We have been seeing this issue quite frequently that on the primary replica of one of the replica sets in our sharded cluster, the disk usage and disk IO utilization just shoots up.

This usually happens under situation of heavy write load. Once the load goes away, the IO utilization comes down but disk storage utilization just stays up. Sometimes, the size is about 10 times the actual size of the database i.e. the size of the secondary replicas. When investigated further, we found that the WireTigerLAS.wt file had grown unbounded and never recovered even when we took away all the write load.

The situation has become more prominent after we upgraded from 3.4 to 3.6, but this could be just because we have more write load since then.