Losing an entire shard is pretty rare and in many cases can be attributed to network partitioning rather than failing processes. When a shard goes down, all operations that would go to this shard will fail. We can (and should) implement fault tolerance in our application level, allowing our application to resume for the operations that can be completed.

Choosing a shard key that can easily map on our operational side can also help; for example, if our shard key is based on location, we may lose the EU shard but will still be able to write and read data regarding US-based customers through our US shard.