Databricks + MLflow

Question 1

What is the main role of Databricks in ML workflows?

Answer 1

To provide a scalable environment for feature engineering, model training, and batch scoring using Spark, Python, and integrated tooling like MLflow.

Question 2

Why is Databricks well-suited for feature engineering at scale?

Answer 2

It uses Spark DataFrames to process large datasets on clusters, enabling joins, aggregations, and transformations over TB-scale data efficiently.

Question 3

How can you decide between distributed training on Spark vs single-node training on Databricks?

Answer 3

Use distributed training when the data or model does not fit in a single machine’s memory; use single-node (e.g., scikit-learn, XGBoost) when data and model fit and distributed overhead is unnecessary.

Question 4

What is MLflow at a high level?

Answer 4

An open-source platform integrated into Databricks that tracks experiments, manages models, and packages them for reproducible deployment.

Question 5

What are the three main components of MLflow most relevant on Databricks?

Answer 5

MLflow Tracking, MLflow Models, and MLflow Model Registry.

Question 6

What is MLflow Tracking used for?

Answer 6

Logging parameters, metrics, artifacts, and code versions for model runs so they can be compared and reproduced.

Question 7

What is an MLflow run?

Answer 7

A single execution of training or evaluation code whose parameters, metrics, and artifacts are logged as a unit in the tracking system.

Question 8

What is an MLflow experiment?

Answer 8

A logical grouping of related runs, often associated with a project, notebook, or modeling task.

Question 9

Why is organizing runs into experiments important?

Answer 9

It makes it easier to compare models, search past runs, and maintain a tidy history of work by project or use case.

Question 10

What kinds of information are typically logged in an MLflow run?

Answer 10

Hyperparameters, training and validation metrics, model artifacts (pickled models, pipelines), plots, and environment information.

Question 11

What is MLflow autologging?

Answer 11

A feature that automatically records parameters, metrics, and models for certain ML libraries without requiring manual log calls.

Question 12

When is MLflow autologging especially convenient?

Answer 12

During iterative experimentation with libraries like scikit-learn, XGBoost, or Spark MLlib, where basic logging can be automated.

Question 13

Why might you still use explicit logging calls with MLflow even when autologging is enabled?

Answer 13

To log custom metrics, artifacts, or metadata that autologging does not capture by default.

Question 14

What is an MLflow Model?

Answer 14

A packaged model format that can include code, environment specifications, and multiple ‘flavors’ (e.g., Python function, Spark, sklearn) for deployment.

Question 15

What is MLflow Model Registry?

Answer 15

A centralized store for managing versions and lifecycle states of models, such as staging and production, with metadata and approvals.

Question 16

Why use a Model Registry instead of just saving models to files or paths?

Answer 16

It provides versioning, stage transitions (e.g., staging → production), audit trails, and a single source of truth for deployment targets.

Question 17

What are typical model stages in MLflow Model Registry?

Answer 17

None, Staging, Production, and Archived.

Question 18

Why is having a ‘Staging’ stage useful?

Answer 18

It allows testing models in non-production environments before promoting them to Production, enforcing a review workflow.

Question 19

How do Databricks Jobs relate to MLflow and model training?

Answer 19

Jobs can run training notebooks or scripts that log runs to MLflow, making training pipelines repeatable and schedulable.

Question 20

What is a typical pattern for a training job on Databricks?

Answer 20

Load features from Delta tables, train a model, evaluate it, log results and artifacts to MLflow, and optionally register or update a model in the registry.

Question 21

How can you implement hyperparameter tuning on Databricks?

Answer 21

By looping over parameter configurations or using libraries like Hyperopt or Spark ML tuning APIs, and logging each configuration as an MLflow run.

Question 22

Why is logging hyperparameter search results to MLflow valuable?

Answer 22

It allows comparison across configurations, visualizing metric trends, and re-running promising setups later.

Question 23

What is batch scoring on Databricks?

Answer 23

Applying a trained model to a large dataset stored in Delta or Parquet to generate predictions, often as a scheduled job.

Question 24

How is batch scoring commonly implemented with Spark on Databricks?

Answer 24

Load features as a DataFrame, apply a UDF or Pandas UDF that wraps the model, and write predictions back to a Delta table or other sink.

Question 25

What considerations are important when doing batch scoring with Python models on Spark?

Answer 25

Serialization and UDF overhead, broadcasting models if possible, and ensuring inference code is efficient and stateless per row or batch.

Question 26

What is online serving of ML models in a Databricks context?

Answer 26

Deploying models behind real-time endpoints using Databricks serving features, MLflow model serving, or external serving infrastructure that loads MLflow models.

Question 27

Why might a team choose external model serving (e.g., on Kubernetes) even when using Databricks?

Answer 27

For tighter integration with existing microservices, custom scaling policies, or when latency and protocol requirements exceed built-in serving options.

Question 28

How can MLflow Models help in external serving scenarios?

Answer 28

They provide a standardized way to package and load models in multiple environments, including Python functions and REST-serving wrappers.

Question 29

What is feature reuse in the context of Databricks and ML?

Answer 29

Using the same feature pipelines and tables across multiple models rather than re-implementing feature logic in each project.

Question 30

Why is feature reuse important?

Answer 30

It reduces duplication, improves consistency, and makes it easier to maintain and evolve feature logic in one place.

Question 31

What is a common pattern for organizing ML projects in Databricks repos?

Answer 31

Separate directories for data prep, model training, evaluation, and deployment scripts, all under version control with shared utils and config files.

Question 32

Why is separating configuration from training code helpful?

Answer 32

It allows the same code to run in different environments, with different data ranges or hyperparameters, without modifying code itself.

Question 33

How does MLflow integrate with experiment tracking in Databricks UI?

Answer 33

Experiments and runs appear in a dedicated Experiments interface where you can filter, compare, and visualize metrics and parameters.

Question 34

Why is experiment tracking crucial in collaborative environments?

Answer 34

It provides a shared history of modeling work, prevents duplicate efforts, and allows others to learn from or build on past runs.

Question 35

What is model lineage in the context of MLflow and Databricks?

Answer 35

The relationship between models, the data and code used to create them, and their deployments, tracked via run metadata and registry entries.

Question 36

How can you track data versions used in ML training on Databricks?

Answer 36

By logging dataset versions or Delta table version numbers and paths as run parameters or tags in MLflow.

Question 37

Why is tracking data versions important for ML reproducibility?

Answer 37

Models depend on the exact data snapshot; being able to reconstruct training data is essential for debugging and audits.

Question 38

What is a typical deployment flow using Model Registry on Databricks?

Answer 38

Train model → log to MLflow → register model → promote to Staging for tests → promote to Production when validated → use in batch/online serving.

Question 39

How can you roll back to a previous model in a registry-based workflow?

Answer 39

By changing which registered version is marked as Production or by directing serving infrastructure to load an earlier version.

Question 40

What is a good practice for ML model monitoring on Databricks?

Answer 40

Log predictions and key features to Delta tables, then use scheduled notebooks or jobs to compute drift and performance metrics over time.

Question 41

In one sentence, what is the core mental model for ML & MLflow on Databricks?

Answer 41

Use Databricks to build feature pipelines and run training at scale, and use MLflow to track experiments, manage model versions, and connect trained models to robust batch or online serving workflows.