What this tests

retry-logic

Point your client at this specimen and find out — before prod does — whether it backs off, retries, and knows when to stop.

Anatomy

An AI provider at capacity: 503s dressed in the house error format, a body that says "overloaded" with the serene detachment of a maître d' turning away the public. No Retry-After worth the name, no queue position, just vibes and a suggestion to try again later — later being a concept, not a timestamp. The trap is the rage-loop. Every retry is another request into a service that's already drowning, so your client's survival instinct is literally the thing making it worse — like a drowning man whose flailing is billed per stroke. Exponential backoff or circuit-break, anything else is a donation. Clients that treat 503 as "wait and try again politely" survive. Clients that treat it as "the button didn't work, press it harder" become part of the load they were complaining about. There's a Zen koan in there somewhere, and it's on fire.

Source: OpenAI API — error codes

Replay it

curl -i https://faultydata.com/r/FD-2026-009

// response appears here

Signature

{Status:503 Headers:map[Retry-After:15] Body:{"error":{"message":"The engine is currently overloaded, please try again later","type":"server_error","code":"overloaded"}} Corrupt: Delay:<nil> Flaky:<nil> Sequence:<nil> Drip:<nil> StallAfter:0 Hang:false Reset:false Raw: Phases:[]}

Keep digging

FD-2026-001 — Stripe-style rate limit FD-2026-004 — GitHub secondary rate limit FD-2026-005 — Token expires mid-session (Entra ID pattern) FD-2026-027 — Signing-key rollover breaks token validation (Azure AD, Mar 2021) FD-2026-032 — Capacity-collapse 503s (Healthcare.gov launch, Oct 2013) FD-2026-033 — Sudden platform rate limit (Twitter/X API, 2023)