Idempotent HTTP requests: why a dropped timeout cannot be guessed
A service fires a state-changing request across an unreliable link, the TCP socket resets before the response arrives, and the retry loop fires again three seconds later. If the server already committed the first attempt, your retry just billed a contractor twice or minted duplicate ledger state. If it didn't commit and your client gave up, the payout silently vanished. A lost HTTP reply does not tell you what happened on the server.
Author: Escherbridge. Published . Last reviewed: , America/Denver.
RFC 9110 (HTTP Semantics, June 2022, §9.2.2) draws the exact protocol line: "A request method is considered 'idempotent' if the intended effect on the server of multiple identical requests with that method is the same as the effect for a single such request." Under the standard, "PUT, DELETE, and safe request methods are idempotent"—which is why a client can safely replay a PUT after a broken socket, whereas "A proxy MUST NOT automatically retry non-idempotent requests."
You can break this exact failure mode on purpose in our interactive Paid-Twice browser simulator, or read the post-mortem on timeouts in Paid twice, or never.
Where protocol idempotence stops and application keys begin
HTTP's §9.2.2 definition is strict about the intended state mutation on the server: logging each duplicate request attempt is fine ("a server is free to log each request separately"), provided the effect stays identical. Crucially, the RFC does not make POST idempotent, and it does not invent "idempotency keys" as a built-in protocol header. When APIs accept POST for payments or multi-step workflow triggers, the application layer itself must enforce deduplication using a stored client key and status check before mutating state.
RFC 9110 also adds an easily overlooked safety brake: "A client SHOULD NOT automatically retry a failed automatic retry." Endless exponential retries on ambiguous network errors turn localized blips into thundering-herd outages. Writing down the exact retry boundary between caller and receiver prevents future engineers from "fixing" timeouts by loosening proxy retries—which is why this decision belongs in architecture decision records.
How this maps to stateful workflow engines
In multi-step orchestration, an interrupted step has to reconcile against persisted state before acting again. AZOA's workflow model for idempotent requests and its companion pattern Settle once model exactly that reconcile-before-retry boundary.
Scope and boundaries
This guide summarizes RFC 9110 §9.2.2 and protocol retry hazards; it does not describe a live PrimusNeo payment rail, financial custody onboarding, or exploit mechanics. AZOA's public workflow sandbox is a simulation and does not move value.
A situation like this
A reverse proxy sees a 504 gateway timeout on an unkeyed POST checkout call and automatically replays it upstream. The original request actually committed 200ms after the proxy cutoff—leaving one customer charged twice because nobody respected the non-idempotent retry rule.
Sources
- RFC 9110, HTTP Semantics, June 2022, §9.2.2 Idempotent Methods. https://www.rfc-editor.org/rfc/rfc9110.html
Also in this section
- What is local-first softwarework that still has to exist when a connection drops.
- BIM and IFC file exchangea file that has to open later is a related honesty problem.
- AI accountability: NIST AI RMF and the GAO frameworka recommendation still needs a name on it.
- Architecture decision records: write down why it was decidedan interface or a dependency is often the thing being decided.
The workflow sibling is AZOA's idempotent-requests page. The modelled "once, or not at all" page is Settle once. AZOA's public site is a simulation and does not move value.
The useful next step is to read §9.2.2, or Paid twice, or never. If something on this page is wrong, write to us.