The footnote on exponential vs. log-normal service times is the part I'd push on.. in production I almost never see exponential, and heavy tails change the picture. Curious if you've looked at how robust this is under realistic distributions