The index creation on the shard collection path is performed using a DBClient, which does not wait for majority write concern. The following scenario might happen:
1. The shard index is created locally
2. The writes to config.chunks and config.collections are sent to the config server
3. A stepdown happens
4. The index creation is rolled backed
If the write on step two succeeds then we'll end up with a sharded collection without the sharded index created on the primary. We should wait for majority write concern of the latest write after the index creation (probably using the same as the sharding catalog manager)