Articles

Resilience and convergence: designing for what breaks, not for what works

Convergence has phases, and the network design decides which one dominates.

Reading: 5 minNetworking

Article cover: Resilience and convergence: designing for what breaks, not for what works

A network is designed against a failure model, not against failure in general. Convergence — the work that returns forwarding to a correct state once something breaks — is a sequence, not an event: detect, propagate, recompute, install. Each mechanism here shortens one step of it. BFD (Bidirectional Forwarding Detection) shortens detection; fast reroute pre-computes a repair a local router switches to without waiting for the network; graceful restart keeps a router forwarding while only its control plane restarts. None is resilience on its own: redundancy tolerates a failure only when the failure modes are independent, and failover is proved by test, never by configuration.

Convergence is four problems in a row

When a link or node fails, “there is inevitably a period of disruption to the delivery of traffic until the network re-converges on the new topology” (RFC 5714, lines 86–88). The disruption has parts (lines 268–304): detect the failure; react locally; propagate the news; recompute the tables; install them in the forwarding hardware. Detection is usually the largest term, and the one a design changes most cheaply.

failure
   │
   ▼
[ DETECT ]  default: the routing protocol's dead timer (seconds)
   │        with BFD or link signalling (milliseconds)
   ▼
[ REACT ]   the adjacent router stops using the failed path
   │
   ▼
[ PROPAGATE → RECOMPUTE → INSTALL ]   the network re-converges
   │
   ▼
traffic back on the optimal path

Detection: the timer you cannot see

A routing protocol already detects failure on the dead interval it agreed with its neighbours — the L3 sheet’s timer. BFD exists because that timer is too slow: in existing protocols the detection times “are no better than a second” (RFC 5880, lines 143–146). BFD is “a simple Hello protocol” (lines 239–241): two systems exchange packets, and if enough are missed the session goes down (lines 287–290). The Detection Time “is not carried explicitly in the protocol” (lines 1770–1775); each side computes it from the agreed interval and the Detect Mult it received (lines 486–490). It runs single-hop on UDP port 3784 and multihop on UDP port 4784 (RFC 5881, lines 151–153; RFC 5883, lines 214–218). Its limits: it is two-sided — “both sides of a session MUST take the ‘Active’ role” (RFC 5881, lines 143–145), so a peer that does not run BFD is not protected; it proves only the path it watches; and haste produces false positives, so the rate must be provisioned “to avoid congestion … and false failure detection” (RFC 5881, lines 127–130).

Repair before the network knows

Shortening detection still leaves the sequence to run. Fast reroute takes it off the critical path: a router pre-computes a path that works before the news arrives — “compute backup routes that allow the failure to be repaired locally by the router(s) detecting the failure without the immediate need to inform other routers” (RFC 5714, lines 105–110). The simplest repairs are an equal-cost path that avoids the failure (lines 408–410) or a loop-free alternate (LFA), “a pre-computed backup IGP next hop for that destination” (RFC 6571, lines 182–185). LFAs need no protocol change and are deployed “router by router” (lines 1499–1501); the backup is installed “in advance of the failure”, which makes a sub-50-millisecond repair possible (lines 1503–1506). The catch is topological: “LFA FRR’s applicability depends on topology” (lines 1689–1690), and its coverage figures are “for illustrative purposes only” (RFC 5714, lines 496–498).

Restarting without dropping the traffic

A control plane can fail while the data plane still forwards. Graceful restart, or non-stop forwarding, exploits that separation so “an OSPF router can stay on the forwarding path even as its OSPF software is restarted” (RFC 3623, lines 32–35, 90–96). Its neighbours keep it advertised as fully adjacent, “but only if the network topology remains static” (lines 130–135); if the topology changes, or a neighbour does not support the feature, the router “automatically reverts to a normal OSPF restart” (lines 36–41) — the safety property: a stale table is worse than a brief restart. BGP works the same way: a session reset causes “transient forwarding blackholes and/or transient forwarding loops” (RFC 4724, lines 83–91), which Graceful Restart mitigates by letting a speaker “preserve forwarding state during BGP restart” (lines 96–98). But it is a trade: attempting it after an unplanned outage “may not be a good idea” (RFC 3623, lines 523–527).

Failure domains, and what redundancy does not buy

Availability is not new — “high reliability” was already an application requirement (NIST SP 800-215, lines 1308–1310). One failure is rarely one failure. A node failure “causes the simultaneous failure of multiple links” (RFC 5714, lines 608–611), and links sharing a route, a conduit or a power feed form a Shared Risk Link Group, “an example of multiple related failures” (lines 604–606). Ask of every redundancy what primary and backup have in common — device, rack, power, physical route, the configuration pushed by the same automation. If they share it, the redundancy is decoration. An area boundary limits how far a change floods: containment, too, is bought by design. Redundancy supplies a second path, not a second failure: a configuration error or shared dependency is replicated to both members. Failover is verified by test; a design that has never failed over has an unproven assumption.

Level and prerequisites

L4 — design-level. It assumes the L3 sheet’s convergence model and the L1–L2 forwarding model (next hop, the table); neither is re-explained. It introduces no configuration and no command: platform syntax is implementation work.

Where to go next

References

  • RFC 5880 — Bidirectional Forwarding Detection (BFD) (Katz, Ward, June 2010, Standards Track): the goals of the protocol, the Asynchronous mode, the Detect Mult and the interval fields, and the calculation of the Detection Time.
  • RFC 5881 — Bidirectional Forwarding Detection (BFD) for IPv4 and IPv6 (Single Hop) (Katz, Ward, June 2010, Standards Track): the single-hop application, the control packets on UDP port 3784 and the TTL 255 rule, and the limits of the mechanism.
  • RFC 5883 — Bidirectional Forwarding Detection (BFD) for Multihop Paths (Katz, Ward, June 2010, Standards Track): the multihop control packets on UDP port 4784 and the authentication recommendation over multiple hops.
  • RFC 5714 — IP Fast Reroute Framework (Shand, Bryant, January 2010, Informational): the phases of a routing transition, the repair paths and the warning that the coverage figures are illustrative.
  • RFC 6571 — Loop-Free Alternate (LFA) Applicability in Service Provider (SP) Networks (Filsfils, Francois, Shand, Decraene, Uttaro, Leymann, Horneffer, June 2012, Informational): the definition of a loop-free alternate and the topological dependence of its applicability.
  • RFC 3623 — Graceful OSPF Restart (Moy, Pillay-Esnault, Lindem, November 2003, Standards Track): non-stop forwarding, the static-topology precondition and the reversion to a normal restart.
  • RFC 4724 — Graceful Restart Mechanism for BGP (Sangli, Chen, Fernando, Scudder, Rekhter, January 2007, Standards Track): the routing flap a restart causes, the Graceful Restart capability and the stale-route timers.
  • NIST SP 800-215 — Guide to a Secure Enterprise Network Landscape (Chandramouli, November 2022): reliability and disaster recovery as drivers of distribution, and the principles that survive each network landscape.
Nodrius Field Kit — cover

Get the Nodrius Field Kit

Enter your email and we'll send you the complete Field Kit: eight technical resources in one download.

Your email address is sent to Nodrius and used to deliver the Field Kit to you by email. Marketing messages are separate and optional: the checkbox does not affect your download, and you are only added to our updates list if you tick it. Privacy Policy.

Privacy & cookies

This site does not use cookies, analytics or tracking, and it does not profile you. There is nothing technical to switch off, so Accept and Reject change nothing about how the site works: either choice lets you browse everything normally.

The only difference is what your browser remembers: two entries in local storage — your choice here, so this notice is not shown again, and a marker that the Field Kit offer has already been made — so the offer is presented once and not on every page. They contain no identifier, they are not shared with anyone, and the site sets no cookie for either.

What we do with your data
How your email address is used when you ask for a Resource, how long it is kept and which rights you have — in the Privacy Policy.
What the site stores in your browser
Two entries in local storage — your choice on this notice and whether the Field Kit offer has been shown — and no cookies, no session storage, no third-party content: in the Cookie Policy.