America First Credit Union optimizes their MLOps stack with Hopsworks
summary
America First converts Oracle SQL centric pipelines to PySpark and replaces Oracle RDBMS with Hopsworks and HopsFS, integrating Redgate, Airflow, Great Expectations, PyTest, MLflow and Label Studio. Training-set feature groups move onto the Hopsworks feature store, enabling fast Spark transforms and versioned training sets. Measured gains include backfills shrink from weeks to ~1 day, job runtimes cut to 30–50% of prior times, and onboarding speeds increase 3–4x, plus improved compliance, reproducibility and online feature serving readiness.
AI-written summary · about
chapters (14)
- 0:00Introduction and background
- 1:03Prehistory and early challenges
- 3:30Transition to CI/CD and SQL tools
- 6:39Feature store journey begins
- 7:41Choosing Hopsworks for on-prem deployment
- 8:03Hopsworks deployment and early ROI
- 9:28Performance gains and backfill efficiency
- 10:32Data scientist onboarding improvements
- 11:59Compliance and model reproducibility
- 13:01Moving to online serving
- 14:04Automating model validation
- 15:31Enhancing data lineage and documentation
- 16:13Feedback loop with Label Studio
- 16:56Future roadmap and conclusion
related talks
14mMLOps at Scale: AI Systems for HR Transformation
32mA feature Store for Machine Learning
53mExperience with an end-to-end MLOps pipeline: Practical insights and 3D architectural perspectives
36mMLOps from the edge to on demand visualization
32mMaking Models Work: How an optimized ML lifecycle can drive better ROI
13mMLOps at Scale
42mApache Hop
49mBuilding a Data-Driven Bank