Dynamic network conditions, such as traffic shifts, device failures, router misconfigurations, and implementation bugs can degrade the performance of end-to-end paths. With the increasing amount of traffic generated by real-time applications (for example, IPTV, VoIP, VoD), operators need to quickly detect, diagnose, and correct degradations in order to avoid storms of calls from unhappy customers. Unfortunately, existing fault monitoring techniques and systems are disconcertingly ad hoc and operators have trouble detecting faults before their customers do. This talk argues that measurements taken at end-hosts are essential for fault detection and identification. We study the practical issues that arise when using network tomography to monitor failures.