A page that failed to load once no longer stays broken

Some documents simply refused to open. Not slowly — instantly, with the same technical message every time: "No response received … the request may have been undeliverable." Reloading changed nothing. Waiting changed nothing. The page stayed broken for hours, while its neighbours in the very same folder opened fine, the search results still listed it, and the folder listing still showed it. Only opening the document itself failed.

That combination is the clue. Nothing was wrong with the document, and nothing was wrong with the server. What was wrong was the memory of a single bad moment.

Behind every open document there is one shared connection to the piece of the system that owns it. When that owner is briefly unreachable — it was asleep and is waking up, a deployment is rolling, the database blinks — the connection fails, and the failure is handed to whoever was waiting. That part is fine and expected; the next visitor should simply get a fresh connection.

Except the failed connection was kept. Every later reader was handed the recorded failure instead of a new attempt, so no new attempt was ever made — which meant the system never learned that the owner had come back. The proof was in the error itself: three attempts eleven minutes apart came back quoting the identical internal request number. It was not failing three times; it was showing one old failure three times. The safety valve designed to notice repeated failures and slow them down could not help either, because from its point of view the failure had only ever happened once.

Two things kept the stale connection alive: the shared connection register, and a second cache one layer below it that handed the same dead connection back even after the first was cleared. Both are now dropped the moment a reader arrives and finds the previous attempt had ended in failure, with nothing currently holding it back on purpose. The next read opens a genuinely new connection to the owner, and a document that failed once opens normally as soon as its owner is reachable again.

The protection against a genuinely broken owner is unchanged, and in fact now works as intended: repeated real failures are still counted and still backed off, briefly and then progressively, so a wedged owner is never hammered. The difference is that "try again" now means trying again.