If two calls to ShardRegistry::reload() overlap, it's possible for one that started sooner and targeted a staler node to finish after one that started later. This means that the ShardRegistry could already have been loaded with newer data, but then throw it out and replace it with older data when the slower reload finishes.