We present COLR-Tree, an abstraction layer designed to support efﬁcient spatio-temporal queries on live data gathered from a large collection of sensors. We use COLR-Treein a publicly-available sensor web portal to separate the concerns of sensor data management from the web portal application. COLR-Tree uses two techniques to optimize end-to-end latencies of users’ queries by minimizing expensive data collection from sensors. First, it uses a novel technique to effectively cache aggregate results computed over sensor data with different expiry times. Second, it incorporates an efﬁcient one-pass sampling algorithm with its range lookup to utilize cached data and compensate for occasional unavailability of sensors. We evaluate our implementation of COLR-Tree on SQL Server 2005 with a real, large workload from Windows Live Local. Our experiments demonstrate that COLR-Tree signiﬁcantly improves both the end-to-end query performance and the number of sensors accessed compared to existing techniques.