While a /reserve is held, no requests complete so workload_served stays
at 0 each metrics tick. The autoscaler sees cur_perf=0 against
max_perf=150, concludes the worker can't deliver claimed throughput,
downgrades it, and gets cautious about scaling up — so additional
/reserve requests pile up behind the held one instead of triggering a
new worker.
Add a 1Hz heartbeat coroutine that, while anything is in flight, sets
workload_served back to TARGET_PERF (150) and flags update_pending. The
metrics tick reads 150 and resets to 0; the heartbeat re-pins it before
the next tick. Net effect: the autoscaler sees a saturated worker
delivering at peak rate, which is the signal it needs to scale a new
worker up rather than queue.
The heartbeat needs the backend instance, which is only created inside
Worker(...) — stash a reference in a module-level dict between Worker()
and .run() so the lifecycle coroutine can reach it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
