How SNCF Cut K8s Waste and Increased Reliability at Scale

SNCF operates France's national rail network and global mobility services (TGV, OUIGO, Eurostar, TER, Transilien, Keolis) with Kubernetes underpinning ticketing, timetables, onboard services, and real-time logistics across hundreds of clusters. Manual rightsizing via Datadog, Prometheus, and FinOps workshops could not scale across 200+ projects and up to 250 clusters, some hosting more than 1,000 workloads. Over-provisioning was widespread because engineers defaulted to safety in production, where optimization carried real risk of disruption. Datadog-based analysis often overestimated usage due to aggregation effects, eroding trust in recommendations. Following the 2024 Rugby World Cup and Olympic Games, leadership mandated a refocus on digital cost efficiency without slowing modernization.

SNCF discovered PerfectScale at KubeCon and adopted it as a production-grade optimization control plane. The key differentiator was risk-aware, in-place rightsizing recommendations that could be safely applied in live production. PerfectScale integrates via Custom Resources and ArgoCD, allowing Autopilot to be activated at the namespace level as a feature flag. SNCF deployed a standard Autopilot configuration across non-production environments automatically, with fine-grained overrides for edge cases. In production, automation was introduced gradually, starting with the cloud-native stack (Datadog, Kyverno, KEDA, AWS Load Balancer Controller, Karpenter), before extending to application namespaces. Optimization became continuous governance rather than a finite project.

SNCF discovered PerfectScale, now PerfectScale by DoiT, at KubeCon. The key differentiator wasn't another dashboard — it was the ability to generate risk-aware, in-place rightsizing recommendations that could be safely applied in live production environments. As Thomas Comtet, Senior Staff Engineer at SNCF, put it: "What convinced us was that PerfectScale did not ask us to trust theory. It showed us exactly what could change without hurting stability."

PerfectScale by DoiT integrates via Custom Resources and ArgoCD so that Autopilot can be activated at the namespace level as a feature flag. SNCF established a standard Autopilot configuration and deployed it across non-production environments automatically, while allowing fine-grained overrides in edge cases. For production, SNCF introduced automation gradually, starting with the entire cloud-native stack (Datadog, Kyverno, KEDA, AWS Load Balancer Controller, Karpenter), and validated reliability before extending to application namespaces.

Rather than running a finite rightsizing initiative, SNCF turned optimization into a continuous operating behavior, enforced automatically through governance with PerfectScale by DoiT. Because recommendations are grounded in observed workload behavior and enforced through automation, they are no longer debated but executed as policy. The PerfectScale team also anticipated that in-place resizing would remove the hidden cost of restart-based optimizations — work that would otherwise have required SNCF to custom-engineer the capability or train dozens of teams to apply it safely.

SNCF plans to continue expanding automation across additional non-production namespaces and gradually into selected production environments for early adopters. The team is also evaluating in-place pod rightsizing to further minimize restart-based disruptions and improve workload stability. Beyond its own adoption roadmap, SNCF has become an active contributor to PerfectScale's product evolution — recent enhancements such as Java workload support and in-place pod rightsizing were directly influenced by SNCF's feedback. "We proved it in production," said Thomas Comtet. "Now we're scaling what works and helping make it even better."