fastapi-disconnect

Cancel FastAPI request handlers when the client disconnects.

By default, when a client disconnects (timeout, closed tab, network failure), FastAPI keeps executing the request handler to completion. For long-running handlers — an AI flow making paid LLM calls, a heavy database query — that means paying for work nobody is waiting for. This library cancels the handler the moment the client goes away, by listening for the ASGI http.disconnect event (no polling of request.is_disconnected()).

Installation

pip install fastapi-disconnect

Usage

App-wide, via middleware — protects every route, no handler changes:

from fastapi import FastAPI
from fastapi_disconnect import CancelOnDisconnectMiddleware

app = FastAPI()
app.add_middleware(
    CancelOnDisconnectMiddleware,
    exclude_paths=[r"/uploads/.*"],   # optional; re.fullmatch against the path
    on_disconnect=log_cancellation,   # optional; sync or async, gets the ASGI scope
    queue_size=64,                    # optional; bounds body buffering (default: unbounded)
)

on_disconnect runs only when a premature disconnect cancelled a request (not on normal connection close) — a natural place for a metric counting the work cancellation saved you.

Per-endpoint, via decorator — applied below the route decorator:

from fastapi_disconnect import cancel_on_disconnect

@app.get("/generate")
@cancel_on_disconnect
async def generate() -> str:
    return await expensive_llm_flow()

Cooperative, via dependency — for handlers that want to decide when to stop instead of being cancelled preemptively (e.g. finish the current step, checkpoint, then exit):

from fastapi_disconnect import Disconnected

@app.get("/agent")
async def agent(disconnected: Disconnected) -> Result:
    for step in plan:
        if disconnected.is_set():        # free local check, no polling
            return partial_result
        await run_step(step)

Disconnected is an annotated alias for Annotated[asyncio.Event, Depends(disconnected_event)]; the event is set the moment the client disconnects and can also be awaited (disconnected.wait()). Use it instead of the decorator/middleware on a route — combined with them, preemptive cancellation fires first (though the event is still set, which is handy inside shielded cleanup). All three mechanisms share a single per-request watcher, so any combination is safe.

On disconnect, the handler receives a regular asyncio.CancelledError at its next await, so finally blocks and async context managers run as usual. The decorator produces a 499 Client Closed Request response (which goes nowhere — the client is gone — but keeps logs meaningful); the middleware simply stops.

Cleanup on cancellation

Catching CancelledError for cleanup is supported, with one rule: the cancellation is level-triggered (anyio cancel-scope semantics). Once the client is gone, every subsequent await inside the handler raises CancelledError again — so asynchronous cleanup must run in a shielded scope:

import anyio

@app.get("/generate")
@cancel_on_disconnect
async def generate() -> str:
    try:
        return await expensive_llm_flow()
    except asyncio.CancelledError:
        with anyio.CancelScope(shield=True):
            await release_resources()  # runs to completion
        raise

An unshielded await in an except CancelledError or finally block is re-cancelled at its first checkpoint. (asyncio.shield() is not a substitute: it detaches the inner call, but the awaiting handler is still re-cancelled, so the cleanup continues without you.) Synchronous cleanup needs no shielding. Re-raise after cleaning up — a handler that swallows the cancellation and returns a value is treated as having completed normally.

Caveats

• asyncio only
• Don't read the raw body (request.body() / request.stream()) inside a guarded handler — the disconnect watcher owns the receive channel while the handler runs. FastAPI-parsed body params are fine, and under the middleware raw reads work normally.
• The decorator guards only the handler — dependencies and request parsing have already run when it starts, and a returned StreamingResponse body runs after it exits (the decorator warns). Use the middleware to cover the whole request lifecycle. Sync (def) handlers can't be cancelled and are rejected with TypeError.
• The middleware buffers request bodies eagerly by default — set queue_size to restore upload backpressure (disconnect detection pauses while the queue is full), exclude upload routes, or limit request size upstream.
• Proxies can hide disconnects — behind a buffering proxy (nginx, ALB) the http.disconnect event may arrive late or never.

License

MIT