Stitcher: Learned Workload Synthesis from Historical Performance Footprints

Database benchmarking and workload replay have been widely used to drive system design, evaluate workload performance, determine product evolution, and guide cloud migration. However, they both suffer from some key limitations: the former fails to capture the variety and complexity of production workloads; the latter requires access to user data, queries, and machine specifications, deeming it inapplicable in the face of user privacy concerns. Here we introduce our vision of learned workload synthesis to overcome these issues: given the performance profile of a customer workload (e.g., CPU/memory counters), synthesize a new workload that yields the same performance profile when executed on a range of hardware/software configurations. We present Stitcher as a first step towards realizing this vision, which synthesizes workloads by combining pieces from standard benchmarks. We believe that our vision will spark new research avenues in database workload replay.