signal-bot/signal_bot/app.py
Krzysztof kuhy Rudnicki 72482629d9 fix: a signal-api that is not up yet no longer kills the bot at boot
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-30 15:23:42 +02:00

96 lines
3.7 KiB
Python

# Copyright (c) 2026 Krzysztof Rudnicki. MIT licence.
"""The ASGI entrypoint, and the one endpoint that says whether the bot works.
Wiring lives in bot.py and the startup sequence in startup.py. What is left
here is the part uvicorn imports: an app that begins serving immediately, so
that "the bot is still connecting" is a state you can ask about rather than a
process that is not there.
"""
from __future__ import annotations
import asyncio
import contextlib
import logging
from typing import TYPE_CHECKING
import httpx
from fastapi import FastAPI, Response, status
from signal_bot.config import load_config
from signal_bot.startup import Starter
from signal_bot.store import Store
from signal_bot.versioning import POPE, __version__
if TYPE_CHECKING:
from collections.abc import AsyncIterator
logger = logging.getLogger(__name__)
HTTP_TIMEOUT_SECONDS = 30.0
# iNaturalist asks automated clients to identify themselves, and Wikimedia's
# policy outright refuses the default "python-httpx/x.y" this used to send.
USER_AGENT = f"signal-bot/{__version__} (+https://github.com/kuhyx/signal-bot)"
@contextlib.asynccontextmanager
async def lifespan(app: FastAPI) -> AsyncIterator[None]:
"""Serve first, connect to Signal afterwards.
Two earlier versions of this got it wrong in opposite directions. The
first awaited asyncio.gather() on two infinite tasks, so uvicorn never
finished starting and never served at all. The second connected to
signal-cli inline, so a container that was not ready yet failed startup
outright and burned the unit's restart limit in 25 seconds.
Config errors still fail here and exit non-zero: load_config() raising is
an answer, not an outage, and no amount of waiting fixes a missing
PHONE_NUMBER. Everything that needs the network is handed to Starter.
"""
config = load_config()
client = httpx.AsyncClient(timeout=HTTP_TIMEOUT_SECONDS, headers={"User-Agent": USER_AGENT})
# Opened once, here, rather than inside the retried connection attempt:
# a Store per attempt would leak one SQLite handle per failed try, for
# as long as the container stayed down.
store = Store(config.database_path)
starter = Starter(config, client, store)
app.state.starter = starter
connecting = asyncio.create_task(starter.run())
logger.info("serving on %s:%d while connecting to Signal", config.bind_host, config.bind_port)
try:
yield
finally:
connecting.cancel()
with contextlib.suppress(asyncio.CancelledError):
await connecting
if starter.runtime is not None:
await starter.runtime.stop()
await client.aclose()
store.close()
app = FastAPI(lifespan=lifespan)
@app.get("/healthz")
async def healthz(response: Response) -> dict[str, str]:
"""Liveness probe, and the only place the bot admits it is not working.
Answering 200 while nothing can reach Signal is what let a three-hour
outage look like a quiet evening, so this must go red before anybody
thinks to look at the logs. It has three answers, not two: a bot that has
not connected yet is "connecting", which is what a booting machine looks
like, and is not the same thing as one that has gone deaf.
"""
starter: Starter = app.state.starter
build = {"version": __version__, "pope": POPE.name}
runtime = starter.runtime
if runtime is None:
response.status_code = status.HTTP_503_SERVICE_UNAVAILABLE
return {"status": "connecting", "reason": starter.reason, **build}
liveness = runtime.bot.liveness
if liveness.healthy:
return {"status": "ok", **build}
response.status_code = status.HTTP_503_SERVICE_UNAVAILABLE
return {"status": "deaf", "reason": liveness.reason, **build}