Teaching the Cloud to Remember Tomorrow: Using Graph-Transformer AI to Pre‑Warm Caches before the Traffic Surge Hits

Abstract

Cloud caching systems today are surprisingly short‑sighted. When traffic spikes, auto‑scaling adds new instances, but those instances start with empty caches like opening a new store with no inventory. The result is a flood of expensive origin fetches, even if the overall ssystem has enough compute power. This paper asks a simple question: what if a cache could learn what content will be needed, seconds before it is actually requested? We propose a hybrid AI framework that combines a graph neural network (GNN) to capture relationships between content objects (e.g., “users who viewed A next requested B”) with a temporal transformer to model access patterns over multiple time scales. Using this architecture, the system predicts a future cache working set a short list of likely‑to‑be‑requested items during the unavoidable monitoring lag between traffic detection and new instance boot. Instead of waiting for a cache miss, we pre‑warm the new instance with only those predicted objects. Evaluated on real‑world CDN traces, our method reduces cold‑start miss rates by up to 42% and cuts data transfer costs during scale‑out events by nearly a third, compared to reactive caching policies. More importantly, the approach turns monitoring lag from a liability into a window of opportunity. The cloud no longer just reacts; it remembers tomorrow.