KubeCon + CloudNativeCon India 2026

As ML workloads move onto Kubernetes, many teams unintentionally turn their clusters into data platforms storing training data, features, and intermediate artifacts alongside compute. While convenient at first, this approach introduces hidden costs that surface over time.

This talk shares real-world lessons from operating ML pipelines on Kubernetes where data management, not models, became the primary source of failures. We’ll explore common anti-patterns involving PVCs, object storage mounts, and ephemeral volumes, and how they led to rising costs, broken reproducibility, and pipelines training on stale or incorrect data. Finally, we’ll discuss practical cloud-native patterns for managing ML data that respect data lifecycles, improve lineage, and keep Kubernetes focused on what it does best.