History of the Internet

TLS Session Ended at the Server, Not at the CGI Script

The TLS Session Ended at the Server, Not at the CGI Script

CGI made the web server’s handoff to an application program legible across implementations. Its 2004 specification also drew a boundary that is easy to miss: the server held the authenticated network session; the script did not inherit it.

IAB Asked for Research Funding, Not a Budgeting Role

The IAB Asked for Research Funding, Not a Budgeting Role

In 2004, the Internet Architecture Board argued that shared Internet foundations needed sustained research money. Its final clarification was just as consequential: the IAB, IETF and IRTF were not asking to handle that money themselves.

Deadline Expired. The SCTP Stack Could Check Later.

The Deadline Expired. The SCTP Stack Could Check Later.

RFC 3758 made partial reliability a policy boundary: an application could ask SCTP to stop trying, but the protocol did not promise a per-message timer firing at the exact deadline. Its generic wire signal could move a receiver past abandoned sequence numbers; when that signal was generated remained a sender-side service decision.

Fast, Smooth, Seamless: RFC 3753 Kept Three Handover Labels Apart

Before engineers could compare a mobile handover, they needed to say what kind of handover they meant. In June 2004, an IETF glossary separated five mostly independent control questions and distinguished latency, packet loss and application-visible continuity—without specifying a mobility protocol or claiming that its vocabulary was universal.

A Search Result Was Not an Identity: RFC 3744’s Canonical Principal

WebDAV’s access-control design had to solve a problem that a display name could hide: one person or software agent might be reachable through several URLs, while an access-control entry needed an unambiguous principal reference. RFC 3744 paired limited human-oriented discovery with a canonical principal URL—and kept both separate from authentication and permission.

One IDN Registration Reserved a Package, Not Just a Label: RFC 3743

When a registry accepts a CJK domain label, the visible string may be only one member of a larger administrative family. RFC 3743’s Joint Engineering Team proposed treating selected character variants as one reservation package, while leaving each zone to decide which forms should actually enter DNS.

One Datagram Changed Its Route. The Other Socket Defaults Should Have Survived: RFC 3542

The most revealing API changes are sometimes changes in the blast radius of an exception. RFC 3542 let one IPv6 datagram depart from one socket default without making that packet an excuse to erase every other choice the application had already made.

Best Route Was Not a Stable Answer: RFC 3345

The Best Route Was Not a Stable Answer: RFC 3345

A router can choose the best path in its own table, advertise it, and still help create a loop that makes the next “best” choice different. RFC 3345 showed how that can happen when BGP’s scaling shortcuts reduce which exit paths each router sees.

AD Bit Reported Validation Policy, Not a Signed Answer: RFC 3655

The AD Bit Reported Validation Policy, Not a Signed Answer: RFC 3655

A DNS reply can contain a one-bit assurance from the resolver that checked it. In 2003, RFC 3655 narrowed what that bit meant—and made clear that trusting the bit depended on trusting the resolver and the path to it.

Route Stopped Flapping. Its Penalty Had Not: RFC 2439

The Route Stopped Flapping. Its Penalty Had Not: RFC 2439

A BGP route can become reachable again before a router has forgotten its instability. RFC 2439 turned recent route history into a temporary penalty: decay could restore a suppressed path, but only after its score crossed a separate reuse threshold.

Control Lost Its Association; Forwarding Still Needed a Rule: RFC 3654

A router can lose contact with the process that programs it without losing every ability to move packets. RFC 3654 treated that interval as an architectural choice: detect the break, decide what the forwarding element should do, and plan how control and state return.

Handle Envelope Sat Outside Its Message Credential: RFC 3652

The Handle Envelope Sat Outside Its Message Credential: RFC 3652

A Handle message had two kinds of work to do: carry an operation that could be authenticated, and arrive in pieces that a client could put back together. RFC 3652 gave those jobs different boundaries. The distinction is easy to miss because both lived inside one protocol message.

Global Registry Held the Service Map: RFC 3650

The Global Registry Held the Service Map: RFC 3650

The Handle System’s “global” namespace did not require every resource record to live in one central store. RFC 3650 placed a registry at the root of a service hierarchy: clients could ask it where a naming authority’s handles were served, then go to that service for resolution.

Query Crossed IPv4. Its Answer Could Still Name IPv6: RFC 3596

The Query Crossed IPv4. Its Answer Could Still Name IPv6: RFC 3596

DNS did not need an IPv6 road in order to talk about an IPv6 address. RFC 3596 kept the packet carrying a question separate from the record the question asked for—a small design boundary that let one naming system span a mixed-version Internet.

TCP Delivered the Trap. It Still Did Not Confirm the Operation: RFC 3430

TCP can deliver every byte of an SNMP message in order and still leave the sender without an answer to the question it actually cares about: did the management application receive, process or queue this operation? RFC 3430 made that boundary explicit. It offered a reliable stream for large management exchanges, but kept transport delivery distinct from an SNMP response that confirms an operation.

OAM Label Survived PHP. The Endpoint Still Had to Understand It: RFC 3429

The OAM Label Survived PHP. The Endpoint Still Had to Understand It: RFC 3429

In 2002, the Internet’s MPLS architecture had an awkward edge case: a network could remove the last forwarding label to save work at the destination hop, yet still need a special packet to reach that destination as OAM rather than ordinary user traffic. RFC 3429 assigned one number for that distinction—and drew a line around which endpoints could actually use it.

First Common AAL2 Profile Still Needed a Service Map: RFC 3441

The First Common AAL2 Profile Still Needed a Service Map: RFC 3441

Two gateways could agree on an AAL2 profile list and still not agree which rows inside the selected profile should carry voice, voiceband data or fax. RFC 3441 kept profile intersection, service preference, ATM bearer setup and a completed call as separate decisions.

A Metric Had a Name. The Measurement Still Varied: RFC 3393

RFC 3393 gave packet delay variation a formal name and a flexible mathematical definition. A decade later, the IETF said its metrics registry still could not identify some measurements precisely enough. The story is not that the metric was wrong; it is how a useful research vocabulary proved harder to turn into an interoperable menu.

836-RTT Recovery Was a Lower Bound: RFC 3742’s Slow-Start Erratum

The 836-RTT Recovery Was a Lower Bound: RFC 3742’s Slow-Start Erratum

In 2004, TCP’s slow-start problem was not simply that it grew quickly. On a connection with a very large congestion window, one round trip could add thousands of segments, and one overshoot could leave both the sender and the traffic sharing its bottleneck paying for the mistake. RFC 3742 proposed a ceiling on that growth. Its later verified erratum changed how high—and how long—the proposal said that ceiling could be.

Signature Survived the Move. The Meaning Might Not: RFC 3741

The Signature Survived the Move. The Meaning Might Not: RFC 3741

Moving signed XML between message envelopes creates an awkward choice: carry every inherited namespace and attribute into the signed bytes, or keep the signature stable and make the application responsible for the context it leaves behind. RFC 3741 chose selective inheritance. Its economy is useful, but it does not make a fragment semantically self-contained.