Batch processing fast-failed on full queues
Batch items now immediately fail instead of continuously retrying when an environment's queue reaches its limit, reducing database strain during overloads.
When a noisy tenant filled their queue to its maximum capacity, batch tasks would historically enter a continuous retry loop. This generated individual pre-failed task records for every item, consuming significant database write capacity for jobs that were guaranteed to fail.
Batch items are now immediately rejected when the queue limit is hit. Retries are completely bypassed, and the individual per-item errors are collapsed into a single aggregate error record per batch. This structural shift bounds database writes to scale with the number of batches rather than the number of individual items, keeping the system stable under load.
Separately, a new realtime service issues run-scoped JWTs, giving browser clients a secure way to subscribe directly to a run's realtime chat streams. These updates are deployed in the run engine and realtime webapp services.