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
When Slack went down in 2021 and shambled back to life, the recovery was the interesting part: services didn't flip from dead to alive, they staggered upright like a junior developer after the Christmas party — vertical, technically, but in no condition to make promises. First request: 503. Second: 503. Third: also 503. Fourth: fine, actually, what did you want. This is where half-open circuit breakers go to be tested. Retry once and surrender? Failure. Retry at a fixed 1ms interval? Congratulations, you've built a hammer that beats on a healing service — the medical equivalent of CPR administered with a mallet. Exponential backoff with jitter, three or four attempts? That's an adult in the room. The sequence is deterministic — three refusals then success — because real recovery is deterministic-ish too. The service IS coming back. It's just doing it at its own pace, and it has never once consulted your pager.
Source: Slack — 2021-01-04 incident summary · 2021-01-04
Replay it
curl -i https://faultydata.com/r/FD-2026-007
// response appears hereSignature
{Status:0 Headers:map[] Body: Corrupt: Delay:<nil> Flaky:<nil> Sequence:0x69e720a52588 Drip:<nil> StallAfter:0 Hang:false Reset:false Raw: Phases:[]}
Keep digging
FD-2026-001 — Stripe-style rate limit FD-2026-002 — CDN edge degradation (Fastly, June 2021) FD-2026-004 — GitHub secondary rate limit FD-2026-005 — Token expires mid-session (Entra ID pattern) FD-2026-009 — Provider overloaded (OpenAI-style 529) FD-2026-021 — S3 us-east-1 subsystem removal (AWS, Feb 2017)