We try to keep the maximum runtime of PR tests to approximately 25 minutes, and recently with unit-test-bucket* tests on Windows have crept up to 35 minutes. This is due to two reasons:

1. When we add new tests the bucketing process moves tests from one bucket to another. Through bad luck this can lead to one bucket having a lot of long running tests and another having only short tests. There's no good way to predict the impact of adding a new python test on PR test runtimes
2. Our python tests can't be run in parallel on Windows. We implement concurrency using the third part tool concurrencytest, however this uses os.fork() which only works on Unix platforms

This ticket is twofold - Firstly to get PR test times down under 25 minutes again, and secondly to identify long term fixes for this issue.