Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

0001 · Target the real NDN wire format, not a private dialect

Status: Accepted

Context

A clean-slate NDN stack could define its own packet encoding — simpler to implement, free to evolve. But NDN already has a published wire specification (NDN Packet Format v0.3) and a decade-old ecosystem of interoperating implementations: NFD and ndn-cxx (C++), python-ndn, NDNts (TypeScript), NDN-DPDK (Go), ndnd (Go), and a running global testbed. A private dialect would be an island.

Decision

Encode and decode the real NDN Packet Format v0.3 and NDNLPv2, byte for byte, and treat interoperation with the existing ecosystem as a correctness requirement rather than a nice-to-have.

Concretely, the type numbers in ndn-packet are the spec’s: Interest 0x05, Data 0x06, LpPacket 0x64, the signed-Interest TLVs 0x2C/0x2E, and so on. The NFD management wire codec in ndn-mgmt-wire is cross-checked against ndn-cxx’s tlv-nfd.hpp; strategy names carry NFD’s /v=N version convention; the trust-schema loader imports python-ndn’s LightVerSec binary format; and the NDNCERT and sync codecs are validated against recorded interop transcripts (testbed/transcripts/, including live pcaps).

Consequences

  • Positive: ndn-rs forwarders and apps can, in principle, join the existing testbed and talk to NFD/ndnd peers. The conformance surface is checkable against an external spec, not just internally consistent (see the conformance matrix).
  • Positive: the wire layer is a stable contract, which makes fuzzing and property-testing it worthwhile (they encode the spec’s invariants).
  • Cost: we inherit the spec’s quirks — evolvability critical-bit rules, non-minimal-encoding rejection, the URI naming conventions — and must implement them faithfully even where a private format would be simpler.
  • Cost: experimental/provisional TLV types must be chosen in the even (non-critical) range so stock forwarders ignore rather than reject them.

Alternatives considered

  • A private, simpler encoding. Rejected: it would forfeit the entire reason to build an NDN stack — interoperation with the data-centric ecosystem — for a one-time implementation saving.
  • Spec-compatible “mostly”. Rejected: partial compatibility is the worst of both worlds; a peer that almost interoperates fails in the field in ways that are expensive to debug.