We present a traffic analysis of the ADU anonymity scheme presented at ESORICS 2008, and the related RADU scheme. We show that optimal attacks are able to de-anonymize messages more effectively than believed before. Our analysis applies to single messages as well as long term observations using multiple messages. The search of a "better" scheme is bound to fail, since we prove that the original Crowds anonymity system provides the best security for any given mean messaging latency. Finally we present D-Crowds, a scheme that supports any path length distribution, while leaking the least possible information, and quantify the optimal attacks against it.