Nokia Architecture Drivers and Implementation

Ahmed Abutaleb, data center architect at Nokia, explains how the problems of an inherited multi-vendor environment shaped the architecture of Nokia IT’s new data centers. He walks through templated building blocks, intent expressed as code in Git, a Kubernetes-native controller, an on-demand digital twin, and continuous drift detection, along with brownfield migration lessons and the agent-driven operating model the team is targeting. The session was part of Nokia’s appearance at Networking Field Day 41.

Abutaleb described a starting point of Cisco, HP, and OmniSwitch fabrics, some past end of service since 2010 and running so hot that engineers were afraid to touch them, alongside a mix of layer 2, layer 3, and EVPN designs and rigid tools. An early audit found a network that was working but wrong, with design documents that no longer matched reality. The new design divides each data center into tested templates for fabrics, server connectivity, firewalls and load balancers, and subnets, with applications choosing from three connectivity models. Engineers express intent in YAML, often through Cursor in natural language, while details such as BGP, EVPN, and BFD stay hidden. Nokia’s internal GitLab serves as the source of truth, and a Kubernetes-native controller treats network objects the same way it treats software services.

Changes go first to a digital twin in GCP that wakes up, mirrors a production fabric, runs automated traffic tests, and generates a method of procedure with test evidence before going back to sleep, though Abutaleb acknowledged the twin covers only the data center, not the WAN, firewalls, or load balancers. The team dropped Argo CD after GCP throttling stretched syncs to 30 or 40 minutes, and built continuous drift detection that flags any unreconciled change, a design he said the product team later built into EDA. Greenfield fabrics come up through zero-touch provisioning in minutes, but brownfield migrations required learning application traffic flows, storage access, and link ordering, since factories tolerate only a few seconds of disruption. Early migrations caused outages, but once application, infrastructure, and network teams cooperated, the final migration went cleanly, setting up the team’s target of a single agent window spanning ServiceNow tickets, fabrics, and alarms.