Reliability a̶n̶d̶ [or] Idempotence?

We’ve spent decades refining reliability patterns around HTTP: retries, idempotency keys, at-least-once delivery, and graceful failure handling. These patterns are so ingrained that many teams implicitly assume the transport will cooperate.

But a growing class of systems doesn’t live in that world.

DNS, DHCP, telemetry, game servers, IoT, industrial control, media pipelines, and emerging real‑time workloads still rely on UDP good reasons. The reality isn’t that UDP is unreliable (as is the common mantra). The problem is that most application developers never learned how to build correct systems on top of it.

This talk reframes reliability as an application-level concern rather than a protocol feature. We’ll explore how concepts like idempotency, deduplication, replay safety, and state modeling apply cleanly to message-oriented and connectionless systems and how serverless execution models force us to get these details right.

The result is a practical mental model for building UDP-backed systems that are observable, scalable, and correct, without turning them into accidental re-implementations of TCP.

As infrastructure shifts toward serverless, edge compute, and real-time workloads, developers are increasingly exposed to protocols and delivery semantics that HTTP was never designed for. The industry needs shared patterns (not folklore) for building correct and efficient systems on top of unreliable transports.