AI
The Operational Squeeze Driving 5G on Kubernetes
Mobile data traffic is predicted to grow three times by 2031. The technology surface is expanding too: 4G, 5G, 5G standalone, edge, and network slicing all run in parallel. Operators at Swisscom and Telstra face a third constraint on top of that growth: maintain and operate this complex network with the same number of people. That gap is what pushed both operators toward running their 5G core on Kubernetes in the first place.
What a 5G Core Actually Looks Like on Kubernetes
The 5G core is a distributed control plane handling authentication, session management, and user data, plus a User Plane Function (UPF) that acts as a high-speed packet forwarding engine. Every connection between components follows strict 3GPP standards. The system must meet carrier-grade reliability targets. Ericsson’s Jessica Østergaard, quoted in the talk, put the tension plainly: the industry deployed these applications on Kubernetes but has not yet harvested the operational benefits without excessive cost.
How Telco Networking Mindset Bypassed Kubernetes
Telco engineers think topology-first: every node, every interface, every link configured individually and kept alive because services depend on specific instances. Kubernetes thinks service-first: declare intent, let the platform schedule, replace failures automatically. When those two models meet, mismatches appear. At Swisscom, throughput requirements on the UPF led engineers to mount SR-IOV NICs directly into pods. That solved the bottleneck, but bypassed Kubernetes networking and IP address management completely. Then node pinning fixed one UPF instance per node, which eliminated scheduler-based orchestration. Then netconf replaced declarative config. One by one, every Kubernetes feature was removed.
The Lifecycle and Reliability Gap
The config file tells the story. A cloud-native custom resource for a network function runs five lines. The netconf equivalent ships over 2,000 parameters, uses an imperative approach, and gets pushed into Git with the GitOps label attached. That is not declarative. The reliability gap is sharper: Kubernetes detects a failed pod and reroutes traffic in seconds. Today’s 5G deployments page an operator, require 5 minutes to acknowledge, 15 minutes of manual validation, and then recovery time that can stretch to hours or phone calls. Every manual touch point is a scaling limit.
What Swisscom and Telstra Are Doing Now
Both operators started pulling external operations into Kubernetes as the control surface. Swisscom built an automation platform using GitOps and the Kubernetes Resource Model, a cloud-native netconf bridge called STC, a NetBox operator for IPAM integration, and Flux CD as the deployment tool. That reference architecture published on the CNCF site roughly a week before this talk. The goal is to stop configuring boxes through out-of-band tooling and start expressing intent through the Kube API, letting operators handle reconciliation.
What 6G Must Get Right from the Start
5G containerized applications on Kubernetes. That was step one. The cloud-native operating model came later, and the industry has paid for that ordering ever since. Vodafone architect Riccardo Gasparetto said it directly: thousands of manual config parameters, zero abstraction, and upgrades too risky to match the Kubernetes release cadence. The speakers’ prescription for 6G: cloud-native by design, declarative by default, resilient by architecture, not as a retrofit. Drop netconf in favor of Kubernetes operators. Build fully automated testing pipelines. Containerized does not mean cloud-native.
Q&A
How are teams handling the kernel networking bottleneck that forced SR-IOV in the first place? Currently using SR-IOV and DPDK to mount NICs directly into pods, with DRA seen as the path forward. ▶ 28:08
Does the skill gap between telco network teams and platform teams explain why 5G deployments bypassed Kubernetes features? Yes, both sides were learning simultaneously, and even hyperscalers were hiring telco engineers who then needed cloud platform education. ▶ 32:33
Should the industry solve today’s cloud-native gaps before pursuing AI-native 6G networks? The speakers said yes: without stable, properly operating Kubernetes foundations, AI-native networks will not work, and the cleanup must happen first. ▶ 36:37
How does session state survive node reboots or failovers in a multi-region 5G core? Long-term subscriber data sits in a database inside Kubernetes, but session state is the hard part: rebooting a node can kill an active phone call, and multi-region active-active sync remains a challenge. ▶ 40:11
Notable Quotes
we containerized these applications. We put it in a appliance level like a network application into a box, labeled it as Kubernetes. We containerized it, but when we start operating the these applications, we realize that we we don’t use the full leverage of the this Kubernetes like the features, etc. Ashan Senevirathne · ▶ 07:06
it looks like completely broken to a platform engineer looking at from the outside world. Ashan Senevirathne · ▶ 09:55
we also put this uh imperative logic into Git and then called it GitOps. Ashan Senevirathne · ▶ 10:52
for many operators running 5G is painful because it’s often just legacy lift and shift. We deal with thousands of manual config parameters and zero abstraction. Riccardo Gasparetto (Vodafone) · ▶ 18:02
Kubernetes is not failing Telco. Uh it’s important to say that out loud. The platform is capable and maturing. Uh the challenge is how we’ve been using it uh to today. Ashan Senevirathne · ▶ 24:20
Key Takeaways
- 5G UPF deployments at Swisscom removed scheduling, IPAM, and declarative config from Kubernetes one feature at a time.
- Pod failure recovery takes seconds in Kubernetes but 15 or more minutes in today’s 5G ops with manual validation steps.
- 6G must be designed cloud-native from the start, not containerized first with cloud-native added later.
About the Speaker(s)
Ashan Senevirathne is a Product Owner at Telstra and a senior technology leader with more than 10 years in telecommunications, cloud-native transformation, and site reliability engineering. He drives design and delivery of cloud-scale architectures for 5G networks, with a focus on automation and zero-touch operations.
Joel Studler is a DevOps Engineer and System Architect at Swisscom, working on the cloud-native 5G core. He specialises in infrastructure automation, software-defined networking, and highly available databases. He holds multiple CK certifications and has authored several Kubernetes operators and CRDs.