The test currently asserts that the percentage of sampled reads/writes is within 5% of some expected percentage. This assertion fails occasionally, and each time the percentage is just over 5% off the expected (we've seen .1-.3% off)- to decrease BF noise, we'll bump the expected percentage to 6, which is still reasonably low.