Databricks Certified Machine Learning Professional Roadmap

Build a rigorous model of Spark ML stages, tuning, evaluation, and batch, streaming, or online inference.

• Choose Spark ML when data, preprocessing, and supported algorithms benefit from distributed execution.
• Construct Pipelines with ordered Estimators and Transformers such as StringIndexer, OneHotEncoder, VectorAssembler, and a classifier or regressor.
• Use evaluators and CrossValidator without consuming the final test set during tuning.
• Evaluate imbalance, threshold behavior, slices, and business costs—not one aggregate metric.
• Use PipelineModel transformation for distributed batch scoring and compatible Structured Streaming paths for streaming inference.
• Select Model Serving for low-latency request-response use cases instead of forcing Spark into one-row calls.

Match parallelism to data, model, framework, and failure behavior while preserving experiment and feature lineage.

• Compare vertical and horizontal scaling, data and model parallelism, Spark and Ray, and grouped pandas function APIs.
• Run bounded distributed hyperparameter tuning with Optuna or Ray and log trials as nested MLflow runs.
• Record custom parameters, metrics, artifacts, signatures, code, data windows, and dependencies without logging secrets.
• Use point-in-time feature lookups to prevent future leakage into historical training rows.
• Automate batch and streaming feature computation with stable keys, timestamps, checkpoints, and idempotency.
• Configure online tables for low-latency feature access and on-demand features for request-time computation.
• Test offline-online parity, stale synchronization, missing keys, late data, and unauthorized access.

Promote code through isolated environments and prove each ML contract before changing production intent.

• Use deploy-code patterns with separate development, test, and production-like data, identities, and configuration.
• Map Workflows, MLflow Tracking, Unity Catalog models, aliases, Feature Engineering, Model Serving, and monitoring to lifecycle stages.
• Move reusable code into versioned modules and write deterministic unit tests for functions and edge cases.
• Design integration tests spanning feature engineering, training, evaluation, deployment, and inference.
• Define jobs, experiments, models, endpoint resources, variables, permissions, and targets with Databricks Asset Bundles.
• Validate and deploy the same reviewed commit with target-specific configuration and machine identities.

Turn data and model signals into governed retraining—not uncontrolled automatic promotion.

• Trigger retraining from sustained drift, degradation, schedule, or new-label conditions with exact context.
• Select challengers against an incumbent using fixed metrics, slices, uncertainty, operational limits, and approval rules.
• Choose snapshot, time-series, or inference monitoring from table semantics.
• Interpret numerical and categorical drift metrics, including distribution tests, with seasonality and effect size.
• Create custom business metrics and purposeful slices while bounding sparse groups and cardinality.
• Track endpoint latency, request rate, error rate, CPU, memory, and prediction outcomes.
• Alert with owner, persistence, context, runbook, and monitor-health state; never treat stale monitoring as green.

Package custom behavior, automate endpoints, control traffic, and demonstrate exact-version recovery.

• Implement an MLflow custom PyFunc and log required artifacts, dependencies, signature, and input example.
• Register immutable versions in Unity Catalog with least-privilege lifecycle ownership.
• Create, update, query, and delete endpoints through Model Serving interfaces, REST, or MLflow Deployments SDK.
• Use authenticated machine clients, validate request schemas, poll readiness, and run deployment smoke tests.
• Compare canary traffic splitting with blue-green environment switching.
• Record exact served versions because aliases are mutable and can change during long evaluations.
• Predefine rollout metrics, minimum samples, duration, stop conditions, rollback target, and cost ceiling.
• Complete all three projects, review 40 flashcards, and answer the independent bank by its exact 22/22/6 distribution.