For enforcing uniqueness in a sharded cluster, the officially recommended approach provided here https://docs.mongodb.com/manual/tutorial/unique-constraints-on-arbitrary-fields/#std-label-shard-key-arbitrary-uniqueness is simplistic and in production environment it brings non-trivial amount of work. Some considerations:

1. Ephemeral issues might cause inconsistencies between the two collections (for example, unique index collection update succeeded but not the main collection) and make some unique keys not useable.
2. There are many changes needed (we're using ORM Mongoose, there are many hooks for it to change) for enforce this universally.

What we ended up doing is to use distributed ephemeral locks (a TTLed MongoDB collection) to lock on the unique keys before adding to the collection. Getting this correct / verifying its correct is non-trivial.