AdaptCacheAI: Adaptive Hybrid Caching with Machine-Learned Eviction for Dynamic Cloud Workloads

Abstract

Dynamic cloud workloads bring challenges such as high variability, multitenant interference, and heterogeneous access patterns, which make it necessary to have effective cache eviction to retain low latency and cost efficiency. Conventional eviction algorithms like LRU, LFU, ARC, and W-TinyLFU, which are based on static heuristics, give good results under stable conditions but their performance decreases significantly when request patterns change rapidly or differ across applications. AdaptCacheAI, to solve these problems, features an adaptive hybrid caching framework that combines machine learning–driven eviction with a multi-tier cache architecture going from memory to SSD layers. A predictive eviction model that, in real time, calculates the probability of reuse of an object is at the center, supported by a workload classification engine that is capable of recognizing patterns like temporal hotspots, bursty traffic, and tenant-specific behaviors. The continuous feedback loop that accounts for signals such as latency, eviction regret, and tier pressure refurbishes the decisions and thus allows AdaptCacheAI to modify policies on a dynamic basis. Tests on diversified traces confirm the substantial improvements that traditional algorithms can make through the employment of AdaptCacheAI, such as increased hit rates, decreased tail latencies, and lower SSD write amplification, all resulting in tangible performance gains and cost savings. AdaptCacheAI serves as a demonstration of the feasibility of employing machine learning for intelligent cache management and also it paves the way for the potential of future enhancements such as reinforcement learning–based optimization and cross-layer coordination in cloud ​‍​‌‍​‍‌systems.