2ndOpinion.FYI中文
genius.wiki

#882 1983 · USC Information Sciences Institute (Paul Mockapetris) · Computer networking

Every computer on the ARPANET lived in one hand-edited file — until naming was delegated

the problem

Every host lived in one global table edited by one institute, and the update load was nearing the limit of manageability

background

Through the early 1980s, every host on the ARPANET was registered with the Network Information Center at SRI and listed in one global table — the file HOSTS.TXT, kept on a single SRI host. Site administrators asked the NIC for a name, the staff edited the file, and every machine on the network periodically downloaded the whole thing. Mockapetris's own 1983 memo states the bind in one line: the size of this table, and especially the frequency of updates to it, were near the limit of manageability.

The instinctive fix was to scale the centre — better staff, faster distribution, a bigger NIC. Paul Mockapetris, working in Jon Postel's group at USC's Information Sciences Institute, went the other way: instead of making the middle stronger, he published RFC 882 and RFC 883 in November 1983, proposing that the middle stop holding the data at all.

what everyone would do

Scale the registry itself: hire more staff at the NIC, automate the file edits, distribute HOSTS.TXT faster and more often. Every variant keeps the whole directory in one place and one team's hands, so every added host still adds work to the same centre — growth makes the bottleneck worse exactly as fast as it makes the network bigger.

what they saw

The bottleneck was not the data but its ownership: a directory one institute must edit cannot grow. So editing rights moved to the organisations that already existed, and the centre kept only a map of who to ask.

the move

Mockapetris designed the Domain Name System as a tree of delegated zones: the centre answers only one question — who knows about this part of the tree — and every organisation runs the naming inside its own zone. Names end in a shared top level, authority is cut along administrative boundaries, and answers are cached with time-to-live values, so the system stays fast while no single site holds or updates the whole directory.

why it works

Delegation converts a global consistency problem into many small local ones: each zone's owner updates its own names immediately, without asking anyone, and the rest of the system only needs to know where to ask. Caching with expiries makes repeated lookups nearly free while tolerating the brief inconsistencies that distribution introduces. The centre's workload stops growing with the network at all — it answers 'who knows', a question whose size is fixed, no matter how many hosts exist.

the payoff

DNS replaced HOSTS.TXT, became one of the original internet standards in 1986, and still resolves the entire web today on the 1983 design.

where it breaks

Delegating authority only works when the boundaries you delegate along are real and stable — cut zones where organisations shift and you get ownership fights instead of scale. It also trades away global atomic updates: the system is eventually consistent, and anything that needs the whole world to see one change at one instant has to be built on top of it, not into it. And somebody must still own the root and its trust — the centre shrank, it did not disappear.

what came after

RFC 882/883 is the rare design document whose subject outgrew the network it was written for: the internet scaled by orders of magnitude nobody in 1983 could have planned, and the directory never had to be rebuilt — because nobody at the centre ever had to keep it. Delegation-as-architecture became the template for how the internet grows: solve the coordination problem by handing out authority, not by enlarging the registry.

references

  1. [1]RFC 882 — Domain Names: Concepts and FacilitiesRFC Editor, 1983rfc-editor.org
  2. [2]Internet Hall of Fame — Paul Mockapetris (documents the HOSTS.TXT bottleneck and DNS adoption)Internet Society, 2012internethalloffame.org

keep it

same kind of clever