There are some hard to reproduce issues using the new in-memory storage of fetched files from the database.
The initial focus is to create a sample application that reproduces the problem, as the current python test is unreliable to reproduce this issue. Possibly a c++ app generating large amount of request quickly would be the way to go.