What this tests
timeoutsretry-logic
Point your client at this specimen and find out — before prod does — whether it has a timeout, and whether the timeout has a timeout.
Anatomy
Your request reached the load balancer, the load balancer reached your instance, and your instance took longer than sixty seconds to respond — so the LB, who has seen this film before and knows how it ends, cuts in with a 504 on the application's behalf. The backend may still be computing. It doesn't matter. The verdict is in. The client-visible signature: a correct 504 arriving long after the client expected anything, possibly long after the client's own timeout fired — a response to a conversation the client already left, like a waiter bringing dessert to a table that paid and left in February. Clients that assume "no response means no response" get surprised when the 504 shows up at T+60s. Clients that gave up and retried at T+10s have now submitted the request twice, which is how a slow query becomes a slow query running twice. There's always a second specimen hiding inside the first.
Source: AWS docs — HTTP 504 Gateway Timeout from Classic/Application LBs
Replay it
curl -i https://faultydata.com/r/FD-2026-049
// response appears hereSignature
{Status:504 Headers:map[Server:awselb/2.0] Body:<html><body><h1>504 Gateway Time-out</h1></body></html> Corrupt: Delay:0x69e720b63570 Flaky:<nil> Sequence:<nil> Drip:<nil> StallAfter:0 Hang:false Reset:false Raw: Phases:[]}
Keep digging
FD-2026-060 — Network congestion cascade (Google Cloud, Jun 2019) FD-2026-088 — DNS provider drowns in DDoS (Dyn, Oct 2016) FD-2026-093 — Routing-table overload blackout (Optus, Nov 2023) FD-2026-097 — Trading halt by infrastructure (Robinhood, Mar 2020) FD-2026-001 — Stripe-style rate limit FD-2026-004 — GitHub secondary rate limit