Databricks ML Professional study plan
Databricks ML Professional · Give each week a clear focus.
8 weeks · About 6 hours per week
Model Development
Map the full Databricks ML workflow surface (workspace, notebooks, clusters, Unity Catalog, MLflow, Feature Store, Model Registry, Jobs, Model Serving) into a one-page diagram you draw yourself from the product documentation
Configure a personal free-tier or trial workspace and run a minimal scikit-learn training job end-to-end on a single-node cluster
Practice reading and writing Delta tables with Spark DataFrames, including schema enforcement and time travel queries
Set up MLflow tracking in a notebook and log params, metrics, and a model artifact for at least two runs
Your checkpoint
A working notebook that trains a scikit-learn model, logs two MLflow runs with artifacts, and registers the better run to the Model Registry, plus your hand-drawn architecture diagram
Model Development
Implement feature engineering with Spark UDFs versus pandas UDFs and benchmark the runtime difference on the same transformation
Build a Feature Store table with a primary key and timestamp key, then look up training features with a point-in-time correct lookup
Train a model with Hyperopt or Optuna on Databricks and log the tuning results to MLflow
Diagnose a skewed join and apply salting or broadcast hints to fix it
Your checkpoint
A Feature Store-backed training notebook that performs point-in-time feature lookup and logs a tuned model, with a short write-up of the UDF benchmark numbers
Model Development
Compare MLflow autologging output against manual logging for the same model and identify what autologging captures by default
Configure a custom MLflow model with a signature and input example, then load it back and score it
Implement a custom pyfunc wrapper that adds preprocessing to a model and verify it loads in a fresh notebook
Register model versions and move one through Staging to Production using registry aliases or stages
Your checkpoint
A registered custom pyfunc model with signature and input example, loadable and scorable from a separate notebook using a registry alias
ML Ops
Design a reproducible training pipeline as a Databricks Job with task dependencies and parameterized widgets
Schedule the job with a cron trigger and inspect run history and failure output
Add retries, timeouts, and notifications to the job definition and force a failure to confirm alerting
Write a training notebook that reads parameters from job base parameters instead of hardcoding paths
Your checkpoint
A scheduled multi-task Databricks Job that trains and registers a model, with a documented failed run and the alert it produced
ML Ops
Implement model validation gates that compare a new model's metrics against the current Production version and block promotion on regression
Automate promotion using the Model Registry API and aliases inside a job task
Set up data quality checks on the training table (nulls, ranges, schema drift) that fail the pipeline
Configure lineage and permissions so a second user can view runs but not promote models
Your checkpoint
An automated promotion pipeline where a deliberately degraded model is rejected by the metric gate and a good model is promoted, with logs proving both paths
ML Ops
Instrument a deployed model with MLflow model serving or inference tables and query the monitoring metrics
Detect training-serving skew by comparing feature distributions between the training table and logged inference data
Build a drift check that compares summary statistics across two time windows and raises when thresholds are crossed
Document a rollback procedure and execute it by moving a registry alias back to a previous version
Your checkpoint
A monitoring notebook plus a written runbook that shows a drift alert firing and a completed rollback to a prior model version
ML Ops & Model Deployment
Deploy a registered model to a Model Serving endpoint and score it over REST with a token
Configure endpoint scale-out and concurrency limits and observe behavior under a small load test
Compare batch scoring via a scheduled job against real-time serving on latency, cost, and freshness, and justify a choice for a stated scenario
Package model code and environment dependencies so the serving endpoint builds without version conflicts
Your checkpoint
A live serving endpoint returning predictions from a REST call, plus a one-page comparison of batch versus real-time serving for two concrete use cases
Model Development & ML Ops & Model Deployment
Complete a timed full-length practice set and tag every miss by domain and sub-topic
Re-study only the weakest three sub-topics identified from the practice results
Rebuild the end-to-end pipeline from scratch without notes: ingest, features, train, tune, register, gate, deploy
Write a one-page cheat sheet of API calls, registry alias commands, and job configuration fields you keep forgetting
Your checkpoint
A scored practice result with a miss-analysis table and a from-scratch rebuilt end-to-end pipeline notebook that runs clean
Milestones
Week 2: You can explain and demonstrate point-in-time feature lookup and why it prevents label leakage, without consulting notes
Week 4: A scheduled job trains and registers a model unattended, and you can show the alert from an intentionally failed run
Week 6: You can trace a model from training run to registered version to deployment, and demonstrate a rollback
Week 8: Practice score is at or above your target passing margin, and the rebuilt end-to-end pipeline runs without errors