End-to-End Principle
A network design principle: functions that only the communicating endpoints can implement completely belong in the endpoints, and the network in between should stay simple and general.
Also known as: End-to-End Argument, End-to-End Arguments in System Design, E2E Principle, Dumb Network
Category: Software Development
Tags: networking, internet, software-architecture, principles, distributed-systems, systems-design
Explanation
The end-to-end principle comes from the 1984 paper "End-to-End Arguments in System Design" by Jerome Saltzer, David Reed and David Clark. Its argument: some functions, such as reliable delivery, encryption or duplicate detection, can only be done correctly with the knowledge that the two ends of a conversation have. Building them into the network in between can't make them complete. At best, it's a performance optimization.
**The classic example**:
You transfer a file from one computer to another. Every network hop can check and retransmit packets, and the file can still arrive corrupted: a disk error, a bug in the file transfer program, a faulty memory chip. The only way to be sure is an end-to-end check: compare a checksum of what was sent with what was stored. Once you have that check, the per-hop guarantees are no longer needed for correctness. They only reduce how often you retry.
**Consequences for the internet**:
- **Dumb network, smart endpoints**: IP only makes a best effort to move packets. TCP, running on the endpoints, adds reliability
- **Permissionless innovation**: A new application only needs new software at the edges. The web, email, video calls and peer-to-peer file sharing were deployed without asking network operators to change anything
- **Generality**: A network that doesn't know about applications can carry applications nobody has invented yet
**Under pressure**:
The modern internet bends the principle a lot. NAT devices, firewalls, content delivery networks, traffic shapers and inspection boxes all put intelligence in the middle. Each one solves a real problem, and each one makes the network harder to change (see Protocol Ossification). End-to-end encryption is partly a way to restore the principle. Net neutrality debates are its political form.
**Beyond networking**:
The same reasoning applies to any layered system. Put a function at the layer that has the knowledge to do it completely. Message brokers can't give you exactly-once processing on their own, so the application has to be idempotent. Checking a backup means restoring it, not trusting each step of the pipeline.
Related Concepts
← Back to all concepts