Re-preparing on UP might fail in the background when the selected connection is not using the correct keyspace yet.
This causes an exception to be logged and the query not being prepared on the newly UP host. As expected, there is no failure yielded to the user.
This mechanism was introduced in CSHARP-506.
I was able to reproduce it in the endurance tests (using .NET Core 2.1):
merged to OSS
merged to dse