A Large-scale Analysis of Hundreds of In-memory Key-value Cache Clusters at Twitter

Modern web services use in-memory caching extensively to increase throughput and reduce latency. There have been several workload analyses of production systems that fueled research in improving the effectiveness systems. However, coverage is still sparse considering wide spectrum industrial cache cases. In this work, we significantly further understanding real-world workloads by collecting traces from 153 clusters at Twitter, sifting through over 80 TB data, sometimes interpreting context business logic behind them. We perform a comprehensive analysis characterize based on traffic pattern, time-to-live (TTL), popularity distribution, size distribution. A fine-grained view different uncover diversity cases: many are far more write-heavy or skewed than previously shown some display unique temporal patterns. also observe TTL an important defining parameter working sets. Our simulations show ideal replacement strategy caches can be surprising, for example, FIFO works best large number workloads.