Commit Graph

3 Commits

Author SHA1 Message Date
Joseph Doherty 4b9bcddbcc feat(drivers): periodic reconcile of desired vs actual subscriptions (#488)
Option A from the issue — the cheap consistency check, no interface change.

The desired ref set was re-applied on a Connected TRANSITION (`ResubscribeDesired`)
and on a deploy's `SetDesiredSubscriptions`, and nowhere else. A subscription lost
while the driver STAYED Connected — a subscribe that failed and was never retried,
a handle dropped down an error path — left the actor subscribed to nothing while
still reporting Healthy, indefinitely.

`ReconcileSubscription` now rides the existing 30 s `health-poll` tick: while
Connected, a driver holding a desired set but no live handle is re-subscribed.

Gated on `ISubscribable`. A driver that cannot subscribe at all can still be handed
a desired set, and without the guard it would be told `Subscribe` every tick and
fail every tick, forever.

This is a state-machine consistency check, NOT a probe — it can only see that this
actor holds no handle, never that a handle is present but dead server-side, because
`ISubscriptionHandle` is opaque. That is option B, and it stays deferred until a
real occurrence shows the handle outliving the subscription; it would need a
liveness member implemented across all eight drivers with different subscription
mechanics.

A redundant `Subscribe` is possible (a tick queued ahead of an already-queued one
observes the same no-handle state) and is harmless: `HandleSubscribeAsync` is a
`ReceiveAsync`, so no tick is processed mid-subscribe, and it drops any prior
subscription before establishing the new one. Suppressing it would need a
pending-subscribe flag threaded through every self-tell site — more state than the
race is worth.

`healthPollInterval` becomes an optional Props parameter, mirroring the existing
`rediscoverInterval` seam, so a test can drive the reconcile without waiting 30 s.

Three tests. The reconcile itself is proven by a stub whose FIRST subscribe throws:
nothing else would ever retry it, so a second attempt can only come from the
reconcile — disabling `ReconcileSubscription` turns it red. The two guard tests are
absence assertions and are bounded by a positive control rather than a sleep: a
counting health publisher makes ticks PROCESSED observable, so "still one subscribe"
means the reconcile ran and declined, not that the timer had not fired.

Explicitly not deploy-time drift (#486) — re-asserting an already-correct desired
set would not have helped there.
2026-07-26 10:30:19 -04:00
Joseph Doherty b0c031f09f fix(drivers): don't tear down field I/O when a deploy's artifact reads back empty (#485)
The driver-side half of the same mistake the address-space fix corrected.
ReconcileDrivers and PushDesiredSubscriptionsFromArtifact both read the
artifact as `...FirstOrDefault() ?? Array.Empty<byte>()`. Their THROW paths
already returned early, but empty bytes flowed onwards as a real answer: zero
driver specs, so DriverSpawnPlanner planned every running child for StopChild,
and then an empty desired set dropped each surviving driver's live
subscription handle. A node lost its entire field I/O and still ACKed Applied.

Same rule as the address space: no bytes is no answer, not "a configuration
with no drivers". Nothing legitimate produces a zero-length blob — deleting
the last driver still deploys a JSON document with empty arrays. Both sites
now skip and keep what is running. The two guards are independent because
PushDesiredSubscriptions does its OWN ConfigDb read, so the row can go missing
between them.

Also corrects the ReconcileDrivers doc comment, which claimed an empty blob
made it "effectively a no-op" — with children running it was the opposite.

Tests: DriverHostActorUnreadableArtifactTests, RED-first (verified failing —
after the empty dispatch the driver list was empty). A zero-length ArtifactBlob
reproduces byte-for-byte what the missing-row case delivers, so the race does
not have to be constructed.

Two controls, both load-bearing:
- a READABLE driverless artifact still stops the driver, so the guard keys on
  "the read gave us nothing", not on "fewer drivers than before";
- dropping a driver's LAST TAG still clears its subscription. This one was
  added after the absence assertion was caught passing on a race: the
  unsubscribe is an async self-tell, so with the second guard deleted the test
  still went green. The control observes that same unsubscribe arriving well
  inside the settle window, which is what makes the absence meaningful — with
  the guard deleted the suite now correctly goes RED.

SubscribableStubDriver gains an UnsubscribeCount for that observation.

Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
2026-07-21 02:40:14 -04:00
Joseph Doherty c03361de1b test(drivers): extract shared stub-driver harness (de-dup) 2026-06-14 22:49:26 -04:00