khone/koʊ.neɪ/
Experimental

Benchmark methodology

Benchmarks compare Khone gateway routes against a standard API Gateway HTTP API baseline under the same workload shape.

For public result numbers and charts, start with Benchmark results. This page explains how those reports are produced and how to read the measurements.

Endpoints

  • steady: Mode A proxy-layer route with a fixed bounded wait.
  • adaptive: Mode A proxy-layer route with rate-aware adaptive wait behavior.
  • target-aware: Mode A proxy-layer route with duration probe-aware wait behavior.
  • standard: API Gateway HTTP API plus Lambda baseline.

All four endpoints use target handlers that call the same benchmark backend Lambda URL. The backend simulates a delayed downstream service: for the public snapshot it sleeps for an 80 ms base delay plus up to 80 ms of item-key-seeded jitter before returning JSON points. That backend delay is separate from the k6 max-delay query value, which was 0ms in the public reports.

This intentionally models an I/O-bound handler, where much of the request duration is spent waiting for another service. The benchmark is less representative of CPU-bound handlers, where each item mostly consumes compute and batching cannot hide wait states.

The public snapshot used 256 MB target handlers on arm64. The Khone gateway ran on LMI with a capacity provider configured with m8g instances, 2048 MB gateway memory, 64 concurrent requests per execution environment, 2.0 GiB/vCPU, and four warm execution environments during the final pass.

Measurements

k6 records client-observed http_req_duration, status codes, endpoint labels, debug batch headers, and target elapsed headers when present.

The reports derive:

  • request counts and error rates
  • latency percentiles
  • effective batch size
  • estimated target invocation count
  • estimated relative cost

Cost estimates are not bills

The estimate uses observed batch size and router-measured target elapsed time when available. It is a workload-efficiency proxy, not Lambda billed duration. Real cost depends on target runtime, memory, architecture, request duration, and AWS pricing. It also excludes gateway LMI capacity cost, Function URL/API Gateway charges, CloudWatch, networking, and other account-level charges.