Reference pilot

Prove the change on your own traffic.

A reference evaluation plan, not a customer success story. Replace assumptions with measurements from your traffic and hardware.

Starting architecture

Place Webship beside the existing web server. Keep the same upstreams, certificates, and representative routes while the current listener remains available for rollback.

Representative load

Replay a documented mix of static, proxy, and protocol traffic at increasing concurrency. Record response correctness, errors, CPU, memory, and tail latency.

Scorecard criteria

Set pass/fail thresholds before the test. Include functional parity and operational behavior, not only maximum throughput.

  • Correct responses and headers across every selected route
  • p95 and p99 latency for representative requests
  • Error rate and failure behavior under load
  • CPU use at sustained target load
  • Memory use and stability during sustained traffic

Rollback plan

Document the old listener, configuration snapshot, traffic split, and the signal that sends traffic back. Test rollback before production cutover.

Go or no-go

Promote only if the agreed gates pass. Keep the incumbent configuration and routing ready until the canary is complete.

This plan contains no measured customer outcome or guaranteed savings. Your team owns the final workload and risk decision.

Plan your own benchmark pilot

Build your pilot checklist

Choose the current proxy, protocol, traffic profile, and environment to create a checklist and JSON template. No result is predicted.

Pilot checklist

  1. Record a baseline on Nginx before changing routing.
  2. Validate the HTTP/2 path end to end.
  3. Replay representative API traffic.
  4. Use comparable resources in Bare metal.
  5. Agree on correctness, p95/p99 latency, errors, CPU, and memory gates.
  6. Keep an immediate rollback route and rehearse it.

JSON pilot template

{
  "schema": "webship-benchmark-pilot/v1",
  "currentProxy": "nginx",
  "protocol": "h2",
  "trafficProfile": "api",
  "environment": "baremetal",
  "gates": [
    "correctness",
    "p95Latency",
    "p99Latency",
    "errorRate",
    "cpu",
    "memory"
  ],
  "rollout": [
    "baseline",
    "shadow",
    "canary",
    "review",
    "rollbackOrPromote"
  ],
  "result": "not_measured"
}

The template records choices only. It does not estimate performance, savings, or production readiness.