High-load cluster Migration from Redis to Apache Cassandra: Architectural Solutions and Performance Analysis

Abstract

The paper presents a case study of migrating a high-load system from Redis 6.2.x to Apache Cassandra 4.1.x in a configuration of two data centers (RF=3+3, 'LOCAL_QUORUM'). A reproducible load testing methodology (YCSBA, Zipf, 100 million keys, 1 KB write, warm-up/metering) and a comparison of p95/p99 latency and bandwidth are given, taking into account the configurations of the operating system, file structure, disk space and the configuration of the JAVA virtual machine. The results of failure tests for node loss and data center loss are shown. Operating procedures, including recovery and backup, are presented. Options for compromise optimization of the choice of data compaction strategy are discussed with the analysis of alternative approaches, such as the LCS uniform compaction strategy and the TWCS time series strategy. The influence of background compaction tasks and data synchronization between nodes in a high-load distributed data storage system on "latency tails" - temporary jumps in latency leading to system performance degradation under high loads is considered.