In a sharded cluster, when listDatabases is executed, the mongoS secuentially sends a listDatabases command to every shard of the cluster.
Currently there is no mechanism to prevent dropping existing databases or creating new ones in the middle of a listDatabases process.
Therefore, the databases view returned to the user may be inconsistent. For instance, we could return a pair of databases that have never co-existed.

You can find a reproducible in 11e4b7dd

The proposed solutions are the following:

1. Read a consistent view from the config server

• First fetch a consistent snapshot from config.collections and config.databases.
• Then, build the listDatabases response with that information.
• Finally, fetch the size of each database for each shard by running listDatabases across shards.

2. Another solution would be to execute listDatabases to all shards using a snapshot read at a specific cluster time. However, listDatabases doesn't support snapshot reads right not, so, unless we enable them somehow, this is not a real option.