Skip to content

Baseline benchmarks (anchor numbers)

This page is the canonical reference for Baseline benchmarks (anchor numbers).

Who this is for

  • Operators and engineers who want “order of magnitude” expectations before running their own load tests
  • Anyone who wants a reproducible harness to compare changes over time (signals, caches, data modes)

What you will get

  • One sample baseline benchmark result (reproducible demo dataset)
  • A command to reproduce the same run in your environment
  • A template to record your own numbers over time

Sample baseline (demo dataset, local Docker)

This is intentionally a small, reproducible baseline. Treat it as an anchor, not a promise: numbers depend heavily on hardware, Docker resource limits, dataset size, and enabled signals.

Environment

  • OS: Linux (x86_64)
  • CPU: AMD Ryzen 7 3700X (8C/16T)
  • Memory: 31 GiB

Dataset + config

  • Dataset: ./scripts/demo.sh (10 events ingested; 4 items in popularity; 4 co-occurrence rows)
  • Endpoint: POST /v1/recommend
  • k=20
  • Algorithm: RECSYS_ALGO_MODE=popularity
  • Data mode: RECSYS_ARTIFACT_MODE_ENABLED=true (manifest in local MinIO)

Result (2026-02-06)

From ./scripts/loadtest.sh (5000 requests, concurrency 50):

  • Throughput: 3557 rps
  • Latency: p50=12.8ms, p95=22.8ms, p99=39.3ms
  • Errors: 0 (200: 5000)

Reproducing these numbers

To reproduce the baseline measurements on your own hardware and dataset, follow: - How to reproduce the baseline benchmarks