How SingleStore Learns: Feedback Reoptimization Explained

This post introduces and explains Feedback Reoptimization (FR), a new capability within SingleStore's query optimizer. FR addresses the problem of stale statistics leading to suboptimal query plans by observing actual query runtime behavior (row counts, CPU, memory) and using this feedback to generate and validate improved execution plans. It contrasts FR with traditional statistics-based optimization, highlighting its effectiveness on skewed data. The post details the four stages of FR: observe, identify, replan, and validate, emphasizing the safety mechanisms in place to prevent regressions. Performance results on TPC-H and JOB benchmarks are presented, showing significant improvements, particularly on skewed datasets.