Urban water supply network is ubiquitous and indispensable to city dwellers, especially in the era of global urbanization. Preventative maintenance of water pipes, especially in urban-scale networks, thus becomes a vital importance. To achieve this goal, failure prediction that aims to pro-actively pinpoint those 'most-risky-to-fail' pipes becomes critical and has been attracting wide attention from government, academia, and industry. Different from classification-, regression-, or ranking-based methods, this paper adopts a point process-based framework that incorporates both the past failure event data and individual pipe-specific profile including physical, environmental, and operational covariants. In particular, based on a common wisdom of previous work that the failure event sequences typically exhibit temporal clustering distribution, we use mutual-exciting point process to model such triggering effects for different failure types. Our system is deployed as a platform commissioned by the water agency in a metropolitan city in Asia, and achieves state-of-the-art performance on an urban-scale pipe network. Our model is generic and thus can be applied to other industrial scenarios for event prediction.