Philipp Kahr and Grzegorz Banasiak discuss the use of Rally for benchmarking Elasticsearch, emphasizing the importance of defining clear goals before conducting tests. Rally should not be run on production clusters due to its potential for destructive actions and data loss. The authors provide a detailed analysis of an ingest-only benchmark test run on three Elasticsearch nodes, outlining various metrics such as store size, throughput, and percentile latencies. They highlight the importance of focusing on ingest-related metrics, caution against using maximum throughput as a reliable measure, and underscore the distinction between latency and service time. The blog also delves into the interpretation of telemetry data and the relevance of different metrics depending on the use case, while advising on optimizations and considerations for accurate results. The post concludes with a look at the potential implications for cluster performance and the importance of examining CPU and other system usages in future analyses.