Latency-Based BGP Route Selection – An Intelligent Failover Engine Optimized with Supervised Learning#

Ernesto Sánchez (Universidad Católica de Salta, Argentina), Henri Alves de Godoy (Universidade Estadual de Campinas, Brasil)

• Why: Performance-aware routing is entering BGP via the IETF (a draft adding a NETWORK_LATENCY TLV to AIGP); this is a concrete, honestly-reported prototype of latency-triggered failover plus ML-based policy tuning.
• What: A failover engine that scores peer/DNS/jitter metrics (weights 0.27/0.50/0.23, packet loss as a severity multiplier), switching providers when the max DNS score exceeds 1.10 sustained over consecutive cycles and returning below 0.80 — rules designed to filter noise (trend, coefficient of variation, p95 deviation, z-score acceleration). Supervised learning over ~62 features and 50,000 simulated cycles (~26 days, 98.7% "normal" class): XGBoost ranks feature importance (DNS dominates at ~30% gain), logistic regression yields interpretable weights, both suggesting DNS should carry ~0.76–0.79 vs the current 0.50. Results: recall 0.95–0.97, precision 0.65–0.72, ROC-AUC ≈0.999, stratified 5-fold cross-validation. The authors are explicit that the dataset is synthetic (Ornstein-Uhlenbeck background noise, Poisson episode starts) calibrated to real measurements, and recommend validating against months of real history.
• How: Containerlab dual-provider ISP topology with DNS nodes, TimescaleDB metric store, Python capture/feature pipeline; the ML calibrates the existing deterministic scoring function rather than replacing it — that is the transferable idea.
• Where applicable: Dual-transit ISPs and enterprises; lab-validated only, not production-proven. Recall-heavy/precision-light tuning means false positives are the operational cost to watch.