Skip to content

Machine Learning Associate · Domain 1: Databricks Machine Learning · 38% of the exam

Task 1.1: Identify the best practices of an MLOps strategy; Identify the advantages of using ML runtimes

Separate development, staging and production environments mapped to catalogs, deploying code rather than models, Git and MLflow in the workflow, and what Databricks Runtime for Machine Learning installs and configures for you.

Study it

  • MLOps on Databricks: environments as catalogs, deploy code, Git and MLflow

    Partly covered by: From Pilot to Production

  • Databricks Runtime for Machine Learning: preinstalled libraries and compute settings

    Lesson coming

Sample questions

Try each one before opening the answer. Every option is explained, with the Databricks documentation page that proves it.

Question 1 · choose 1

A multitask Databricks job trains a churn model in its first task and registers a new version of prod.ml_team.churn in Unity Catalog. The second task must validate exactly the version that the first task registered in the same job run, even when another training run registers a version of the same model at about the same time. How should the validation task find that version?

  1. AThe training task sets a task value with the model URI, and the validation task reads that task value
  2. BThe validation task calls search_model_versions ordered by version number and takes the newest version
  3. CThe validation task loads the model URI models:/prod.ml_team.churn@Champion
  4. DThe validation task loads the version that holds the Challenger alias with get_model_version_by_alias
Show the answer and why
  • AThe training task sets a task value with the model URI, and the validation task reads that task value

    Correct

    The recommended workflow passes the model URI that the training task yields to the validation task through task values, so it always gets the version from its own run.

  • BThe validation task calls search_model_versions ordered by version number and takes the newest version

    Incorrect

    order_by is not supported in search_model_versions for models in Unity Catalog, and the newest version could come from the other run.

  • CThe validation task loads the model URI models:/prod.ml_team.churn@Champion

    Incorrect

    Champion points to the version currently in production, not to the new version that still has to pass validation.

  • DThe validation task loads the version that holds the Challenger alias with get_model_version_by_alias

    Incorrect

    In the recommended workflow the Challenger alias is assigned only after the version passes validation, so it does not yet point to the new version.

In the recommended MLOps workflow on Databricks, the model training task yields a model URI that the model validation task can use, and task values pass that URI between the tasks. The validation pipeline assigns the Challenger alias only to a version that passes its checks, and Champion marks the production version.

Question 2 · choose 2

A team creates classic compute on Databricks Runtime 17.3 LTS ML to train models with the Python xgboost package on features that it reads from Unity Catalog tables. Which two statements about this compute are true? (Choose TWO.)

  1. ASelecting the Machine learning checkbox sets the access mode to Dedicated, which Databricks Runtime ML needs to read Unity Catalog data
  2. BStandard access mode is required for Databricks Runtime ML to read Unity Catalog tables
  3. CGraviton instance types cannot be used with Databricks Runtime ML
  4. DPhoton can speed up the Spark DataFrame feature engineering, but not the training done by the Python xgboost package
  5. EMLflow must be installed with %pip before the training runs can be tracked
Show the answer and why
  • ASelecting the Machine learning checkbox sets the access mode to Dedicated, which Databricks Runtime ML needs to read Unity Catalog data

    Correct

    The checkbox sets Dedicated access mode automatically, and Dedicated is required to access Unity Catalog data on Databricks Runtime ML.

  • BStandard access mode is required for Databricks Runtime ML to read Unity Catalog tables

    Incorrect

    The page requires Dedicated access mode for Unity Catalog data on Databricks Runtime ML.

  • CGraviton instance types cannot be used with Databricks Runtime ML

    Incorrect

    Databricks Runtime 15.4 LTS ML and above support Graviton instance types.

  • DPhoton can speed up the Spark DataFrame feature engineering, but not the training done by the Python xgboost package

    Correct

    Photon improves Spark SQL, DataFrames and feature engineering, and is not expected to improve Python packages such as XGBoost.

  • EMLflow must be installed with %pip before the training runs can be tracked

    Incorrect

    MLflow is pre-installed on Databricks Runtime ML clusters.

Databricks Runtime for Machine Learning sets Dedicated access mode when you select the Machine learning checkbox, and that mode is what lets it read Unity Catalog data. Photon (15.2 ML and above) helps Spark DataFrame and feature engineering work but not Python libraries such as XGBoost, PyTorch or TensorFlow. Graviton is supported from 15.4 LTS ML, and MLflow comes pre-installed.

Question 3 · choose 1

Databricks describes MLOps as a set of processes and automated steps for managing code, data and models. Which three practices does it say MLOps combines?

  1. ADataOps, FinOps and SecOps
  2. BDevOps, FinOps and ModelOps
  3. CDevOps, DataOps and platform operations
  4. DDevOps, DataOps and ModelOps
Show the answer and why
  • ADataOps, FinOps and SecOps

    Incorrect

    FinOps and SecOps are not part of Databricks' definition of MLOps.

  • BDevOps, FinOps and ModelOps

    Incorrect

    FinOps is not one of the three; the definition names DataOps instead.

  • CDevOps, DataOps and platform operations

    Incorrect

    Platform operations is not one of the three; the definition names ModelOps.

  • DDevOps, DataOps and ModelOps

    Correct

    Databricks defines MLOps as combining DevOps, DataOps and ModelOps.

MLOps is a set of processes and automated steps for managing code, data and models to improve the performance, stability and long-term efficiency of ML systems; it combines DevOps, DataOps and ModelOps.

Question 4 · choose 1

In the staging stage of Databricks' recommended MLOps workflow, integration tests run every pipeline, including model training. Training the model is expensive and slow. What does Databricks suggest to keep the integration tests affordable?

  1. AUse small subsets of data or fewer training iterations in the tests
  2. BSkip the model training pipeline in staging and test it in production
  3. CRun only unit tests and merge to the main branch
  4. DRun the integration tests in the development workspace on a laptop sample
Show the answer and why
  • AUse small subsets of data or fewer training iterations in the tests

    Correct

    Integration tests can trade fidelity for speed or cost, for example with small data subsets or fewer training iterations.

  • BSkip the model training pipeline in staging and test it in production

    Incorrect

    Integration tests run all pipelines, including model training, in staging.

  • CRun only unit tests and merge to the main branch

    Incorrect

    The CI process runs integration tests after the unit tests.

  • DRun the integration tests in the development workspace on a laptop sample

    Incorrect

    Integration tests run in a staging environment that should match production as closely as is reasonable.

In the staging stage, the CI process runs integration tests of all pipelines (feature engineering, training, inference and monitoring) in an environment that matches production as closely as is reasonable. To reduce their time, some steps can trade fidelity for speed or cost, for example using small subsets of data or fewer training iterations when models are expensive to train.

Question 5 · choose 1

In Databricks' recommended MLOps practices, where should ML pipelines and their code be kept, and which Databricks feature connects them to the workspace?

  1. AIn MLflow experiments, as artifacts of the training runs
  2. BIn Git, synced with Databricks Git folders
  3. CIn Delta tables in the production catalog
  4. DIn Unity Catalog volumes, one per environment
Show the answer and why
  • AIn MLflow experiments, as artifacts of the training runs

    Incorrect

    MLflow tracks runs, parameters, metrics and models; pipelines and code belong in Git.

  • BIn Git, synced with Databricks Git folders

    Correct

    Databricks recommends Git for version control and Git folders to sync notebooks and source code with the workspace.

  • CIn Delta tables in the production catalog

    Incorrect

    Delta tables are recommended for raw data and feature tables, not for code.

  • DIn Unity Catalog volumes, one per environment

    Incorrect

    The recommendation for code and pipelines is Git with Git folders.

Databricks recommends storing pipelines and code in Git for version control, so that moving ML logic between stages is moving code between branches, and using Databricks Git folders to integrate with the Git provider. Data and feature tables live in Delta tables, and MLflow tracks model development.

Question 6 · choose 1

Databricks designates a subset of the libraries in Databricks Runtime for Machine Learning as top-tier libraries. What does that designation give them?

  1. AThey are the only libraries that can be imported in notebooks
  2. BThey are pinned to one version for the whole life of the runtime line
  3. CThey can be removed in any minor release without notice
  4. DA faster update cadence with each runtime release, plus advanced support, testing and embedded optimizations
Show the answer and why
  • AThey are the only libraries that can be imported in notebooks

    Incorrect

    Databricks Runtime ML includes many other libraries, and you can install more.

  • BThey are pinned to one version for the whole life of the runtime line

    Incorrect

    Top-tier libraries are updated to the latest package releases with each runtime release.

  • CThey can be removed in any minor release without notice

    Incorrect

    Top-tier libraries are added or removed only with major releases.

  • DA faster update cadence with each runtime release, plus advanced support, testing and embedded optimizations

    Correct

    Databricks updates top-tier libraries with each runtime release, barring dependency conflicts, and gives them advanced support and optimizations.

Databricks Runtime ML comes with popular ML libraries pre-installed. For the top-tier subset, Databricks provides a faster update cadence (the latest package releases with each runtime release), advanced support, testing and embedded optimizations; top-tier libraries are added or removed only with major releases.

Question 7 · choose 1

A team plans to fine-tune a deep learning model on GPU compute. Which compute does Databricks recommend for custom deep learning workloads like this one?

  1. AA CPU cluster on Databricks Runtime ML with Graviton instances
  2. BAI Runtime, the serverless GPU compute offering
  3. CA standard Databricks Runtime cluster without the ML runtime
  4. DA SQL warehouse with a large cluster size
Show the answer and why
  • AA CPU cluster on Databricks Runtime ML with Graviton instances

    Incorrect

    Graviton instances are CPU instances; for GPU deep learning Databricks points to AI Runtime.

  • BAI Runtime, the serverless GPU compute offering

    Correct

    For custom GPU deep learning such as fine-tuning, Databricks recommends AI Runtime instead of Databricks Runtime ML.

  • CA standard Databricks Runtime cluster without the ML runtime

    Incorrect

    The standard runtime lacks the ML libraries, and Databricks points GPU deep learning to AI Runtime.

  • DA SQL warehouse with a large cluster size

    Incorrect

    SQL warehouses run SQL workloads, not deep learning training.

Databricks recommends AI Runtime, a serverless GPU compute offering optimized for deep learning with simpler setup and faster provisioning, instead of Databricks Runtime ML for custom GPU workloads such as model fine-tuning or training.

Question 8 · choose 1

A team of four data scientists wants to share one Databricks Runtime ML compute resource with Dedicated access mode while each person keeps only the team's permissions on it. How does Databricks describe this setup?

  1. AAssign the compute to a group; each user's permissions down-scope to the group's
  2. BSwitch to Standard access mode, which Databricks Runtime ML needs for sharing
  3. CCreate four separate compute resources, because Dedicated compute serves one user only
  4. DGive every member workspace admin rights on the compute
Show the answer and why
  • AAssign the compute to a group; each user's permissions down-scope to the group's

    Correct

    A Dedicated resource assigned to a group down-scopes each user to the group's permissions, so members can share it securely.

  • BSwitch to Standard access mode, which Databricks Runtime ML needs for sharing

    Incorrect

    Sharing does not need another access mode; a Dedicated resource can be assigned to a group.

  • CCreate four separate compute resources, because Dedicated compute serves one user only

    Incorrect

    Dedicated compute can be assigned to a single user or to a group.

  • DGive every member workspace admin rights on the compute

    Incorrect

    Sharing is handled by assigning the compute to a group, not by admin rights.

When you select the Machine learning checkbox, the access mode is set to Dedicated with your account as the user, and you can assign the resource to a different user or a group. When it is assigned to a group, each user's permissions automatically down-scope to the group's permissions, so the resource can be shared securely with the other members.

Practice domain 1 →Practice all domains →