Engineers at Google developed linear elastic caching to solve the inefficiency of fixed-size memory allocation. Traditional caches often waste money by over-provisioning for peak demand or hurt performance by under-provisioning. This new approach treats memory as a variable utility, dynamically adjusting size based on real-time workload costs rather than static limits.

By applying the "ski rental" logic, the system decides whether to keep data in RAM or evict it to save costs. A shallow decision tree implemented in C++ predicts the optimal time-to-live for each page. This lightweight model considers data size and miss costs to balance memory footprint against I/O penalties.

Integration into Spanner showed a 15.5% reduction in memory usage with only a minor increase in cache misses. Because the algorithm targets cheap-to-fetch data for eviction, actual I/O costs rose by only 0.5%. This shift lowered the total cost of ownership by roughly 5% across production servers.

Testing against public benchmarks confirmed that elastic policies outperform fixed-size caches across diverse workloads. Savings increase as the relative cost of memory rises compared to cache miss penalties. The system proves that cost-aware eviction policies maintain high performance while cutting infrastructure spend.