Investigation // active

API latency under load

Three providers, one data pipeline, and a controlled stress test to find where the latency curve breaks. Results are unfiltered.

3

providers tested

14d

run duration

~2M

requests logged

Hypothesis + method

The setup

Instrumentation ran through a thin Go proxy that stamped each request with provider ID, timestamp, and response code before writing to a local ClickHouse instance. No SDK wrappers — raw HTTP only, so the numbers reflect wire latency, not client overhead.

The hypothesis: provider-advertised p99 latency numbers collapse under sustained concurrent load. Each provider was hit with identical payloads at 50, 200, and 500 requests per second across a 14-day window.

Final verdict

Provider C breaks at 200 rps. The others hold — barely.

Advertised p99 figures held at 50 rps. Above 200, Provider C degraded to 4× its baseline. Providers A and B stayed within spec but showed measurable drift past 400 rps. Raw logs and the proxy source are in the repo.