Summary

• PrometheusMiddleware: add queue_wait_seconds histogram (end-to-end queue latency via _taskiq_enqueue_timestamp label injected in pre_send), task_errors_by_type counter
  (error count broken down by exception class name), and idempotent timestamp injection so multiple middlewares can coexist
• ReceiverObserver protocol: a new typing.Protocol (taskiq/receiver/observer.py) that lets external implementations observe receiver internals — prefetch queue depth, semaphore
  availability, active task count, unknown task lookups, and deserialization errors — with zero overhead when no observer is set (if self.observer is not None guards)
• PrometheusReceiverObserver: concrete implementation using Prometheus Gauges/Counters, wired automatically via PrometheusMiddleware.set_broker() → broker._receiver_observer →
  Receiver(observer=...)

New metrics

Some decisions decisions

• ReceiverObserver is a Protocol (not ABC) so there's no coupling — any class with matching methods satisfies it
• queue_wait_seconds gracefully degrades: if the sender doesn't use PrometheusMiddleware, the timestamp label is absent and the metric is simply not observed
• Negative queue wait values (from clock skew) are clamped to 0. If needed we can change this behavior.

Testing

Honestly still no idea how to test this via py test. But I'll be working on it.