Two replicas installing one module no longer store it twice

An installation that runs several replicas has them all watch the same plugin registry, so when a module is published they all adopt it — usually within the same second. Each replica downloaded the module and wrote it into a folder named with a fresh random id. The bytes were identical; the folder names were not.

That difference propagated. The installation records one module set — the exact folder of every module it runs, so every replica runs the same thing — and each replica derived that set from the folder it had just written. Two replicas therefore proposed two different sets for the same moment in the same installation. One was picked (deterministically, so nobody disagreed about which), nothing was lost, and the other replica's copy stayed on the volume forever. The rejected proposal stayed too, and every replica that started afterwards re-read it and reported the same already-settled conflict again — an error line on every start, about a state that was correct.

On the public instance this had accumulated to 100 such conflicts, 687 bookkeeping records and 843 module folders that nothing could clear.

A module folder is now named after its contents. Two replicas adopting the same published module compute the same name, so the second one finds the module already there and adopts it instead of storing a second copy. They derive the same module set, the second proposal is simply "nothing changed", and no conflict is recorded — because there is no disagreement to record. Two replicas that genuinely adopt different modules still produce two sets, and that conflict is still reported: it now means something.

A conflict that does occur is also settled once. The rejected proposal is retired by the same housekeeping pass that clears superseded records, so it is reported where it is decided rather than on every start of every replica from then on. Nothing is lost by retiring it: the next update wave derives the set from what is actually installed, never from the previous proposal.

One consequence is worth knowing: if a module is republished with a new version number but byte-identical contents, its folder does not change — the new version is recorded, and what runs is what already ran, because the bytes are the same.

See Module Set Convergence for the mechanism.