RFC 4838 describes the architecture for Delay-Tolerant Networking (DTN), a framework for networks where the comfortable assumptions of TCP/IP simply do not hold: continuous end-to-end paths, low round-trip times, low loss rates, and symmetric data rates. When Vint Cerf — yes, that Cerf — and a team that included JPL engineer Scott Burleigh sat down to design this, they were thinking literally about interplanetary communication. Light from Mars takes 4–24 minutes to reach Earth one-way. TCP's three-way handshake becomes absurd.

The core problem: the Internet protocol stack assumes that if a packet can't be delivered now, it should be dropped. Routers don't have meaningful storage. End-to-end ACKs run on millisecond timers. None of this works when:

• A spacecraft passes behind a planet for an hour.
• A sensor on a buoy uplinks only when a satellite is overhead.
• A village's internet kiosk syncs via a "data mule" motorcycle that visits weekly.
• Battlefield radios are jammed for unpredictable windows.

The design move: introduce an overlay layer above the transport called the Bundle Protocol (specified in the companion RFC 5050, later RFC 9171). Bundles are large, self-contained application data units carrying their own routing metadata and lifetime. Intermediate nodes practice store-carry-forward: if the next hop isn't reachable, persist the bundle to non-volatile storage and forward when a contact opportunity arises, possibly minutes, hours, or days later.

Several design decisions are worth dwelling on:

• Custody transfer. Reliability cannot rely on end-to-end ACKs over hour-long RTTs. Instead, a node can take "custody" of a bundle, accepting responsibility for its onward delivery, similar to a postal service relay. This pushes retransmission state closer to the failure.
• Late binding of names. DTN endpoints use URI-like Endpoint Identifiers (e.g., dtn://mars-rover-1/telemetry). Resolution to a transport-level address can be deferred until a bundle nears its destination — useful when topology evolves over the bundle's multi-hour journey.
• Time synchronization is assumed. Bundle lifetimes are absolute timestamps, so DTN nodes need reasonable clock sync — a non-trivial assumption that the architecture explicitly calls out.
• Asymmetric and unidirectional links are first-class. Unlike TCP, the protocol does not require a return path of comparable bandwidth, or any return path at all in some configurations.

Why it matters today: DTN powers NASA's deep-space communications. The Bundle Protocol runs on the ISS and was used in the Deep Impact spacecraft experiments in 2008. NASA's Lunar Gateway and Artemis architecture rely on DTN. But the ideas have leaked into terrestrial engineering too: opportunistic networking for disaster response, vehicular ad-hoc networks, IoT systems with intermittent uplinks, and even mobile messaging apps that queue messages locally and sync on connectivity. If you've ever designed a mobile app's offline-first sync layer with idempotent operations and conflict-free merging, you've reinvented a piece of DTN.

Interesting history: The work began in the late 1990s as the InterPlanetary Internet (IPN) Special Interest Group, then broadened. Cerf — fresh off being declared "father of the internet" — pivoted to designing a protocol stack for off-world use. The IRTF DTN Research Group's draft architecture became RFC 4838, and the IETF's DTN Working Group has continued evolving it (BPv7 in RFC 9171, 2022). It's one of the rare protocol families designed with explicit awareness that the speed of light is a real engineering constraint.