From Chaos To Control: Building ML Platform
Talkat KubeCon + CloudNativeCon Europe 2025 · on Friday, 2025-04-04· at 13:45· Level 1 | Hall Entrance S10 | Room A·
summary
Abacus is an internal ML platform built on Kubernetes that automates onboarding (creates GitHub repo, Kubernetes namespace, Harbor image registry, LakeFS repo, Azure AD group) and surfaces 30‑day costs in the UI; it runs production workloads for ~3 years with 200 monthly active users and 48 contributors. Workflows split into insights (notebooks, Spark, LakeFS) and ML product (Kubeflow pipelines, Katib, CI that builds/tag/uploads components), and deployments use Argo CD + KServe; system works, but operators would now replace some bespoke services.
AI-written summary · about
chapters (17)
- 0:00ML platform Abacus intro
- 1:05Technology stack & enterprise context
- 3:11Platform scale & community
- 3:33Chaos to control with ML
- 4:37Abacus lifecycle overview
- 7:07Control vs flexibility in design
- 7:28First mile onboarding
- 11:02Onboarding lessons learned
- 13:08Day-to-day operations & tiers
- 17:28Insights tier analytics
- 18:11ML product tier workflows
- 21:44Platform implementation learnings
- 23:08Final mile model deployment
- 26:36Production deployment challenges
- 29:03Integration & maintenance
- 30:08Community & platform as product
- 30:54Q&A
related talks
12mKeynote: Building Enterprise AI/ML Platform on Kubernetes
44mEngaging the Kubeflow Community: Building an Enterprise-Ready AI/ML Platform
28mMachine Learning on Kubernetes at Shell: A Kubeflow Journey
22mBuilding a MLOps Platform with Kubeflow on GKE
08mWelcome + Opening Remarks
46mDeploying scalable Machine Learning models with Kubernetes
37mA Day in the Life of a Data Scientist. Conquer ML Lifecycle on Kubernetes
21mMastering the DevOps Machine Learning Pipeline