A team is using Optuna to tune hyperparameters for a model. They want to log each trial's parameters and metrics to MLflow. They are running the tuning on a Databricks cluster. What is the recommended way to integrate Optuna with MLflow?

Manually call mlflow.log_params() and mlflow.log_metric() inside the objective function.

Use Ray Tune's MLflow integration instead.

Use mlflow.autolog() which automatically integrates with Optuna.

Use the MLflowCallback provided by Optuna's integration module.

A team is training a very large transformer model with 100 billion parameters. They have a cluster with 64 GPUs, each with 40 GB memory. The model does not fit on a single GPU. They need to choose between data parallelism and model parallelism. What is the primary consideration?

A combination of both is required, but model parallelism is the primary strategy.

Model parallelism is required because the model is too large for a single GPU.

Data parallelism is preferred because it is easier to implement and scales well with the number of GPUs.

Use pipeline parallelism, which is a form of model parallelism.

A team needs to train a custom reinforcement learning (RL) environment that involves complex simulation and requires frequent communication between the simulation and the model. The training loop is not easily expressible as a Spark job. They are considering using Ray or Spark. Which framework is more suitable?

Use Spark for data preprocessing and Ray for training, combining both.

Both are equally suitable; the choice depends on the team's familiarity.

Ray, because it supports flexible task scheduling and stateful actors, which are well-suited for RL.

Spark, because it is integrated with Databricks and can handle large-scale data processing.

A company has a Databricks workload that trains a model on a dataset that grows by 10% each month. Currently, the training job runs on a single node with 64 GB RAM and takes 8 hours. The team expects the data to double in the next year. They have a fixed budget and need to keep the training time under 12 hours. They are considering two options: (1) upgrade to a node with 128 GB RAM, or (2) add a second node and use Spark to distribute the training. The model training algorithm is not easily parallelizable. Which option is more appropriate?

Switch to a different algorithm that is parallelizable.

Add a second node and use Spark to distribute the training.

Use a combination of both: upgrade RAM and add a node.

Upgrade to a node with 128 GB RAM.

DatabricksCertified Machine Learning Professional

Domain 1Objective 2

Scaling and Tuning MACHINE-LEARNING-PROFESSIONAL Practice Questions (Page 4)

Part of the Model Development domain, which accounts for 44% of the MACHINE-LEARNING-PROFESSIONAL exam. Databricks does not publish an official question count, but from its 120-minute exam (~50–80 total, ~22–35 in this domain), expect 6–9 from this objective — we provide 23 practice questions to prepare you well beyond it. (estimate)

23questions here

5free pages

7concepts

44%of the exam

Questions 16–20

16application · medium
A team is using Optuna to tune hyperparameters for a model. They want to log each trial's parameters and metrics to MLflow. They are running the tuning on a Databricks cluster. What is the recommended way to integrate Optuna with MLflow?
Select an answer first
17expert · hard
A team is training a very large transformer model with 100 billion parameters. They have a cluster with 64 GPUs, each with 40 GB memory. The model does not fit on a single GPU. They need to choose between data parallelism and model parallelism. What is the primary consideration?
Select an answer first
18expert · hard
A team needs to train a custom reinforcement learning (RL) environment that involves complex simulation and requires frequent communication between the simulation and the model. The training loop is not easily expressible as a Spark job. They are considering using Ray or Spark. Which framework is more suitable?
Select an answer first
19foundation · easy
Which component of the Spark MLlib pipeline is responsible for transforming raw data into features that can be used by a distributed ML algorithm?
Select an answer first
20expert · hard
A company has a Databricks workload that trains a model on a dataset that grows by 10% each month. Currently, the training job runs on a single node with 64 GB RAM and takes 8 hours. The team expects the data to double in the next year. They have a fixed budget and need to keep the training time under 12 hours. They are considering two options: (1) upgrade to a node with 128 GB RAM, or (2) add a second node and use Spark to distribute the training. The model training algorithm is not easily parallelizable. Which option is more appropriate?
Select an answer first

Finished these 5 questions?

Review the revealed explanations, or continue through the curriculum.

Next 5 questions →Back to MACHINE-LEARNING-PROFESSIONAL

Free Basic Practice is a study aid with revealable answers — not a scored exam. Examers.io is independent and not affiliated with or endorsed by Databricks. “MACHINE-LEARNING-PROFESSIONAL” is a trademark of its owner, used for identification only.

Scaling and Tuning MACHINE-LEARNING-PROFESSIONAL Practice Questions (Page 4)

Questions 16–20

A team is using Optuna to tune hyperparameters for a model. They want to log each trial's parameters and metrics to MLflow. They are running the tuning on a Databricks cluster. What is the recommended way to integrate Optuna with MLflow?

A team is training a very large transformer model with 100 billion parameters. They have a cluster with 64 GPUs, each with 40 GB memory. The model does not fit on a single GPU. They need to choose between data parallelism and model parallelism. What is the primary consideration?

Which component of the Spark MLlib pipeline is responsible for transforming raw data into features that can be used by a distributed ML algorithm?

Keep studying — Model Development