What assumptions does the BH procedure actually need?

The original proof assumes independence (or a form of positive dependence, PRDS) among test statistics. In practice tests are often correlated in messy ways. How worried should practitioners be, and when is the BY (more conservative) correction worth the lost power?