The explosive growth of modern short-video platforms like WeChat Channels — approaching 1.4 billion monthly active users (MAU) while serving over one billion queries per second (QPS) — has exposed fundamental limitations in conventional content delivery network (CDN) architectures, particularly their inability to handle highly dynamic content popularity patterns. The traditional cache-centric approach proves to be catastrophic when faced with ephemeral content exhibiting a "million-item waist" distribution. This phenomenon results in severe cache pollution (manifesting as 5–40% origin miss ratios) and introduces prohibitive back-to-origin bandwidth costs. Our TritonFlow overcomes these challenges by introducing access point scheduling as a novel origin traffic compensator, which reduces the operational cost of back-to-origin fetches by enhancing CDN cache affinity. Through analysis of production traces (peak >1B QPS), we demonstrate that TritonFlow reduces scheduling fluctuations by 96% and decreases back-to-origin traffic caused by non-compulsory cache misses by 35%, all while maintaining sub-200ms fetch latency.

Data centers heavily rely on scale-out architectures like fat-tree, BCube and VL2 to accommodate a large number of commodity servers. Since the traditional electrical network is demand-oblivious and cannot perfectly respond to the bursty traffic generated by big data applications, a growing trend is to design demand-aware topologies via introducing the converters with adjustable optical links, instead of adding more wiring links. However, little is known today about how to fully exploit the potential of the flexibility from the converters: the joint optimization on adjusting the optical links inside the converters and the routing in the whole network remains algorithmically challenging. In this paper, we design a set of customized converters for Diamond, VL2 and BCube topologies and initiate the optimization study in hybrid switching data center networks. As a case study, we introduce demand-aware load balancing problem (DLBP), i.e., a joint optimization on the physical layer (how the optical links interconnect inside the converter) and the network layer (how to determine the route especially for elephant flows in the whole network). We prove that DLBP is not only NP-hard, but also$\rho $-inapproximation. Accordingly, we design two algorithms: the first one is an intuitive greedy algorithm and the second one uses reinforcement learning to improve upon the solution of the first one. Trace-driven evaluations show that our algorithms can reduce the traffic congestion by 12% on average.