Scaling and Tuning MACHINE-LEARNING-PROFESSIONAL Practice Questions (Page 2)

Questions 6–10

1. 6foundation · easy

   What is a primary advantage of horizontal scaling over vertical scaling for a distributed ML training workload?

   • AIt is always cheaper than vertical scaling.
   • BIt provides a way to scale beyond the maximum size of a single machine.
   • CIt simplifies the model training code.
   • DIt eliminates the need for data partitioning.
   Reveal Solution Discussion Select an answer first
2. 7foundation · easy

   What is a primary strength of Spark compared to Ray for machine learning workloads?

   • AIts ability to run arbitrary Python code with low latency and high flexibility.
   • BIts native support for deep learning frameworks like PyTorch.
   • CIts mature ecosystem for data processing, feature engineering, and standard ML algorithms on large tabular datasets.
   • DIts ability to manage stateful actors for long-running services.
   Reveal Solution Discussion Select an answer first
4. 8foundation · easy

   A team wants to use Ray Tune for hyperparameter tuning. They have a cluster of multiple machines. What is the primary mechanism Ray Tune uses to scale the tuning process across these machines?

   • AIt schedules each trial as a separate Ray task that can be executed on any available node in the cluster.
   • BIt requires the user to manually partition the hyperparameter space and assign each partition to a node.
   • CIt launches a separate JVM process on each node to run the trials.
   • DIt uses Spark's executors to run the trials in parallel.
   Reveal Solution Discussion Select an answer first
5. 9application · medium

   A team is training a linear regression model on a dataset with 10 million rows and 100 features. They are using SparkML on a Databricks cluster with 10 worker nodes. The training is taking too long. Which change is most likely to improve training speed?

   • AIncrease the number of worker nodes to 20.
   • BCache the training DataFrame in memory.
   • CUse a solver that is optimized for the dataset size, such as 'l-bfgs' instead of 'auto'.
   • DIncrease the number of partitions in the training DataFrame.
   Reveal Solution Discussion Select an answer first
6. 10application · medium

   A team is using Optuna to tune hyperparameters for a gradient-boosting model. They have a Databricks cluster with 8 worker nodes, each with 4 cores. They want to run 32 trials in parallel to speed up the tuning process. They plan to use Optuna's distributed mode with a MySQL database as the storage backend. What is the correct way to configure the number of parallel trials?

   • ACreate a single study with the MySQL storage, then launch 32 separate processes or notebooks, each calling study.optimize() with n_jobs=1.
   • BSet n_jobs=32 in the study.optimize() call and run the optimization on the driver node.
   • CSet the environment variable OPTUNA_N_TRIALS=32 and run a single optimize() call.
   • DUse Spark's map function to run study.optimize() on each of the 32 partitions of the training data.
   Reveal Solution Discussion Select an answer first