Compared with its 2D counterpart, the scalability problem is greatly exacerbated in a 3D wireless sensor network. In this paper, we propose a scalable routing algorithm, dubbed Bubble Routing. It preprocesses global knowledge via a distributed algorithm, such that a node only needs to store a small constant information to make correct and efficient local routing decisions and achieve guaranteed delivery at the same time. More specifically, the proposed bubble routing algorithm first decompose a 3D network into a set of hollow spherical cells (HSCs). A continuous and one-to-one mapping is applied and a virtual tree structure is established inside each HSC to enable greedy routing. On the other hand, routing across HSCs is guided by a small routing table whose size is bounded by the number of interior holes. Our simulation results show that bubble routing can achieve guaranteed data delivery, low stretch factor, and well balanced traffic load.