Add /release control endpoint to null pyworker

The held /reserve now waits on an asyncio.Event and resolves when the local
queue consumer POSTs /release on the internal control port (127.0.0.1:18999
by default). This produces a 200 success in metrics instead of the 499
cancellation you got from disconnecting the client. The duration cap stays
as a safety net for stuck consumers.

The internal aiohttp server is now unconditional and hosts /release always;
the stub /health route is added only when BACKEND_HEALTH_URL is unset.
NULL_STUB_HEALTH_PORT is renamed to NULL_CONTROL_PORT to reflect the
broader role.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

workers/null/README.md

@@ -2,8 +2,9 @@
 
 A PyWorker that does **nothing** — it does not forward requests to any model
 server. Each HTTP POST to `/reserve` simply marks the worker as busy and holds
-the request open until the caller disconnects (or a configured timeout
+the request open until the user's queue consumer (running locally on the
-elapses).
+instance) calls `/release` on the internal control port — or a safety
+timeout elapses.
 
 ## When to use it
 
@@ -18,11 +19,12 @@ Use this worker when you want to drive Vast Serverless autoscaling but you do
   Serverless autoscaler to spin instances up and down based on demand on
   *your* side.
 
-For each job your side wants to run on a Vast instance, you POST once to
+For each batch of work your side wants on a Vast instance, you POST once to
-`/reserve`. The autoscaler will provision a worker if none is free; the
+`/reserve`. The autoscaler provisions a worker if none is free; the request
-request stays open, keeping that worker counted as busy, until you close the
+stays open, keeping that worker counted as busy. When your queue consumer
-connection. When you close, the worker goes idle and the autoscaler is free
+finishes its work it POSTs `/release` on `127.0.0.1:18999` and the held
-to scale it down.
+`/reserve` returns `200`, so the request is recorded as a normal success in
+Vast metrics (not a cancellation).
 
 ## How it works
 
@@ -33,19 +35,22 @@ to scale it down.
 - `lifecycle` is used instead of `model_log_file`, so there is no log to tail
   and no model server to start. The worker reports itself ready immediately
   after the (trivial) benchmark.
-- The handler is a `remote_function` rather than an HTTP proxy, so the
+- The `/reserve` handler is a `remote_function` rather than an HTTP proxy, so
-  framework never tries to forward the request anywhere.
+  the framework never tries to forward the request anywhere — it just awaits
+  an internal `asyncio.Event`.
+- An internal aiohttp control server, bound to `127.0.0.1`, hosts
+  `/release` (and, when no external healthcheck URL is provided, a stub
+  `/health`).
 
 ## Healthchecking
 
 The framework periodically GETs a healthcheck URL after startup; if it ever
 fails after the first success, the worker is marked errored and the
-autoscaler can decommission it. The null worker exposes two modes:
+autoscaler can decommission it. Two modes:
 
-- **Stub (default)** — a tiny HTTP server runs on
+- **Stub (default)** — the internal control server also answers
-  `http://127.0.0.1:18999/health` (override the port with
+  `GET /health` with `200`. This is just enough to satisfy the framework
-  `NULL_STUB_HEALTH_PORT`) and always returns `200`. This is just enough to
+  while you wire up real consumers.
-  satisfy the framework while you wire up real consumers.
 - **Point at your queue consumer (recommended)** — set
   `BACKEND_HEALTH_URL=http://127.0.0.1:9090/health` (absolute URL) and the
   pyworker will healthcheck *your* consumer instead. If your consumer
@@ -57,39 +62,60 @@ your template.
 
 ## API
 
-### `POST /reserve`
+### `POST /reserve`  (external port, signed by the autoscaler)
 
-Holds the worker busy for the lifetime of the request.
+Holds the worker busy until the reservation ends.
 
 Request body (all fields optional):
 
 ```json
-{ "duration": 60 }
+{ "duration": 600 }
 ```
 
-- `duration` (seconds, optional): how long to hold the reservation if the
+- `duration` (seconds, optional): safety cap on how long to hold the
-  client does not disconnect first. Capped by `MAX_RESERVATION_SECONDS` (env
+  reservation if no `/release` arrives. Capped by `MAX_RESERVATION_SECONDS`
-  var, default 3600). If omitted, defaults to the cap.
+  (env var, default 3600). If omitted, defaults to that cap.
 
 Behavior:
 
-- Returns `200` with `{"released": "duration_elapsed", "duration": <n>}` when
+- Returns `200` with `{"released": "explicit", ...}` when the local consumer
-  the duration elapses normally.
+  POSTs `/release` on the internal port. **This is the intended happy path
-- Returns `499` when the client disconnects (the reservation is released
+  — the request is counted as a success in metrics.**
-  immediately).
+- Returns `200` with `{"released": "duration_elapsed", "duration": <n>}` if
+  the duration cap fires (safety net for a stuck consumer).
+- Returns `499` if the external client disconnects (counted as cancelled in
+  metrics — avoid this; use `/release` instead).
 - Returns `429` if the worker is already busy and queue wait would exceed
   `max_queue_time` (30s by default).
 
+### `POST /release`  (internal port, localhost-only)
+
+Marks the active reservation as done. No body required. Idempotent:
+
+```bash
+curl -X POST http://127.0.0.1:18999/release
+```
+
+Responses:
+
+- `200 {"released": true}` — active reservation was released; the held
+  `/reserve` will return `{"released": "explicit"}`.
+- `200 {"released": false, "reason": "no active reservation"}` — nothing was
+  in flight, no-op.
+
+Only processes on the Vast instance can reach this port. There is no
+authentication on it.
+
 ## Environment variables
 
 - `MAX_RESERVATION_SECONDS` — upper bound on how long a single `/reserve`
-  call can hold a worker. Defaults to `3600`. Set lower if you want a tighter
+  call can hold a worker if `/release` is never called. Defaults to `3600`.
-  safety cap against stuck clients.
 - `BACKEND_HEALTH_URL` — absolute URL the framework should healthcheck
-  (e.g. `http://127.0.0.1:9090/health`). When set, the stub server does not
+  (e.g. `http://127.0.0.1:9090/health`). When set, the stub `/health` route
-  run. When unset, the built-in stub is used.
+  is not registered on the internal server. When unset, the built-in stub
-- `NULL_STUB_HEALTH_PORT` — port for the built-in stub healthcheck server.
+  is used.
-  Defaults to `18999`. Only used when `BACKEND_HEALTH_URL` is unset.
+- `NULL_CONTROL_PORT` — port for the internal control server (hosts
+  `/release` and optionally `/health`). Defaults to `18999`.
 
 ## Deploying on Vast Serverless
 
@@ -100,26 +126,35 @@ Behavior:
 3. There is no model server to configure; you can omit model-related env vars
    entirely.
 4. Run your own queue-consumer process on the instance alongside the
-   PyWorker (e.g. as a separate supervisor service started by the template).
+   PyWorker. When the consumer finishes its work it should:
+   ```bash
+   curl -X POST http://127.0.0.1:18999/release
+   ```
+   so the held `/reserve` returns success and the autoscaler can scale the
+   worker down cleanly.
 
 ## Client example
 
 ```bash
-python -m workers.null.client --endpoint <ENDPOINT_NAME> --duration 300
+python -m workers.null.client --endpoint <ENDPOINT_NAME> --duration 600
 ```
 
-This will POST once to `/reserve`, which causes exactly one worker to be
+This POSTs once to `/reserve`, which causes exactly one worker to be
-provisioned (if none is free) and held busy for up to 300 seconds. Killing
+provisioned (if none is free) and held busy. To exercise the full flow,
-the client process (Ctrl-C) drops the connection and releases the worker
+shell into the worker and run `curl -X POST http://127.0.0.1:18999/release`
-early.
+— the client will return with `{"released": "explicit", ...}`.
 
 ## Notes and caveats
 
-- The HTTP connection must stay open for the full reservation. Make sure
+- The HTTP connection from the external caller must stay open for the full
-  your client and any intermediate proxies allow long-lived requests
+  reservation. Make sure your client and any intermediate proxies allow
-  (disable idle timeouts, retries, and connection reuse if necessary).
+  long-lived requests (disable idle timeouts, retries, and connection
+  reuse if necessary).
 - If your client retries on timeout, you may end up provisioning duplicate
-  workers. Use idempotent semantics in *your* queue, or set `duration` to a
+  workers. Configure `duration` generously and rely on `/release` from the
-  finite value and accept release-on-elapse as the normal exit.
+  consumer to end reservations promptly.
+- Avoid disconnecting the external `/reserve` request as a way to release —
+  that produces a `499` and is counted as a cancellation in Vast metrics.
+  Always release via `POST /release` on the internal port.
 - There is no streaming / heartbeat in the response; the request returns
   exactly once, when the reservation ends.

workers/null/worker.py

@@ -2,6 +2,7 @@ import asyncio
 import logging
 import os
 from contextlib import asynccontextmanager
+from typing import Optional
 from urllib.parse import urlsplit
 
 from aiohttp import web
@@ -16,8 +17,8 @@ from vastai import (
 
 log = logging.getLogger(__file__)
 
-# Safety cap: if a client never disconnects and never sets `duration`, the
+# Safety cap: if the user's queue consumer never calls /release, the
-# reservation is auto-released after this many seconds so a stuck client
+# reservation is auto-released after this many seconds so a forgotten /release
 # can't pin a worker indefinitely. Override with MAX_RESERVATION_SECONDS.
 MAX_RESERVATION_SECONDS = float(os.environ.get("MAX_RESERVATION_SECONDS", 3600))
 
@@ -25,20 +26,19 @@ MAX_RESERVATION_SECONDS = float(os.environ.get("MAX_RESERVATION_SECONDS", 3600))
 # immediately during capacity estimation instead of sleeping.
 BENCHMARK_SENTINEL = "__null_worker_benchmark__"
 
-# Healthcheck wiring. The framework periodically GETs
+# Internal control server. Hosts:
-# `<model_server_url>:<model_server_port><model_healthcheck_url>` and marks the
+#   * POST /release  — always available, marks the active reservation as
-# worker errored if that ever fails after the first success. For the null
+#     done so the held /reserve returns 200 (success in metrics, not a
-# worker we either:
+#     cancellation).
-#   * point at a URL the user supplies via BACKEND_HEALTH_URL — typically
+#   * GET  /health   — only when no external BACKEND_HEALTH_URL is set; the
-#     their own queue-consumer's health endpoint, so the autoscaler sees the
+#     framework's healthcheck loop polls it so the worker has a live signal.
-#     worker as broken if the consumer dies, or
+# Bound to 127.0.0.1 so only processes on the instance can reach it.
-#   * run a tiny built-in stub that always returns 200, so the framework has
+INTERNAL_HOST = "127.0.0.1"
-#     something live to talk to until the user wires up a real consumer.
+INTERNAL_PORT = int(os.environ.get("NULL_CONTROL_PORT", 18999))
-BACKEND_HEALTH_URL = os.environ.get("BACKEND_HEALTH_URL", "").strip()
-STUB_HEALTH_HOST = "127.0.0.1"
-STUB_HEALTH_PORT = int(os.environ.get("NULL_STUB_HEALTH_PORT", 18999))
 STUB_HEALTH_PATH = "/health"
 
+BACKEND_HEALTH_URL = os.environ.get("BACKEND_HEALTH_URL", "").strip()
+
 if BACKEND_HEALTH_URL:
     _parsed = urlsplit(BACKEND_HEALTH_URL)
     if not _parsed.scheme or not _parsed.hostname:
@@ -48,43 +48,73 @@ if BACKEND_HEALTH_URL:
     HEALTH_BASE_URL = f"{_parsed.scheme}://{_parsed.hostname}"
     HEALTH_PORT = _parsed.port or (443 if _parsed.scheme == "https" else 80)
     HEALTH_PATH = _parsed.path or "/"
-    USE_STUB = False
+    USE_STUB_HEALTH = False
 else:
-    HEALTH_BASE_URL = f"http://{STUB_HEALTH_HOST}"
+    HEALTH_BASE_URL = f"http://{INTERNAL_HOST}"
-    HEALTH_PORT = STUB_HEALTH_PORT
+    HEALTH_PORT = INTERNAL_PORT
     HEALTH_PATH = STUB_HEALTH_PATH
-    USE_STUB = True
+    USE_STUB_HEALTH = True
+
+
+# Singleton active reservation. `allow_parallel_requests=False` on the
+# /reserve handler guarantees the framework only runs one at a time per
+# worker, so a single slot is enough.
+_active_reservation: Optional[asyncio.Event] = None
+
+
+def _build_internal_app() -> web.Application:
+    app = web.Application()
+
+    async def release_handler(_request: web.Request) -> web.Response:
+        event = _active_reservation
+        if event is None:
+            return web.json_response(
+                {"released": False, "reason": "no active reservation"},
+                status=200,
+            )
+        event.set()
+        return web.json_response({"released": True}, status=200)
+
+    app.router.add_post("/release", release_handler)
+
+    if USE_STUB_HEALTH:
+        async def stub_health(_request: web.Request) -> web.Response:
+            return web.Response(status=200, text="ok")
+
+        app.router.add_get(STUB_HEALTH_PATH, stub_health)
+
+    return app
 
 
 @asynccontextmanager
 async def null_lifecycle():
-    runner = None
+    app = _build_internal_app()
-    if USE_STUB:
+    runner = web.AppRunner(app)
-        async def stub_health(_request: web.Request) -> web.Response:
+    await runner.setup()
-            return web.Response(status=200, text="ok")
+    site = web.TCPSite(runner, INTERNAL_HOST, INTERNAL_PORT)
+    await site.start()
 
-        app = web.Application()
+    lines = [
-        app.router.add_get(STUB_HEALTH_PATH, stub_health)
+        f"Null pyworker internal control server: http://{INTERNAL_HOST}:{INTERNAL_PORT}",
-        runner = web.AppRunner(app)
+        f"  POST /release  - end the active reservation (call from your queue consumer)",
-        await runner.setup()
+    ]
-        site = web.TCPSite(runner, STUB_HEALTH_HOST, STUB_HEALTH_PORT)
+    if USE_STUB_HEALTH:
-        await site.start()
+        lines.append(
-        log.info(
+            f"  GET  {STUB_HEALTH_PATH}      - stub healthcheck (override with BACKEND_HEALTH_URL)"
-            f"Null pyworker stub healthcheck listening on "
-            f"http://{STUB_HEALTH_HOST}:{STUB_HEALTH_PORT}{STUB_HEALTH_PATH} "
-            f"(override by setting BACKEND_HEALTH_URL)"
         )
     else:
-        log.info(f"Null pyworker healthcheck pointing at {BACKEND_HEALTH_URL}")
+        lines.append(f"Framework healthcheck pointed at: {BACKEND_HEALTH_URL}")
+    log.info("\n".join(lines))
 
     try:
         yield
     finally:
-        if runner is not None:
+        await runner.cleanup()
-            await runner.cleanup()
 
 
 async def reserve_worker(**params: object) -> dict:
+    global _active_reservation
+
     if params.get(BENCHMARK_SENTINEL):
         return {"ok": True, "benchmark": True}
 
@@ -97,17 +127,27 @@ async def reserve_worker(**params: object) -> dict:
         except (TypeError, ValueError):
             duration = MAX_RESERVATION_SECONDS
 
+    event = asyncio.Event()
+    _active_reservation = event
     log.info(
-        f"Reservation acquired; holding worker busy for up to {duration:.1f}s "
+        f"Reservation acquired; awaiting POST /release on "
-        f"(release early by disconnecting the HTTP request)"
+        f"http://{INTERNAL_HOST}:{INTERNAL_PORT}/release "
+        f"(auto-release after {duration:.1f}s)"
     )
     try:
-        await asyncio.sleep(duration)
+        try:
-        log.info("Reservation duration elapsed; releasing worker")
+            await asyncio.wait_for(event.wait(), timeout=duration)
-        return {"released": "duration_elapsed", "duration": duration}
+            log.info("Reservation released via /release")
-    except asyncio.CancelledError:
+            return {"released": "explicit", "duration_cap": duration}
-        log.info("Reservation released by client disconnect")
+        except asyncio.TimeoutError:
-        raise
+            log.warning(
+                f"Reservation hit duration cap of {duration:.1f}s without "
+                f"explicit /release; releasing automatically"
+            )
+            return {"released": "duration_elapsed", "duration": duration}
+    finally:
+        if _active_reservation is event:
+            _active_reservation = None
 
 
 worker_config = WorkerConfig(