In sharding, dropping the database does not go through database versioning. So if you have more than 1 mongos and are round robining requests (in the way all drivers do) any mongos that did not explicitly send the drop or create will not discover that the database exists, thus sending incorrect results. For example after inserting data and then performing a count, you'll see the count is 0. See this ticket for details: HELP-9914

We're working on potential solutions on the server-side, but I want to update all the transactions examples that we'd like to reference in the general public beta.

I do not believe the node examples have this issue: https://gist.github.com/mbroadst/2a719a7adeb8b0476f50ac5643d70166

Python does: https://github.com/mongodb/mongo-python-driver/blob/3.7.0b0/test/test_examples.py#L849

Java also does: https://gist.github.com/jyemin/7031cc235759e4619ab0ab342f22db37#file-betaexamples-java-L53

Changing Java to the below works

       database = client.getDatabase("test");
        // create the collections in advance so there is no attempt to implicitly create them within the transaction
        database.createCollection("inventory");
        database.createCollection("shipment");

        inventory = database.getCollection("inventory");
        shipment = database.getCollection("shipment");

        inventory.deleteMany(new Document());
        shipment.deleteMany(new Document());