Part 5 · Tools of the trade · 6 of 6
Data Science and ML Platforms: MLflow, SageMaker, Vertex AI, Dataiku, Notebooks
Where models are built, tracked, and served, and what a Data PM needs from them: reproducibility, a registry, and evaluation before anything reaches a decision.
2 min read
A Data PM rarely trains models, but often owns the data products they consume and the decisions they feed. Know the parts.
| Tool | What it does | What a DPM asks of it |
|---|---|---|
| Jupyter / Databricks notebooks / Hex / Deepnote | Interactive Python (the pandas you practised) with charts and prose; where analysis and prototyping happen. | Is the notebook reading a governed table or a CSV someone exported in March? |
| MLflow (open source, built into Databricks) | Experiment tracking (parameters, metrics, artifacts per run), a model registry with stages (staging, production), and packaging for deployment. | Which model version is live, what data was it trained on, and what were its metrics? |
| Amazon SageMaker | AWS's end-to-end ML platform: notebooks, training jobs, feature store, model hosting, pipelines, monitoring. | Who owns the feature definitions, and do they match the warehouse's? |
| Google Vertex AI | GCP's equivalent: AutoML and custom training, feature store, pipelines, model monitoring, plus the Gemini model APIs. | Same, plus cost per prediction. |
| Dataiku | Visual data science platform for mixed teams: drag-and-drop recipes and code side by side, deployment and governance included. | Good for analyst-heavy teams; check that "recipes" don't become an ungoverned second transformation layer. |
| Feature stores (Feast, Tecton, platform-native) | Define a feature (e.g. days_in_negotiation) once and serve the same value to training and to production. | The ML version of the semantic layer: one definition, two consumers. |
The Data PM's checklist for anything model-shaped
- Lineage: which data product does it read? Is that product's SLO good enough for this decision?
- Reproducibility: can we retrain the live version from tracked code, data, and parameters?
- Evaluation: a held-out set with agreed answers, run before every change (week 6's agent lesson applies to classical models too).
- Monitoring: drift in inputs and outputs, not just uptime.
- Cost per correct prediction, the Lean AI metric.
Source: Synthesized for DPM Lab from vendor documentation and public positioning (as of 2026).
Where this shows up