Massive video analytics systems, comprised of many densely-deployed cameras and supporting edge servers, are driving innovation in many areas including smart retail stores and security monitoring. To support such systems the challenge lies in collecting video footage in a way that maximizes end-to-end application goals, and scales this performance as camera density increases to meet application needs. This paper presents Spider, a multi-hop, millimeter-wave (mmWave) wireless relay network design that meets these needs. To mitigate physical mmWave link blockage, Spider integrates a low-latency Wi-Fi control plane with a mmWave relay data plane, allowing agile re-routing around blockages. Spider proposes a novel video bit-rate allocation algorithm coupled with a scalable routing algorithm that works hand-in-hand toward the application-level objective of maximizing video analytics accuracy, rather than simply maximizing data throughput. Our experimental evaluation uses a combination of testbed deployment and trace-driven simulation and compares against both Wi-Fi and mmWave mesh schemes that operate without Spider's algorithms. Results show that Spider is able to support camera densities up to 176% higher (gains of 2.76x) than the best-performing comparison scheme, allowing it alone to meet real-world camera density targets (4-250 cameras/1,000 sq. ft., depending on application). Further experiments demonstrate Spider's scalability in the presence of failures, with a 5.4-100x reduction in average failure recovery time.