Exploring and Preprocessing Data for ML PROFESSIONAL-MACHINE-LEARNING-ENGINEER Practice Questions (Page 2)

Questions 6–10

1. 6expert · hard

   A streaming media company needs to preprocess clickstream data for a real-time personalization model. The data arrives as a stream of 100,000 events per second. The preprocessing requires windowed aggregations (e.g., 'clicks in last 5 minutes per user') and enrichment with a static lookup table of content metadata. The team needs to minimize operational overhead and wants a fully managed solution. What should they use?

   • AApache Spark Streaming on a self-managed Dataproc cluster
   • BCloud Dataflow with a streaming pipeline that performs windowing and joins with the lookup table
   • CBigQuery with a scheduled query that runs every minute on the raw stream
   • DCloud Pub/Sub with a Cloud Function that processes each event individually
   Reveal Solution Discussion Select an answer first
2. 7application · medium

   A media company receives user interaction logs as a continuous stream of JSON events. The ML team needs to build a real-time feature for 'average session duration per user' to serve a recommendation model with sub-second latency. The team also needs to ensure that user IDs are hashed before the feature is stored. Which combination of tools should they use?

   • ACloud Pub/Sub alone to store the events and compute features on-demand in the serving layer
   • BCloud Dataflow for stream processing with a hash transform, writing to Gemini Enterprise Agent Platform Feature Store
   • CApache Spark on Dataproc with micro-batches, writing to a Cloud SQL database
   • DBigQuery for batch loading the logs daily, then a scheduled query to compute the feature
   Reveal Solution Discussion Select an answer first
4. 8foundation · easy

   A dataset contains a column with customers' email addresses. To protect PII during model training, which technique is most appropriate?

   • AEncrypt the email addresses with a reversible encryption algorithm.
   • BRemove the column entirely from the dataset.
   • CApply a hash function to the email addresses to create a pseudonymous identifier.
   • DReplace email addresses with the customer's full name.
   Reveal Solution Discussion Select an answer first
5. 9expert · hard

   A company has a 50 TB dataset of user behavior logs in Cloud Storage. The ML team needs to create a set of aggregate features for a recommendation model. The team has a tight budget and wants to minimize compute costs. The features need to be updated daily. What is the most cost-effective approach among the following?

   • AUse Cloud Dataflow with a batch pipeline and a large number of workers
   • BUse BigQuery to run SQL aggregation queries on the external data and write results to a feature store
   • CUse a single large VM with in-memory Python to process the data daily
   • DProvision a Dataproc cluster with 100 nodes and run Spark jobs daily
   Reveal Solution Discussion Select an answer first
6. 10foundation · easy

   A machine learning engineer is starting a new project and needs to decide how to organize the available data for efficient experimentation. The data consists of a CSV file with customer transactions, a folder of JPEG images, and a set of text documents. Which approach best organizes this data for ML workflows?

   • ALoad all data into memory as Python objects for direct access during experimentation.
   • BKeep tabular data in a structured format like CSV or BigQuery, store images in a cloud storage bucket, and store text documents in a format suitable for NLP, such as JSON lines or a document database.
   • CConvert all data to a single format, such as JSON, to ensure consistency across types.
   • DStore all data in a single BigQuery table, regardless of type, to simplify access.
   Reveal Solution Discussion Select an answer first