Uploaded June 2026 | Updated September 2026, 1 week ago
Abstract
This presentation outlines the technical journey and operational best practices for migrating from a high-throughput, bespoke stateless HW load balancer to an open, standardized solution using MetalLB in BGP mode for Kubernetes LoadBalancer services. We detail the architecture of using BGP to advertise service IPs and leverage Equal-Cost Multi-Path (ECMP) routing across the network fabric for per-connection load balancing. The core focus is on solving the critical networking challenge inherent to stateless ECMP: mitigating the widespread connection resets that occur when a backend node is removed. We present the network-side solution, including the implementation of resilient ECMP configuration on ToR switches to stabilize the hashing algorithm, alongside essential operational considerations for ingress node physical distribution and the resultant shift to a stricter port exposure security model. This talk provides actionable insights for network engineers and Kubernetes platform teams seeking to build scalable, production-ready, BGP-native load balancing infrastructure.
Sergey Kolobov:
Sergey is a Network Engineering professional supporting hyperscale infrastructures for AI, DC, DCI and Enterprise customers as Tech Lead Systems Engineer at Arista Networks. He appreciates good documentation and intentional clear communication. He plays ice hockey, appreciates prime backcountry skiing and splendid mountain biking.
Marcin Wojciechowski:
Marcin is a Platform and SRE engineering professional specializing in the management of large-scale Kubernetes clusters at Arista Networks. His infrastructure career spans over two decades, evolving from configuring early Linux-based Frame-Relay gateways in Poland to modern, cutting-edge DevOps practices. Off the clock, he recharges by cycling, walking forest and beach trails, and listening to a great audiobook.
nanog.org/events/nanog-97/content/5740
Abstract
This presentation outlines the technical journey and operational best practices for migrating from a high-throughput, bespoke stateless HW load balancer to an open, standardized solution using MetalLB in BGP mode for Kubernetes LoadBalancer services. We detail the architecture of using BGP to advertise service IPs and leverage Equal-Cost Multi-Path (ECMP) routing across the network fabric for per-connection load balancing. The core focus is on solving the critical networking challenge inherent to stateless ECMP: mitigating the widespread connection resets that occur when a backend node is removed. We present the network-side solution, including the implementation of resilient ECMP configuration on ToR switches to stabilize the hashing algorithm, alongside essential operational considerations for ingress node physical distribution and the resultant shift to a stricter port exposure security model. This talk provides actionable insights for network engineers and Kubernetes platform teams seeking to build scalable, production-ready, BGP-native load balancing infrastructure.
Sergey Kolobov:
Sergey is a Network Engineering professional supporting hyperscale infrastructures for AI, DC, DCI and Enterprise customers as Tech Lead Systems Engineer at Arista Networks. He appreciates good documentation and intentional clear communication. He plays ice hockey, appreciates prime backcountry skiing and splendid mountain biking.
Marcin Wojciechowski:
Marcin is a Platform and SRE engineering professional specializing in the management of large-scale Kubernetes clusters at Arista Networks. His infrastructure career spans over two decades, evolving from configuring early Linux-based Frame-Relay gateways in Poland to modern, cutting-edge DevOps practices. Off the clock, he recharges by cycling, walking forest and beach trails, and listening to a great audiobook.
nanog.org/events/nanog-97/content/5740
