Expertise on distributed systems is critical for system maintenance and improvement. However, it is challenging to keep the up-to-date knowledge from distributed systems due to the complexity and continuous updates. Hence, computing platform providers study on how to extract knowledge directly from system behavior. In this paper, we propose a methodology called KEREP to automatically extract knowledge on distributed system behavior through request execution path. Technologies are devised to construct component structures, to depict the in-depth dynamic behavior and to identify the heartbeat mechanisms of target distributed systems. Experiments on two real-world distributed systems show the KEREP methodology extracts accurate knowledge of request processing and discovers undocumented features with good execution performance.