OPTIMIZATION OF SERVER CACHING USING ADAPTIVE TTL BASED ON MACHINE LEARNING

Abstract

Static definition of time-to-live (TTL) for cached server data often leads to inefficient resource usage or serving stale information. An approach for caching optimization through dynamic TTL determination using machine learning is proposed. Two main methods are considered: batch training of a model (e.g., gradient boosting) based on historical data about source change frequency, access patterns, and data characteristics; and online learning, particularly using reinforcement learning (RL), which allows the system to continuously adapt to current conditions. The RL agent learns to select the optimal TTL by analyzing positive (cache hits) and negative (misses, stale data) outcomes of its actions. Adaptive TTL is expected to increase the cache hit rate, reduce latency and database load, while ensuring higher data freshness.