FixedReproduced on demandcause established — the ordering is now harmless

test_subscriber_waits_for_ready

test_R05_event_set_after_wait.py · failed in 70% of captured runs

Forcing subscribe to start before publish_ready reproduced the failure on every attempt, so that ordering is a sufficient condition for the failure.

The two operations in `proven_inversion` name the same `resource`, and one has `access: "write"` while the other has `access: "read"`. The reader observed state the writer had not yet published.

Where the two runs diverge

The same test, twice, on a shared time axis. The outlined pair is the ordering that differs.

The same test twice on a shared time axis. Each operation is anchored at the moment it started. In the failing run subscribe#0 starts before publish_ready#0, and publish_ready never ran at all.
Passing run3 operations in 0.30 ms
publish_ready
subscribe
assert
Failing run2 operations in 0.15 ms
subscribe
assert
publish_ready — never ran
0 ms0.15 ms0.30 ms

Outlined: subscribe#0 starts before publish_ready#0 in the failing run, and after it when the test passes.

publish_ready never started in the failing run. The run flushed its spans normally, so that absence is evidence: the operation had not happened by the time the assertion read the state.

Evidence

Suspicion comes from comparing runs. The decision comes from forcing the ordering and seeing what happens.

OrderingSuspiciousnessWhen forcedVerdict
subscribe#0 → publish_ready#0the assertion depends on this1.00fails 100%reproduces the failure every time it is forced

One scheduling constraint is enough to reproduce this failure.

Policy gate

Every check the proposed patch had to pass before it was allowed to run.

All 15 checks passed. A patch runs only when every one of them does.

Refused shortcuts

  • no sleep calls introduced
  • no aliased sleep imports introduced
  • no timeout marker added or inflated
  • no retry or flaky decorator added
  • no retry loop wrapped around the assertion
  • no assertion removed or weakened
  • no exception handler swallowing the failure
  • no test skipped, xfailed or renamed out of collection

Required substance

  • a real synchronization primitive was added
  • the primitive is reachable from both the signal and the wait site
  • the patch is not a no-op

Structural safety

  • the proposed wait edge creates no wait-for cycle
  • no production-scope file modified without opt-in
  • no third-party or vendored file modified
  • the patched module parses

Verification

What was established, and at which strength. A weaker check is never presented as proof.

  • Reproduced the exact interleaving on demand

    before the fix it failed every time under the forced ordering; after the fix it passed every time under the identical ordering (confirmed)

  • Adversarial schedules

    not attempted for this incident, and so not claimed

  • Residual flake check

    20 of 20 ordinary runs stable, against 9 failures in the same number of runs before the fix

Measured overhead
-0.564 msno fixed delay introduced
Isolation
one process per run
Reproduction seed
random_seed_base 1729random_seed_sweep 1729..1733pythonhashseed 0python_version 3.12.13forced_order subscribe#0 -> publish_ready#0
Regression guard (executed and confirmed)
benchmark/cases/R05_event_set_after_wait/test_R05_event_set_after_wait.py::test_chronotrace_regression_ef292eae

Proposed change

Nothing is merged automatically. This is a diff for a human to review.

--- a/benchmark/cases/R05_event_set_after_wait/test_R05_event_set_after_wait.py
+++ b/benchmark/cases/R05_event_set_after_wait/test_R05_event_set_after_wait.py
@@ -6,6 +6,18 @@
 
 from benchmark.support import io_latency
 from chronotrace.capture.instrument import assertion, operation
+from chronotrace.schedule.harness import ScheduleHarness, force_order
+
+_chronotrace_gate_broker_ready = asyncio.Event()
+
+
+@pytest.fixture(autouse=True)
+def _chronotrace_reset_broker_ready():
+    """Provide a fresh synchronization gate for each test."""
+    global _chronotrace_gate_broker_ready
+    _chronotrace_gate_broker_ready = asyncio.Event()
+    yield
+
 
 BROKER: dict[str, object] = {}
 
@@ -14,11 +26,13 @@
 async def publish_ready() -> None:
     """Mark the broker as ready to accept subscribers."""
     BROKER["ready"] = True
+    _chronotrace_gate_broker_ready.set()
 
 
 @operation("subscribe", resource="broker.ready", access="read")
 async def subscribe() -> bool:
     """Subscribe, which is only valid once the broker is ready."""
+    await _chronotrace_gate_broker_ready.wait()
     return bool(BROKER.get("ready", False))
 
 
@@ -37,3 +51,19 @@
     with assertion("broker.ready"):
         assert accepted is True
     await startup
+
+
+@pytest.mark.asyncio
+async def test_chronotrace_regression_ef292eae() -> None:
+    """Reproduce the interleaving that used to fail, deterministically.
+
+    Generated by ChronoTrace for incident ef292eae. Before the repair this
+    race appeared in roughly 70% of runs; this guard forces the
+    exact ordering that caused it, so a regression fails here on every run
+    rather than once in a while.
+    """
+    forced_order = ['subscribe#0', 'publish_ready#0']
+    harness = ScheduleHarness(forced_order, timeout_s=5.0)
+    with force_order(harness):
+        await test_subscriber_waits_for_ready()
+    assert harness.reached == forced_order
← All incidents