A page that loses its answer now asks again instead of hanging

Occasionally a page would sit on a spinner for a full minute and then report that its content could not be found — content that plainly existed, and that loaded fine on the very next attempt.

The cause was a moment of bad timing. Content in the platform is served by a small worker dedicated to that item, and there are legitimate reasons for the platform to retire one and start a fresh copy — most commonly right after the item's type has been rebuilt, so the next visitor gets the new version rather than the old one. That swap is quick and normally invisible.

The problem was what happened to a request that arrived during the swap. The worker accepted it, started fetching, and then went away mid-answer. Nothing was left to reply, and nothing said so — so the person who asked simply waited, all the way to the one-minute limit, and was then told the content did not exist. That last part was the most misleading piece: the platform reported what it could see a minute later ("nobody is serving this"), not what had actually happened.

Now the hand-over is explicit. A request whose worker retires mid-answer is told exactly that — "this item is being refreshed, ask again" — and the platform asks again straight away, landing on the fresh copy. In practice the page loads normally instead of stalling, and the rare genuine failure now says what went wrong rather than blaming a missing item.

The same fix removes a whole class of silent waits: any read that loses the thing it was reading from now ends with an answer, never with silence.