Before we return a sharded connection we check that the version is up-to-date on all connections to that namespace. This check does not throw a stale config exception, and prevents further stale config exceptions.

In write operations, we first find a chunk for the update with given config data then assume we'll get a stale config exception if this config data changes before we use it. Since the config data changes without throwing this exception and the config data is also up-to-date, we don't get an exception or a writeback.

Could be other complications.