A backend can fail mid-request, and the transport returns an error. The active count must still come back down, or a flaky backend slowly looks permanently overloaded. Today you pin that the lease is released even on error.
Ensure Dispatch releases the lease and returns the error when the transport fails, leaving the active count at zero.
Connections do not only end cleanly - a backend can time out, reset, or return an error partway through. When that happens the balancer still has to release the lease, because the request is no longer in flight. Forget this and every failed request permanently inflates the backend’s active count; least-connections would then steer traffic away from a backend that has actually recovered, and the counts drift further from reality with each failure.
If you wrote Dispatch with a defer release() last lesson, this behavior is
already correct - the deferred call runs on every return path, success or error - so
today is a payoff: add the failing-transport test and watch the active count
return to 0 even though the request failed. If instead you released only on the
success path, this is the lesson to move the release into a defer. Either way, the
error itself is propagated to the caller; releasing the lease and reporting the
failure are two separate obligations, and Dispatch owes both.
// No new code should be needed if Dispatch already uses `defer release()`:// the defer runs whether the transport returns a response or an error.// This lesson adds the failing-transport test and pins Active() == 0.// If your Dispatch released only on success, move the release into a defer now.