In a two-shard cluster, an empty collection sharded with a hashed key should create two chunks and each shard should own one chunk. When there is a concurrent setFCV operation, the two chunks could both live on only one of the shards. 

This could be due to shardCollection checking the value of an unstable featureFlag twice during its execution. It is mentioned during discussion that gOneChunkPerShardEmptyCollectionWithHashedShardKey feature  flag is data-racy in presence of an FCV downgrade.