A reply that cannot be delivered now says so instead of hanging

When the portal runs on more than one server, a request may be answered by a different server from the one your session is on. If that server could not complete the request, it sent back a failure notice — and that notice travelled on a channel with no delivery guarantee. Publishing to a channel nobody is listening on succeeds, so the notice was simply discarded. Nothing failed, nothing was logged, and your request waited out its full minute-long budget for an answer the server believed it had already sent.

That was the worst possible shape for it: a lost error is indistinguishable from a slow one, so a page that should have shown "not found" in a fraction of a second instead looked frozen, and there was nothing in the logs to explain it.

Failure notices now travel the same directed route that ordinary messages take, which either lands or reports back. So an error reaches you as an error, at once. In the one case where that route is not available — during a rolling upgrade, while some servers are still on the previous release — the old channel is still used, but it now checks for a listener first and says loudly in the logs when there is none, instead of pretending the message went through.