I observed that these tests are slow because we wait after the split until the parent has been compacted away. It's not necessary to do so (unless we wanted to test that specific HBase logic and HBase already has many tests for that). We can change the test to only wait until the daughter regions are created. When they exist they will be used.
Avoiding that waiting cuts the runtime of each test from 17mins to 2min (on my machine) so it saves at least 30mins of runtime (and more on the slower jenkins machines)