||Peer-to-peer (P2P) streaming promises scalable streaming to large group. In order to achieve stream robustness against node churns, a streaming mesh is often used. Traditionally, such mesh is constructed in a rather ad-hoc manner, where peers connect to multiple parents through a gossip mechanism. Because stream delay is given by the longest path of the packets, the mesh constructed in this way is often of high delay. In this work, we address how to design and optimize a mesh to achieve and robustness to peer churns. We first formulate the mesh optimization problem which is to minimize worst-case user delay. The problem is shown to be NP-hard, and hence we propose a centralized heuristic which is applicable to small centrally-controlled network and serves as a benchmark. We then propose a novel distributed algorithm called RMesh, which adapts to network dynamic by continuously reducing delay. Using simulation and real Internet from PlanetLab, we show that RMesh indeed achieves low delay and is robust against node churns by offering high video continuity. Moreover, there is an optimal number of parents to make good use of peer bandwidth to achieve minimal delay.