Summary

  • At APNIC 16, Masasi Eto argued that an Internet Routing Registry could accept the syntax of two routing policies while missing a contradiction between one AS's export and its peer's import.
  • His presentation reported at least one inconsistency for 55.8% of 11,696 registered ASes examined. APNIC recorded the work as an informational presentation with no further action required.

Imagine a warehouse with two dispatch books. One says a parcel leaves for the next depot; the other says that depot will not receive it. Both books are filled out correctly. A spelling checker is pleased. The parcel has a problem. That was the routing-registry flaw Masasi Eto took to APNIC 16 in August 2003.

An Internet Routing Registry, or IRR, stores statements about how autonomous systems exchange routes. Operators can use those statements to build router filters. Eto's original presentation showed the trap in treating a valid statement as an agreed one. An AS could say it exports routes to a peer while the peer's record omits their import; the reverse mismatch also appears in his examples. The IRR checked the grammar of each policy. It did not ask whether the two policies could both be true.

The deck went beyond an elegant diagram. It described a policy checker comparing peers' statements and an inspection of a unified IRR database. On its result slide, 11,696 ASes are counted as registered and 55.8% are said to have at least one inconsistency. That percentage describes the inspected records in 2003. It is no estimate of current routing failure or lost traffic. Still, it is a sharp answer to anyone who mistakes a clean-looking registry entry for a dependable operating agreement.

The Routing SIG minutes give the proposal its useful resistance. Attendees questioned whether loading even prefix-and-origin pairs onto routers would scale, and whether the approach handled the security of the information it relied on. Another participant described the maintenance trap: networks neglected IRR data because others were not using it. Why spend the morning repairing a book nobody opens? Then why open a book nobody repairs?

Eto discussed a policy checker on JPIRR and notifications to networks with inconsistent entries as future work. APNIC's proposal record says the presentation required no further action; the minutes list no action item. The public record here does not show a completed rollout. Its durable result is the test Eto put on the table: compare what neighbours claim about each other before using their entries to configure a network. A registry can hold both sides of an argument. It has to notice when they disagree.