Sequencing a machine learning project from raw data to served predictions
A standard ML project on Google Cloud follows a repeatable lifecycle: define the business problem, collect and prepare data, train and evaluate a model, then deploy it for predictions and monitor performance. Google Cloud provides both AutoML (low-code, for users without ML expertise) and custom training (via Vertex AI, for data scientists) to support this workflow. Understanding the sequence and purpose of each phase is key for the exam, even without deep ML implementation knowledge.
Must-know
- The typical ML workflow order is: define objective/problem, extract/collect and prepare data, train the model, evaluate the model, deploy the model, and serve predictions (batch or online).
- Data collection and preparation (cleaning, labeling, splitting into training/validation/test sets) typically consumes the majority of project time and directly determines model quality.
- Vertex AI is Google Cloud's unified platform for the entire ML lifecycle, offering AutoML for automated model building and custom training for code-based model development.
- Model evaluation uses metrics appropriate to the problem type (e.g., accuracy, precision, recall, F1 for classification; RMSE/MAE for regression) and should be assessed against a held-out test set never used in training.
- Predictions can be served as batch predictions (large datasets, no immediate response needed) or online predictions (low-latency, real-time inference via deployed endpoints).
- A production ML project is iterative and includes ongoing monitoring for data/model drift, requiring periodic retraining, not a one-time train-and-deploy event.
A retail company's data practitioner has curated a dataset of customer purchase history that lives entirely inside BigQuery. The team needs to build a model that predicts whether a customer will make a repeat purchase within 30 days. They have strong SQL skills but no dedicated ML engineers, and leadership wants to avoid moving the data out of BigQuery. Which approach should the practitioner use to train the model?
What you have tried across GCP ADP's objectives, not a readiness score.
Data Preparation and Ingestion
- When to load first and when to transform first, and what sits between the two
- Picking a way to move existing data into Google Cloud
- Judging whether a dataset is trustworthy enough to build on
- Fixing messy records before they reach a report
- Telling CSV, JSON, Parquet, Avro, and relational tables apart, and where each fits
- Picking how to pull data out of a source system
- Matching a workload to the right storage or database service
- Getting files and tables loaded with a CLI, a transfer service, or a client library
Data Analysis and Presentation
- Writing BigQuery SQL that answers a reporting question
- Exploring and charting data inside a hosted notebook
- Turning a question from the business into an analysis that settles it
- Building a dashboard and getting it in front of the right people
- Deciding whether a job calls for Looker or for Looker Studio
- Editing LookML to change what a model exposes
- Spotting a problem worth solving with BigQuery ML or AutoML
- Calling a hosted Google language model straight from BigQuery
- Sequencing a machine learning project from raw data to served predictions
- Building, fitting, and scoring a model with SQL alone
- Running predictions against a model you already trained
- Keeping trained models catalogued in one place
Data Pipeline Orchestration
- Matching a transformation job to Dataproc, Dataflow, Dataform, or a managed alternative
- Weighing whether the transform belongs before or after the load
- Assembling the services a simple transformation pipeline needs
- Putting a query on a schedule and keeping it running
- Watching a Dataflow job and spotting where it stalls
- Reading logs and metrics to work out what a pipeline actually did
- Choosing what should drive a multi-step workflow
- Streaming messages into BigQuery as they arrive rather than in batches
- Wiring a trigger so one event starts the next step
Data Management
- Granting only the access a person or service actually needs
- Controlling who can read a bucket, and what uniform access changes
- Sharing a dataset with another team or company without copying it
- Matching a storage class to how often the data gets read
- Expiring old data automatically so it stops costing money
- Picking somewhere to park data that must be kept but is rarely read
- Comparing the managed backup and restore options across services
- Working out when a second copy is worth what it costs
- Regions, dual-regions, multi-regions, and zones as redundancy choices
- Deciding who should hold the encryption keys
- What a key management service does for creating, rotating, and revoking keys
- Protecting data on the wire versus data sitting on a disk
Coverage checked against the published exam guide on Aug 12, 2026.
These are independent practice questions, written against this certification's published exam guide. They are not the certification vendor's own questions, and not the real exam.