Skip to content

BenchForge

Research-grade, scenario-based database benchmark platform for DB researchers and engineers.

Compare different DB access stacks - driver, ORM, language - under identical workloads with statistical rigor suitable for academic publication (VLDB, SIGMOD, OSDI) and professional engineering evaluation.

:material-rocket-launch: Get started in minutes :material-github: View on GitHub

Feature Highlights

  • :material-chart-line: Statistical rigor

Run multi-iteration experiments with bootstrap confidence intervals instead of single-run guesswork.

  • :material-recycle-variant: Reproducibility by design

Use seed control, setup/teardown isolation, and full environment capture for defensible results.

  • :material-file-chart: Publication-quality output

Generate ECDF plots, CI error bars, and paper-ready reports with colorblind-safe visual defaults.

  • :material-compare-horizontal: Apples-to-apples comparison

Benchmark the exact same workload across drivers, ORMs, and language stacks.

  • :material-timer-outline: High-fidelity latency tracking

Record latency distributions with an HDR histogram implementation tuned for benchmark workloads.

  • :material-cog-sync: Scenario-driven workflows

Encode workload, load profile, experiment controls, and targets in one versionable YAML scenario.

Quick Install

pip install benchforge

Then follow the Quick Start guide to run your first benchmark and generate a report.

Why BenchForge?

Most benchmark scripts are ad-hoc and difficult to reproduce. BenchForge is purpose-built for trustworthy comparisons:

  • Multi-iteration measurements with statistical confidence
  • Controlled experiment design via scenario DSL
  • Repeatable runs with deterministic seed behavior
  • Result artifacts designed for both academic and engineering review

If you are new to BenchForge, start with Quick Start, then read Concepts and Methodology.