You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When a committor intent fails, later intents that depend on the same committed accounts can stay in blocked_intents forever. They never run, and callers waiting for their results never hear back.
Reproduction
Schedule an intent that fails, followed by one that shares a committed key with it and also uses a second key. Then schedule another intent using that second key. After the first failure, inspect the blocked queue and result subscriptions: the dependency extends beyond intents that directly share the original key.
The scheduler cascade and execution-engine tests in #1524 cover this scenario in source.
Expected
Fail the waiting dependents too, rather than executing work that relies on a failed intent. Remove them transitively from the blocked queue, report the original error for the failed intent, and send exactly one PoisonedIntentError for each removed successor—even when several dependency paths reach it. Unaffected predecessors and unrelated work can still run.
The failed intent's committed keys remain poisoned until the process restarts. Reject new intents touching those keys and poison every key in each rejected bundle.
There's an important difference for callers: already-scheduled successors receive failure results, but new intents rejected at admission receive no result broadcast. Callers must handle those paths differently.
Expose poisoned-key and cascade-size metrics so we can see how much work is affected.
Context
The changes are in magicblock-committor-service/src/intent_engine/intent_scheduler.rs and intent_execution_engine.rs in #1524. Poison state is process-local; it doesn't survive a restart.
Activity