https://docs.mongodb.com/manual/tutorial/build-indexes-on-replica-sets/
If one follows the round-robin method for building an index with a unique constraint, it is possible to end up with different data on different nodes (see SERVER-30276). There are workarounds to avoid this situation:
1. Build the index normally using a foreground or background index build.
Or
2. Stop writes to the collection (or avoid inserting unique-constraint-violating data) during the round-robin index building.