Kubeflow in Practice: ML Pipelines on Kubernetes

Your platform remembers runs. Every question you actually get asked — by an auditor, by an incident reporter, or by your future self — is about a product. A live build of ML pipelines on Kubernetes using Kubeflow and Mnemosyne (the NASA JPL-born OODT, modernised), treating provenance as a first-class output. Two questions drive the whole thing: can you prove your best model never saw its test set, and can you answer whether a batch of source data was faulty? Chapters 0:00 Two questions about your best model 1:04 2,212 papers. 62 survived. Zero were usable 2:37 The EU AI Act is coming 5:06 Brain drain and shared knowledge 5:43 A run is a verb. A product is a noun 6:51 Training is the easy part 8:18 Blast radius: who consumed the output 9:57 Lifetime: pipelines that outlive the cluster 11:45 What Kubeflow actually is 17:10 Kubeflow graduates from the CNCF 20:51 The record layer, and what nobody does 22:40 OODT: built at NASA JPL 26:25 Mnemosyne: file manager, catalog, workflow 30:28 Lineage: naming your parents 34:25 Demo: the OPSUI walkthrough 37:20 The pipeline, end to end 40:03 The provenance gate in action 41:35 Inside the Kubeflow run 45:20 Serving the model 49:49 Provenance is the product The demo, end to end: https://github.com/spiculedata/kubeflowdemo Mnemosyne by Chris Mattmann: https://github.com/chrismattmann/mnemosyne Kubeflow: https://kubeflow.org — CNCF graduated, August 2026 Tom Barber: https://linkedin.com/in/tombarber Concept To Cloud builds production data and ML platforms for regulated, data-heavy organisations. https://concepttocloud.com #kubeflow #kubernetes #mlops #dataengineering #machinelearning