlib import Path
|
||||
|
||||
import uvicorn
|
||||
from uvicorn._subprocess import get_subprocess
|
||||
from uvicorn.importer import import_from_string
|
||||
|
||||
|
||||
logger = logging.getLogger("uvicorn.error")
|
||||
|
||||
|
||||
class _ReadinessGate:
|
||||
def __init__(self, app, check):
|
||||
self.app, self.check = app, check
|
||||
|
||||
async def __call__(self, scope, receive, send):
|
||||
if scope["type"] != "lifespan":
|
||||
return await self.app(scope, receive, send)
|
||||
|
||||
async def checked_send(message):
|
||||
if message["type"] == "lifespan.startup.complete" and not await self.check():
|
||||
await send({"type": "lifespan.startup.failed", "message": "API readiness check failed"})
|
||||
else:
|
||||
await send(message)
|
||||
|
||||
await self.app(scope, receive, checked_send)
|
||||
|
||||
|
||||
class _ReadyServer(uvicorn.Server):
|
||||
def __init__(self, config: uvicorn.Config, control: Connection):
|
||||
super().__init__(config)
|
||||
self.control = control
|
||||
|
||||
async def startup(self, sockets=None) -> None:
|
||||
await super().startup(sockets=sockets)
|
||||
if self.started and not self.should_exit:
|
||||
self.control.send("ready")
|
||||
|
||||
async def shutdown(self, sockets=None) -> None:
|
||||
# Acknowledge that the old worker has actually stopped accepting. Its
|
||||
# established streams may still run for minutes after this handshake.
|
||||
for listener in self.servers:
|
||||
listener.close()
|
||||
for sock in sockets or []:
|
||||
sock.close()
|
||||
self.control.send("draining")
|
||||
await super().shutdown(sockets=sockets)
|
||||
|
||||
|
||||
def _serve_worker(*, config: uvicorn.Config, control: Connection, readiness_check: str | None, sockets) -> None:
|
||||
config.reload = False
|
||||
config.workers = 1
|
||||
if readiness_check:
|
||||
config.load()
|
||||
config.loaded_app = _ReadinessGate(config.loaded_app, import_from_string(readiness_check))
|
||||
server = _ReadyServer(config, control)
|
||||
|
||||
async def receive_commands() -> None:
|
||||
while True:
|
||||
if control.poll():
|
||||
try:
|
||||
command = control.recv()
|
||||
except EOFError:
|
||||
command = "drain"
|
||||
if command == "drain":
|
||||
server.should_exit = True
|
||||
return
|
||||
await asyncio.sleep(0.05)
|
||||
|
||||
async def serve() -> None:
|
||||
receiver = asyncio.create_task(receive_commands())
|
||||
try:
|
||||
await server.serve(sockets=sockets)
|
||||
finally:
|
||||
receiver.cancel()
|
||||
await asyncio.gather(receiver, return_exceptions=True)
|
||||
control.close()
|
||||
|
||||
config.setup_event_loop()
|
||||
asyncio.run(serve())
|
||||
|
||||
|
||||
@dataclass
|
||||
class Worker:
|
||||
process: BaseProcess
|
||||
control: Connection
|
||||
|
||||
def drain(self, *, wait_timeout: float = 0) -> bool:
|
||||
if self.process.is_alive():
|
||||
try:
|
||||
self.control.send("drain")
|
||||
if wait_timeout:
|
||||
deadline = time.monotonic() + wait_timeout
|
||||
while time.monotonic() < deadline:
|
||||
if self.control.poll(0.05) and self.control.recv() == "draining":
|
||||
return True
|
||||
if not self.process.is_alive():
|
||||
return True
|
||||
return False
|
||||
except (BrokenPipeError, EOFError, OSError):
|
||||
return not self.process.is_alive()
|
||||
return True
|
||||
|
||||
|
||||
class HandoffSupervisor:
|
||||
"""Start the replacement before draining the worker with existing connections."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: uvicorn.Config,
|
||||
*,
|
||||
startup_timeout: float = 60,
|
||||
reload_dirs: list[Path] | None = None,
|
||||
poll_interval: float = 0.25,
|
||||
readiness_check: str | None = None,
|
||||
):
|
||||
self.config = config
|
||||
self.startup_timeout = startup_timeout
|
||||
self.reload_dirs = reload_dirs or [Path("app"), Path("samples/peripheral")]
|
||||
self.poll_interval = poll_interval
|
||||
self.readiness_check = readiness_check
|
||||
self.current: Worker | None = None
|
||||
self.retiring: list[Worker] = []
|
||||
self.socket = config.bind_socket()
|
||||
self.socket.listen(config.backlog)
|
||||
self.should_exit = threading.Event()
|
||||
self._files: dict[Path, tuple[int, int]] = {}
|
||||
|
||||
def _candidate(self) -> Worker | None:
|
||||
parent = child = process = None
|
||||
ready = False
|
||||
try:
|
||||
parent, child = multiprocessing.get_context("spawn").Pipe()
|
||||
process = get_subprocess(
|
||||
config=self.config,
|
||||
target=partial(_serve_worker, config=self.config, control=child, readiness_check=self.readiness_check),
|
||||
sockets=[self.socket],
|
||||
)
|
||||
worker = Worker(process, parent)
|
||||
process.start()
|
||||
child.close()
|
||||
deadline = time.monotonic() + self.startup_timeout
|
||||
while not self.should_exit.is_set() and time.monotonic() < deadline:
|
||||
if parent.poll(0.05):
|
||||
try:
|
||||
ready = parent.recv() == "ready" and process.is_alive()
|
||||
except EOFError:
|
||||
break
|
||||
if ready:
|
||||
return worker
|
||||
if not process.is_alive():
|
||||
break
|
||||
logger.error("Replacement worker did not become ready; keeping the current worker")
|
||||
return None
|
||||
except Exception:
|
||||
logger.exception("Replacement startup or readiness handshake failed; keeping the current worker")
|
||||
return None
|
||||
finally:
|
||||
if child is not None:
|
||||
child.close()
|
||||
if not ready:
|
||||
# Only an unready candidate is terminated. The serving worker
|
||||
# and workers draining established connections are untouched.
|
||||
if process is not None and process.pid is not None:
|
||||
if process.is_alive():
|
||||
process.terminate()
|
||||
process.join(timeout=2)
|
||||
if process.is_alive():
|
||||
process.kill()
|
||||
process.join()
|
||||
if parent is not None:
|
||||
parent.close()
|
||||
|
||||
def start(self) -> bool:
|
||||
self._files = self._snapshot()
|
||||
self.current = self._candidate()
|
||||
return self.current is not None
|
||||
|
||||
def reload(self) -> bool:
|
||||
self.reap()
|
||||
candidate = self._candidate()
|
||||
if candidate is None:
|
||||
return False
|
||||
old, self.current = self.current, candidate
|
||||
logger.info("Worker %s is ready to accept new connections", candidate.process.pid)
|
||||
if old is not None:
|
||||
acknowledged = old.drain(wait_timeout=2)
|
||||
self.retiring.append(old)
|
||||
if not acknowledged:
|
||||
logger.error("Worker %s has not yet acknowledged draining", old.process.pid)
|
||||
return False
|
||||
logger.info("Worker %s is draining existing connections", old.process.pid)
|
||||
return True
|
||||
|
||||
def reap(self) -> None:
|
||||
remaining = []
|
||||
for worker in self.retiring:
|
||||
if worker.process.is_alive():
|
||||
remaining.append(worker)
|
||||
else:
|
||||
worker.process.join()
|
||||
worker.control.close()
|
||||
self.retiring = remaining
|
||||
|
||||
def _snapshot(self) -> dict[Path, tuple[int, int]]:
|
||||
files = {}
|
||||
for directory in self.reload_dirs:
|
||||
for path in directory.rglob("*.py"):
|
||||
try:
|
||||
stat = path.stat()
|
||||
files[path.resolve()] = (stat.st_mtime_ns, stat.st_size)
|
||||
except OSError:
|
||||
continue
|
||||
return files
|
||||
|
||||
def run(self) -> None:
|
||||
signals = (signal.SIGINT, signal.SIGTERM)
|
||||
original = {sig: signal.signal(sig, lambda *_: self.should_exit.set()) for sig in signals}
|
||||
try:
|
||||
if not self.start():
|
||||
raise RuntimeError("The initial API worker failed to start")
|
||||
pending_reload = False
|
||||
while not self.should_exit.wait(self.poll_interval):
|
||||
self.reap()
|
||||
files = self._snapshot()
|
||||
if files != self._files:
|
||||
self._files = files
|
||||
pending_reload = True
|
||||
active_failed = self.current is None or not self.current.process.is_alive()
|
||||
if active_failed:
|
||||
logger.error("The active worker exited; starting a replacement")
|
||||
pending_reload = True
|
||||
# Coalesce rapid edits while old SSE connections drain, rather
|
||||
# than keeping an unbounded number of application generations.
|
||||
if pending_reload and (active_failed or len(self.retiring) < 2):
|
||||
pending_reload = False
|
||||
self.reload()
|
||||
finally:
|
||||
try:
|
||||
self.close()
|
||||
finally:
|
||||
for sig, handler in original.items():
|
||||
signal.signal(sig, handler)
|
||||
|
||||
def close(self) -> None:
|
||||
self.should_exit.set()
|
||||
workers = [*self.retiring, *([self.current] if self.current else [])]
|
||||
for worker in workers:
|
||||
worker.drain()
|
||||
self.socket.close()
|
||||
for worker in workers:
|
||||
worker.process.join()
|
||||
worker.control.close()
|
||||
self.current = None
|
||||
self.retiring = []
|
||||
Reference in New Issue
Block a user