because any reasonable implementation of the remote execution API will have long-ish lived workers, and local caches of artifacts that have been fetched.