Skip to content

Auto-stop status route reflects Service Bus queue depth

Motivation

The previous change made the auto-stop beat driver keep a Running AKS cluster alive while the Service Bus request queue still held undrained work, but left the autostop status route (/api/aks/.../autostop/status) unaware of the queue. That route is what the SPA polls to render the "Stops in mm:ss" countdown, so the dashboard could show an idle countdown ticking down while the beat had already decided to keep the cluster — a confusing decide-vs-display mismatch flagged as the out-of-scope follow-up (issue #52, step 4).

The reason it was deferred was cost: a naive implementation issues one Service Bus admin call (get_queue_runtime_properties) per status poll per cluster.

User-facing change

The autostop status route now passes the Service Bus request-queue depth into the evaluator, so the SPA countdown matches the beat decision: a Running cluster with queued work shows the kept state (reason=sb_queue_pending:N) instead of a shrinking idle countdown. The AutoStop panel renders this as "N requests queued in Service Bus — staying running.".

The admin-call cost is bounded by a deployment-global TTL cache (default 5s): the request queue is a single deployment-wide entity, so one admin read serves every cluster card and every concurrent browser within the window, regardless of how many clusters are polling. Same env gate as before (AKS_AUTOSTOP_RESPECT_SB_QUEUE, default on); an unreadable/disabled queue degrades to the prior job-count decision (additive only).

API / IaC diff summary

No HTTP contract change (the reason field already documents free-form codes). Internal changes:

  • New api/services/auto_stop_sb_signal.py — pending_queue_signal(power_state, *, ttl_seconds=5.0): the shared, gated, TTL-cached request-queue signal. ttl_seconds <= 0 bypasses the cache. Never raises.
  • api/tasks/azure/idle_autostop.py — _sb_pending_signal now delegates to the shared signal with ttl_seconds=0 (the beat act-path keeps reading the live queue); the inline gating + import os were removed.
  • api/routes/aks/autostop.py — _status_pending_queue_depth(power_state) feeds the cached signal into the status route's evaluate_cluster call.
  • web/src/components/ClusterItem/AutoStopPanel.tsx — reasonText renders the sb_queue_pending:N code; web/src/api/aks.ts doc comment lists the new code.

Validation evidence

  • uv run ruff check api — clean.
  • uv run pytest -q api/tests/test_auto_stop_sb_signal.py api/tests/test_idle_autostop_sb_queue.py api/tests/test_tasks_facade_contract.py api/tests/test_auto_stop_evaluator.py api/tests/test_aks_autostop_route.py — 128 passed. New test_auto_stop_sb_signal.py covers the gates, cache-hit (one admin call within TTL), cache-bypass (ttl=0), cache re-read, and degrade.
  • cd web && npm run build — succeeds.