metadata should ideally use REPL, while the larger chunks of data should use DIST,
distributed locks should use REPL, but should never use a permanent store.
(SKIP_STORE is used, but still during a rehash they might get written to the store).