Consider the following sequence:

1. A client requests a dropDatabase(dbName) to a sharded cluster
2. The DDL request is served by the primary shard, until the point on which all the collections under dbName get removed from the local catalog; secondary nodes replicate the collection removals;
3. Before reaching the stage on which the dropDatabase op entry is written onto the oplog, the primary shard receives a shutdown (for restart) signal, so that the DDL is interrupted;
4. Upon restart, the newly spawned process is elected as the RS primary
5. As part of the initialisation routine, the newly spawned process fills up its DatabaseHolder (the in-memory cache of existing database names, according to the collection names retrieved from the local catalog): due to step 2, dbName is no longer present;

When the sharding DDL coordinator resumes the execution of dropDatabase(dbName), the absence of dbName from DatabaseHolder causes the function in charge of executing step 1 to early exit, skipping the instruction to generate the op entry about the completion of the DDL on the local node.