test(historian): bounded-wait durability read; honest nondeterminism notes on the #491 controls

2026-08-17 04:07:08 -04:00
parent 20022083df
commit 6b46c485ef
2 changed files with 56 additions and 1 deletions
@@ -353,6 +353,18 @@ public sealed class ContinuousHistorizationRecorderTests : TestKit
    ///         The yield is the whole point: every other test here uses an outbox that returns
    ///         <c>ValueTask.CompletedTask</c>, so the continuation runs inline with the context still
    ///         installed and the defect is invisible. The production FasterLog outbox fsyncs per entry.
+    ///         The outbox double must therefore complete on a POOL thread
+    ///         (<c>Task.Delay(1).ConfigureAwait(false)</c>); a bare <c>Task.Yield()</c> posts back to the
+    ///         CAPTURED ActorTaskScheduler context and is just as vacuous as completing synchronously.
+    ///     </para>
+    ///     <para>
+    ///         ⚠️ The falsification control is NONDETERMINISTIC UNDER LOAD. Re-introducing the defect is
+    ///         only observable when the thread that resumes the continuation happens to carry no Akka
+    ///         actor cell in its ThreadStatic slot; on a loaded parallel test sweep the resuming pool
+    ///         thread can already be servicing this ActorSystem, the <c>Self</c> read succeeds, and the
+    ///         mutation passes. So a SINGLE green run of the intended-red mutation is not proof this
+    ///         regression is caught — re-run the mutation several times, ideally on an otherwise idle
+    ///         machine, before concluding the test is falsifiable.
    ///     </para>
    /// </summary>
    [Fact]
@@ -931,6 +943,15 @@ public sealed class ContinuousHistorizationRecorderTests : TestKit
    ///         <c>Task.Yield()</c>, which posts back to the CAPTURED ActorTaskScheduler context — leaves the
    ///         context installed and the defect invisible.
    ///     </para>
+    ///     <para>
+    ///         ⚠️ The falsification control is NONDETERMINISTIC UNDER LOAD. Re-introducing the defect is
+    ///         only observable when the thread that resumes the continuation happens to carry no Akka
+    ///         actor cell in its ThreadStatic slot; on a loaded parallel test sweep the resuming pool
+    ///         thread can already be servicing this ActorSystem, the <c>Self</c> read succeeds, and the
+    ///         mutation passes. So a SINGLE green run of the intended-red mutation is not proof this
+    ///         regression is caught — re-run the mutation several times, ideally on an otherwise idle
+    ///         machine, before concluding the test is falsifiable.
+    ///     </para>
    /// </remarks>
    [Fact]
    public async Task A_yielding_batch_append_neither_faults_the_actor_nor_loses_values()