Summary

  • RFC 5004 retains an existing external best path when it and a challenger are still tied immediately before the BGP-Identifier comparison. The incumbent is not declared faster or safer; the router has merely run out of operationally meaningful reasons to replace it.
  • Enke Chen and Srihari Sangli converted an observed oscillation into a deliberately small extension. Its value depends on evidence that the equality predicate actually held, the forwarding table did not move, and later policy or reachability changes can still end the retention.

A forwarding change with no new reason

Imagine a border router with external path A already installed. Path B arrives. The normal decision process compares the routes through locally assigned preference, AS path, origin, the MED rules that apply, external versus internal learning, and the interior cost to the next hop. Every comparison remains tied. Only then does B win because its advertising speaker has a numerically lower BGP Identifier.

Nothing in that last number says that B carries more capacity, has a better commercial settlement, crosses fewer fragile facilities or currently delivers packets. It is a four-octet identifier used to make the procedure terminate consistently. RFC 6286 even removed the need for it to be an address assigned to the speaker: it is a non-zero unsigned value that should be unique within the AS. Its dotted-quad appearance is not a measurement of a path.

Yet under the base RFC 4271 procedure, that weak distinction can replace the selected route. Control-plane state changes, the Loc-RIB may change, the FIB may be rewritten and downstream speakers may receive new announcements. The traffic could continue without loss; it could also meet transient reordering, convergence or a different failure domain. The decision algorithm alone cannot prove the packet outcome.

RFC 5004 inserts restraint at precisely this boundary. Let A be the existing external best path and B the external challenger that the ordinary procedure would select. If neither route has been eliminated before the BGP-Identifier step, retain A. The extension does not make A newly superior. It says that B has supplied no sufficiently meaningful reason to disturb an already running choice.

The equality predicate is the real feature

The useful word in the mechanism is not “existing”; it is “tied.” A route with higher LOCAL_PREF still wins. So can a route that wins through AS_PATH, ORIGIN, an applicable MED comparison, EBGP over IBGP, or lower IGP cost to NEXT_HOP. A withdrawal or loss of next-hop reachability also ends the question. RFC 5004 is not permission to keep a stale route against contrary evidence.

Its exclusions are equally important. The rule does not apply when either candidate comes from a BGP Confederation peer. It also does not apply when both peers use the same BGP Identifier, a situation in which the peer-address comparison across parallel sessions may encode a deliberate operational choice. These are not implementation footnotes. They define where “keep the incumbent” would erase information that an operator may have intended to matter.

This makes an enablement checkbox poor proof of behavior. To say the extension retained a path, an operator needs a decision trace containing both candidates, the values of every earlier comparison, the exact comparison at which the ordinary algorithm would have switched, and the implementation's recorded selection reason. The resulting FIB entry and packet observations are separate records. A route table showing the same next hop before and after the event does not, by itself, show whether RFC 5004 ran, no challenger arrived or policy rejected the challenger earlier.

What one 1998 incident can and cannot establish

Chen and Sangli wrote that the idea was inspired by a route-oscillation case observed in the BBN/Genuity network in 1998, and that the algorithm was implemented and deployed there. That is valuable provenance: the mechanism began with an operational failure rather than an abstract wish for calm. It is not a census of current deployment, a universal efficacy claim or proof that today's route-reflector topology reproduces the same condition.

The worked example matters because route reflection changes who can see which alternatives. RFC 4456 removes the need for every IBGP speaker to maintain a full mesh by allowing a route reflector to readvertise selected routes. The scaling gain comes with partial visibility. A client generally receives the reflector's chosen route, not a complete account of every external candidate available elsewhere.

In the RFC 5004 example, MED and this incomplete view allow a sequence in which a newly selected external route changes what is advertised, which changes a neighbor's selection, which removes the condition that supported the first change. Retaining the existing external route at the late tie can break that cycle. But the document expressly does not claim to solve every oscillation catalogued in RFC 3345. Some topologies and MED comparison patterns require design or policy changes beyond local incumbent retention.

The distinction protects the article from a familiar retrospective error. An RFC that treats one observed failure mode is not the owner of all failures that share a symptom. Repeated UPDATE and withdrawal activity can come from physical instability, policy churn, malformed state, route-reflector visibility, MED ordering or other causes. The diagnosis must reconstruct the actual candidate sets and decisions.

Stability purchased with history

The extension reduces needless changes, but it does not create network-wide determinism. Two equally configured routers can retain different paths if they encountered those paths in a different order. RFC 5004 acknowledges this history dependence, sometimes described as non-determinism. The term should not be used as an accusation. It is an explicit allocation of state: when all stronger comparisons tie, the local past is allowed to decide the present.

That bargain is often sensible at an exchange point with multiple external peers, where a meaningless last tiebreak could otherwise cause avoidable forwarding changes. It can also surprise an operator comparing two devices and expecting identical route selection from identical current inputs. The missing input is the incumbent itself.

If a network needs one route to win everywhere, it should author that preference in the authority it controls—often LOCAL_PREF, or MED where its comparison semantics fit—rather than hoping a remote BGP Identifier happens to order the paths as desired. Explicit policy is portable evidence of intent. Incumbency is evidence only of sequence.

The later answer: show more paths

ADD-PATH, specified in RFC 7911, attacks a different part of the information problem. Negotiating the capability lets a speaker advertise multiple paths for the same prefix by attaching locally meaningful Path Identifiers. This can expose alternatives that ordinary BGP would suppress. It does not say which paths must be sent, guarantee a complete set, select the best route, create ECMP or make a Path Identifier durable across a restart.

RFC 7964 later described using ADD-PATH to advertise all paths or a Group Best set separated by neighboring AS, addressing persistent oscillations caused by incomplete MED comparisons. More visibility costs memory, update volume and convergence work; the RFC advises limiting the technique to affected prefixes when possible. RFC 5004 spends almost the opposite resource. It keeps state small and preserves one existing choice inside a narrow equality class.

Neither approach is a slogan for “stability.” One reduces transitions by giving the past limited authority. The other distributes more candidate state so remote decisions can be better informed. A deployment can need one, the other, both or neither. The evidence is the topology and failure mechanism, not the feature name.

Enke Chen's bounded place in the record

Enke Chen co-authored RFC 5004 with Srihari Sangli. His wider IETF record includes route reflection, the revised BGP Identifier, ADD-PATH and work on persistent oscillation. That continuity makes him a useful person through whom to read how BGP designers moved among three scarce resources: stable forwarding, complete information and bounded control-plane state.

It does not make Chen the sole inventor of BGP, the owner of an operator's policy or the person responsible for any present implementation. The RFC authors describe a mechanism. Vendors decide how to implement and expose it. Network teams decide whether to enable it and what explicit preferences surround it. Route reflectors determine which alternatives are propagated. The FIB installs a result, and packets reveal whether delivery followed.

A University of Michigan Engineering profile provides a dated institutional account of Chen's education and career in Internet technologies. It is appropriate evidence for identity and biography. It is not evidence that a particular network currently uses RFC 5004 or that Chen controls that network's routing choices.

The proof chain for a path that did not move

The hardest event to audit is often an event that did not happen. A useful record would connect six layers:

  1. Adj-RIB-In evidence shows both external candidates and their received attributes at the relevant time.
  2. Import-policy output shows the local attributes after policy, especially LOCAL_PREF and any modified MED.
  3. The decision trace proves that the routes survived through IGP-cost comparison and identifies the BGP-Identifier result that would otherwise replace A.
  4. The implementation records incumbent retention as the actual selection reason, including exclusions that were checked.
  5. Loc-RIB and FIB evidence shows whether control and forwarding state remained unchanged.
  6. Packet and service telemetry shows what traffic did; it must not be inferred from the route-selection log.

The same record needs an exit condition. A policy change, candidate withdrawal, next-hop failure, changed applicable MED, lower IGP cost or loss of the incumbent must be able to produce a new decision. Continuity without a tested escape becomes hidden lock-in.

RFC 5004 therefore offers a modest but durable lesson. A protocol need not let the weakest available distinction exercise the strongest operational consequence. It may preserve running state when the challenger has not earned displacement. But the retained path occupies its seat on a narrow procedural ground, not on a certificate of quality. Good operations preserve that distinction in telemetry, review and language.

Sources