← Interview Prep

BFD, HA & Fast Convergence

Sub-second convergence: the detect/propagate/compute/FIB budget, BFD (async/echo/micro-BFD), LFA/rLFA/TI-LFA, BGP PIC and next-hop tracking, graceful restart vs NSF/SSO, MPLS FRR, and ISSU.

When a link or node fails, how fast does traffic reroute? "Convergence" is a budget you can decompose — and every fast-convergence feature attacks one term of it. BFD shrinks detection; LFA/PIC/FRR shrink repair; graceful restart removes control-plane restarts from the equation entirely.

The convergence budget

Total = detect + propagate + compute + update FIB. Sub-second convergence means attacking each term: fast detection (BFD), fast flooding and SPF (throttling/iSPF), and — the big one — a precomputed backup so repair is a local FIB switch, not a network-wide recompute.

Default protocol hello/hold timers (OSPF 40s dead, BGP 180s hold) are far too slow. You either tune timers (costly in CPU/stability) or offload detection to BFD.

BFD — Bidirectional Forwarding Detection

A tiny, protocol-independent hello that runs in the forwarding path to detect a path failure in milliseconds, then tells its client protocols (OSPF/IS-IS/BGP/static/PIM) to tear the neighbor down immediately — far faster than their own timers.

• Async mode (default): both ends send control packets every tx interval; a session is down after multiplier missed packets. Detection ≈ rx × multiplier (e.g. 50 ms × 3 = 150 ms).
• Echo mode: send packets the neighbor loops back through its data plane — tests the actual forwarding path and lets you slow the async rate.
• Demand mode: stop periodic packets once up; poll only when needed.
• Single-hop vs multihop (e.g. multihop for iBGP over a loopback); micro-BFD per-member-link for LAG/port-channels.

BFD offloads liveness so you keep the routing protocol's timers relaxed (low CPU) yet still fail over fast. Related: BGP, OSPF.

IGP fast convergence

• Detection: BFD, or carrier-delay/debounce tuning + tuned hellos.
• Reaction: LSA/LSP generation and SPF throttling (exponential backoff: fast on the first event, damped on flaps) and incremental SPF (recompute only what changed).
• Precomputed repair — LFA: Loop-Free Alternate installs a backup next-hop in the FIB before failure. On failure the router swaps to it locally in <50 ms. Remote-LFA (a targeted LDP tunnel) and TI-LFA (Topology-Independent, via Segment Routing) extend coverage to 100% with the post-convergence path. See Segment Routing.

BGP convergence

• Next-hop tracking: react to IGP next-hop changes immediately instead of on the scan timer.
• BGP PIC (Prefix-Independent Convergence): precompute a backup path so failover is one FIB edit regardless of how many prefixes use it — PIC Edge (backup PE) and PIC Core (backup IGP path). Needs a backup path present: add-paths or best-external.
• BFD for BGP: sub-second peer-down detection instead of the 180s hold.
• Graceful Restart / GR helper: a router restarting its BGP control plane keeps forwarding from the last-known FIB while peers hold routes marked Stale and help it relearn — a control-plane restart with no data-plane loss. Pairs with NSF/SSO.
• Graceful shutdown (the well-known community): de-prefer a session before maintenance so traffic drains before you tear it down — make-before-break for planned work.

MPLS fast reroute

MPLS FRR pre-signals a backup LSP so a Point of Local Repair switches in <50 ms: link protection (backup around a link) and node protection (around a whole router), with facility backup sharing one bypass for many LSPs. In SR, TI-LFA gives the same local-repair guarantee without RSVP. See RSVP-TE.

Hardware HA: NSF/SSO, ISSU

NSF vs Graceful Restart: NSF is the local ability to keep forwarding during an RP/control-plane restart; GR is the protocol signalling that asks neighbors to help and hold routes while it happens. They work together.