Summary
- Danny Cohen's IEN 137 showed that big-endian and little-endian can each be internally coherent; failure begins when a public interface leaves order or grouping implicit.
- The durable settlement was not a universal victory for one camp. It was a governance rule: one declared representation per shared format, explicit conversion at the edge and tests that preserve the meaning of the original value.
The number that changes without changing
Imagine an incident review with only a four-byte hex dump on the screen. One engineer reads the first byte as the most significant part of an integer. Another starts from the opposite end. Both decoders are faithful to their own machines. Both return a valid number. Only one can match what the sender meant.
That is the compact problem behind Danny Cohen's April 1980 Internet Experiment Note 137, On Holy Wars and a Plea for Peace. The memo is famous for borrowing the Big-Endian and Little-Endian names from Swift's quarrel over how to open an egg. Its more consequential move is analytical. Cohen refuses to treat a familiar local representation as a law of nature. He separates memory order from transmission order, byte order from bit order, and the sequence itself from the size of the chunks used to inspect it.
The distinction matters because a network joins machines that were not designed together. Inside one computer, an implementation can make a coherent choice and live comfortably with it. Once a value crosses an interface, however, private comfort becomes a public claim. The receiver needs to know which unit was serialized, in which direction and under which conversion rule. Without that agreement, the bytes survive while the value does not.
When the chunk becomes part of the protocol
IEN 137 describes two fully consistent ways to serialize a structured value. If the order remains consistent across bit, byte, word and larger boundaries, observers may group the same sequence differently without silently reversing parts of it. A stream can be viewed as bits, bytes or words and still retain one direction of significance.
Mixed conventions are more dangerous. A system may number bits one way inside a byte, arrange bytes another way inside a word and then transmit words under a third convention. In that case, “first” no longer has a stable meaning. The chunk size becomes hidden protocol state. Two endpoints can honestly agree on the bytes and disagree on how they compose.
Cohen notes that ARPA Internet datagrams were handled under a consistent Big-Endian assumption. That historical choice is important, but it is not the philosophical conclusion of the memo. The conclusion is that logic alone cannot crown a winner. What protects communication is an order known to all parties. The public convention wins only in the modest sense that everybody can implement against it.
This is why “network byte order” is best understood as an interface contract, not as praise for one processor family. A little-endian host does not become less legitimate when it converts before sending. The conversion is precisely what allows internal design freedom to coexist with a common wire image.
Opposite formats, same discipline
Two later specifications make Cohen's point more clearly than a slogan can. RFC 2083 defines the PNG image format and places multi-byte integers in network order, with the most significant byte first. RFC 1951 defines DEFLATE and stores multi-byte quantities with the least significant byte first, while separately spelling out how bits are packed.
Those choices point in opposite directions. Both work because each format says exactly what its octets mean. A PNG decoder must not substitute the host's native order. A DEFLATE implementation cannot invoke network convention to reverse the format it was given. Correctness sits at the named boundary.
This contrast also exposes a common engineering error: carrying an in-memory structure straight onto the wire. Compiler padding, native integer width, alignment and host order are not portable semantics. A durable format assigns widths, positions and order independently of the machine that happens to encode it. The encoder produces that public representation; the decoder reconstructs a local one. Those two transformations are part of the protocol, even when a convenient library makes them disappear.
An exception needs a fence
RFC 1071 offers a useful warning against overgeneralization. For the Internet checksum, a particular algebraic property permits the sum to be computed in either byte order, provided the swapped result is handled correctly. That is a bounded invariant of that checksum calculation. It does not make arbitrary headers or payload fields order-independent.
Good specifications identify such invariants precisely. Bad implementations remember only the comforting half of the sentence. If code relies on a byte-order exception, the exception needs a named operation, a test vector and a clear exit back to the canonical representation. Otherwise an optimization becomes folklore, and folklore eventually becomes a corrupted record that no longer reveals which interpretation was intended.
The cost of letting both sides choose
In 1999, Chris Newman revisited the subject in an individual Internet-Draft about network byte order. The document expired and has no formal standing in the IETF standards process, so it should not be cited as a rule. Its operational observation nevertheless remains sharp: selecting byte order at run time creates two variants of a protocol or media type, and both variants need independent testing. In practice, one is often neglected.
Negotiation therefore spends more than a flag bit. It multiplies parsers, fixtures, failure modes, security review and forensic ambiguity. If peers can choose order per session, caches and signed records must bind that choice to the bytes. If a file can declare either order, readers must exercise both paths against malformed and boundary values. Optionality is support debt with a long tail.
A single fixed order is not always possible. Existing formats, zero-copy constraints or hardware interfaces may make plurality rational. Cohen's framework still applies: make the selection explicit, scope it to one boundary and keep the conversion receipt. The dangerous state is not diversity; it is diversity that has lost its label.
The engineer behind the metaphor
The Internet Hall of Fame credits Cohen with coining the computing terms Big Endian and Little Endian and also records his early packet-video and packet-voice work. USC's memorial places him within the long research life of the Information Sciences Institute. That wider career helps explain the force of IEN 137. Real-time traffic makes representational mistakes observable: a wrong interpretation is not an abstract philosophical flaw but a failed exchange between independently built systems.
The egg satire gave the memo its memory. The engineering contribution was the peace treaty beneath the joke. Local architectures could remain different, provided the shared surface became exact enough that difference did not leak into meaning.
Sources
- IEN 137: https://www.ietf.org/ietf-ftp/rfc/ien/ien137.html
- IEN index: https://www.ietf.org/ietf-ftp/rfc/ien/ien-index.html
- Expired individual Internet-Draft: https://datatracker.ietf.org/doc/html/draft-newman-network-byte-order-01
- RFC 2083, PNG: https://www.rfc-editor.org/rfc/rfc2083.html
- RFC 1951, DEFLATE: https://www.rfc-editor.org/rfc/rfc1951.html
- RFC 1071, Internet checksum: https://www.rfc-editor.org/rfc/rfc1071.html
- Internet Hall of Fame biography: https://www.internethalloffame.org/inductee/danny-cohen/
- USC memorial: https://viterbischool.usc.edu/news/2019/08/in-memoriam-danny-cohen-distinguished-computer-scientist-and-long-time-isi-staff-member-dies-at-81/
- Public portrait reference: https://www.internethalloffame.org/wp-content/uploads/2012/04/Danny-Cohen.jpg
Member Briefing
Deeper Profile Context
Sign in with the right membership level to unlock the full briefing and source notes.
Only for Strategic Circle
Strategic Circle
Open to all readers. Unlock profile briefings after joining and signing in.
Join Strategic CircleOnly for Leadership Alliance
Leadership Alliance
For qualified IP-asset owners and management; sign in to unlock alliance briefings.
Join Leadership Alliance
