homelab-codex-ws/services/control-plane/tests/test_supervisor_loop_resilience.py

156 lines
5.8 KiB
Python

"""Tests for supervisor loop resilience: a single cycle must never be able to
freeze the process forever, and the loop must survive exceptions.
Regression coverage for the 2026-07-15 incident: control-plane-supervisor
reported "Up 24h (healthy)" while the reconcile loop had silently stopped
ticking after processing one ha_websocket_dead event, with zero exceptions
and zero further log lines — because reconcile() had no per-cycle timeout
and loop() had no try/except around the cycle body.
"""
from __future__ import annotations
import sys
import time
from pathlib import Path
import pytest
sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
import supervisor as supervisor_module
from supervisor import Supervisor
def _make_supervisor(tmp_path: Path, monkeypatch) -> Supervisor:
actions = tmp_path / "actions"
events = tmp_path / "events"
world = tmp_path / "world"
state = tmp_path / "state"
repo = tmp_path / "repo"
for d in (actions, events, world, state, repo / "hosts"):
d.mkdir(parents=True, exist_ok=True)
monkeypatch.setattr(supervisor_module, "ACTIONS_DIR", actions)
monkeypatch.setattr(supervisor_module, "EVENTS_DIR", events)
monkeypatch.setattr(supervisor_module, "WORLD_DIR", world)
monkeypatch.setattr(supervisor_module, "REPO_ROOT", repo)
sup = Supervisor()
sup.desired_state = {"services": {}}
sup.actual_state = {"services": {}, "nodes": {}, "incidents": {}}
return sup
# ---------------------------------------------------------------------------
# 1. Loop processes many cycles without hanging
# ---------------------------------------------------------------------------
def test_loop_processes_multiple_cycles(tmp_path, monkeypatch):
sup = _make_supervisor(tmp_path, monkeypatch)
monkeypatch.setattr(time, "sleep", lambda s: None)
calls = []
monkeypatch.setattr(sup, "reconcile", lambda: calls.append(1))
sup.loop(interval=0, max_cycles=5)
assert len(calls) == 5
# ---------------------------------------------------------------------------
# 2. An exception in one cycle does not kill the loop
# ---------------------------------------------------------------------------
def test_exception_in_one_cycle_does_not_kill_loop(tmp_path, monkeypatch, caplog):
sup = _make_supervisor(tmp_path, monkeypatch)
monkeypatch.setattr(time, "sleep", lambda s: None)
calls = []
def flaky_reconcile():
calls.append(1)
if len(calls) == 2:
raise RuntimeError("simulated blowup mid-cycle")
monkeypatch.setattr(sup, "reconcile", flaky_reconcile)
with caplog.at_level("ERROR"):
sup.loop(interval=0, max_cycles=4)
# All 4 cycles ran despite cycle #2 raising.
assert len(calls) == 4
assert any("unhandled exception" in r.message for r in caplog.records)
def test_run_cycle_safely_never_raises(tmp_path, monkeypatch):
sup = _make_supervisor(tmp_path, monkeypatch)
monkeypatch.setattr(sup, "reconcile", lambda: (_ for _ in ()).throw(ValueError("boom")))
# Must not raise.
sup._run_cycle_safely()
# ---------------------------------------------------------------------------
# 3. A blocking cycle is bounded by a timeout; the loop continues afterward
# ---------------------------------------------------------------------------
def test_blocking_cycle_times_out_and_loop_continues(tmp_path, monkeypatch, caplog):
"""A reconcile() that hangs on a blocking call must not freeze loop().
Uses a single, genuinely-blocking cycle (real time.sleep in a background
thread — this is what an unkillable blocked syscall like fsync() looks
like from Python) bounded by a short reconcile_timeout. If the fix
regresses to the old bare `while True: self.reconcile()`, this test
times out the whole test run instead of failing fast.
"""
sup = _make_supervisor(tmp_path, monkeypatch)
calls = []
def hanging_reconcile():
calls.append(1)
time.sleep(0.5) # much longer than reconcile_timeout below
monkeypatch.setattr(sup, "reconcile", hanging_reconcile)
start = time.monotonic()
with caplog.at_level("ERROR"):
sup.loop(interval=0, max_cycles=1, reconcile_timeout=0.1)
elapsed = time.monotonic() - start
# loop() returned control well before the 0.5s hang finished.
assert len(calls) == 1
assert elapsed < 0.5, f"loop should not block for the full hang duration, took {elapsed}s"
assert any("did not complete within" in r.message for r in caplog.records)
def test_loop_keeps_advancing_across_repeated_timeouts(tmp_path, monkeypatch, caplog):
"""The outer loop must reach a second iteration even while a prior cycle
is still stuck in the background — proving the control loop itself (tick
counting, logging, sleeping) is never blocked by a wedged cycle."""
sup = _make_supervisor(tmp_path, monkeypatch)
monkeypatch.setattr(sup, "reconcile", lambda: time.sleep(0.5))
start = time.monotonic()
with caplog.at_level("ERROR"):
sup.loop(interval=0, max_cycles=3, reconcile_timeout=0.05)
elapsed = time.monotonic() - start
assert elapsed < 0.5, f"outer loop should complete 3 iterations fast, took {elapsed}s"
timeout_logs = [r for r in caplog.records if "did not complete within" in r.message]
assert len(timeout_logs) == 3
# ---------------------------------------------------------------------------
# 4. Existing behavior is preserved: reconcile() itself still works normally
# when called directly (not just through the loop).
# ---------------------------------------------------------------------------
def test_reconcile_still_works_standalone(tmp_path, monkeypatch):
sup = _make_supervisor(tmp_path, monkeypatch)
heartbeat = tmp_path / "state" / "supervisor.heartbeat"
sup.reconcile()
assert heartbeat.exists()