Professional ML Engineer
Free Google Cloud Professional Machine Learning Engineer Practice Test
12 exam-style questions with full explanations — no sign-up. Score yourself, then close your gaps with the full course.
Exam fee ~$2006 exam domainsLevel Advanced2 timed practice tests in the course
✅ Free practice — no sign-up📝 Real exam-style questions💡 Detailed explanations💸 30-day money-back via Udemy
A retail company receives thousands of customer support emails daily and wants to automatically classify them by sentiment (positive, negative, neutral) with minimal ML engineering effort and no model training or labeled dataset. Which approach BEST meets this requirement?
The Natural Language API provides a pre-trained, ready-to-use sentiment analysis endpoint (analyzeSentiment) that requires zero training data, zero model management, and can be called immediately, matching the 'minimal effort, no training' requirement exactly.
A retail company built a fraud-detection model that must score transactions online with p99 latency under 10ms. The features (customer lifetime spend, rolling 7-day transaction count) are computed by a batch pipeline and must be identical between training and serving to avoid training-serving skew. Which solution BEST meets these requirements?
Vertex AI Feature Store is purpose-built for this scenario: it stores precomputed feature values with an online serving layer optimized for low-latency (single-digit ms) lookups, and the same feature definitions are used for both batch training reads and online serving, eliminating training-serving skew.
A team trains an image classification model on Vertex AI using 8 NVIDIA A100 GPUs spread across 4 worker nodes with synchronous data-parallel training in TensorFlow. Profiling shows that most of the step time is spent on gradient synchronization across nodes (all-reduce), not on compute. Which change should they make to reduce this communication bottleneck with minimal code changes?
Vertex AI Reduction Server is purpose-built to accelerate all-reduce operations in multi-node, multi-GPU synchronous data-parallel training. It runs dedicated reduction worker VMs that offload and parallelize the gradient aggregation, cutting network bottleneck time without requiring changes to the training script (it works transparently with NCCL-based frameworks like TensorFlow and PyTorch when the appropriate worker pool is added to the job spec).
You deploy a TensorFlow model to a Vertex AI online endpoint. Traffic is bursty overnight and near-zero for several hours each day, and you want to minimize idle cost while keeping p99 latency low during traffic spikes. Which deployment configuration should you use?
Vertex AI online endpoints require min_replica_count of at least 1 (they cannot scale to zero), and letting max_replica_count and the default 60% CPU utilization target handle bursts is the standard cost/latency balance for variable traffic.
A data science team built a Vertex AI Pipeline that retrains a fraud model daily using data that lands in BigQuery overnight. They want the simplest solution with the least additional infrastructure to run the pipeline on this fixed daily cadence. Which approach BEST meets this requirement?
Vertex AI Pipelines supports native recurring schedules (PipelineJobSchedule, exposed via the SDK's create_schedule or gcloud ai pipeline-job-schedules), letting you specify a cron expression to trigger runs without provisioning Composer environments, Cloud Functions, or VMs.
A company deploys a fraud detection model to a Vertex AI endpoint. They want to compare the distribution of incoming request features against the distribution of features used during model training, and receive an alert if the divergence exceeds a threshold. Which type of Vertex AI Model Monitoring should they enable?
Training-serving skew detection specifically compares the statistical distribution of features seen in production requests against the distribution of features in the original training dataset, which is exactly what the scenario requires.
A finance team needs to extract structured key-value pairs — invoice number, vendor name, total amount, due date — from scanned PDF invoices that arrive in many different layouts, without building or training a custom model. Which service should they use?
Document AI's specialized Invoice Parser is a pre-built processor trained specifically to extract structured, labeled fields (invoice number, totals, vendor, dates) from varied invoice layouts out of the box, with no training required.
A data engineering team has an existing, mature Apache Spark codebase performing complex ETL and feature engineering on petabytes of log data stored in Cloud Storage. The ML team wants to reuse this exact Spark logic with minimal rewrite to prepare training data on Google Cloud. Which service should they use?
Dataproc is a managed Hadoop/Spark service designed exactly for this use case — running existing Spark/Hadoop code on Google Cloud with little to no code rewrite, at scale.
A research team needs to train a very large Transformer model on TensorFlow using enormous matrix multiplication workloads, with a training budget that favors maximum throughput per dollar for sustained large-batch training over many days, and the model is written to be XLA-compatible from the start. Which hardware choice on Vertex AI is MOST appropriate?
TPUs are custom ASICs optimized specifically for large matrix multiplications and are most cost-efficient for sustained, large-batch, XLA-compiled TensorFlow training at scale, which matches every constraint given in the scenario.
A retailer needs to score 50 million user profiles once per day to generate recommendations for an overnight email campaign. There is no requirement for sub-second responses. Which Vertex AI approach minimizes cost and operational overhead?
Batch prediction jobs are designed for large, non-latency-sensitive scoring workloads, running only for the duration of the job with no standing endpoint cost, and integrate directly with BigQuery for input and output.
In Vertex ML Metadata, what does an Execution node represent within a pipeline's lineage graph?
Vertex ML Metadata models lineage using three node types: Artifacts (data/model outputs), Executions (a run of a step/component with its parameters), and Contexts (logical groupings like pipeline runs or experiments). An Execution specifically represents a single component run.
A team wants to detect when the relative importance of input features to a deployed model's predictions changes significantly over time, which could indicate the model is now relying on different signals than it did during training, even if raw feature values look similar. Which Vertex AI Model Monitoring capability should they enable?
Feature attribution drift monitoring, built on Vertex Explainable AI, tracks changes in feature attribution scores over time and alerts when the relative influence of features shifts, independent of raw feature value distributions.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$129.99 $34.99 with code FREETEST33 — valid through September 7.
Get my $34.99 deal →Professional ML Engineer exam — quick answers
How much does the Professional ML Engineer exam cost?
The exam fee is approximately $200 and varies by region — confirm current pricing with the certification vendor before you book.
What topics are on the exam?
It covers 6 domains: Architecting low-code AI solutions (13%), Collaborating to manage data and models (14%), Scaling prototypes into ML models (21%), Serving and scaling models (18%), Automating and orchestrating ML pipelines (18%), Monitoring AI solutions (16%). The full course has a dedicated chapter, lab and practice-test coverage for each.
Is this practice test really free?
Yes — all questions on this page are free with explanations and no sign-up. The paid Udemy course adds two full-length timed exams, video lessons and hands-on labs.
How do I get the discount?
Use code FREETEST33 at checkout for $34.99 (list $129.99) through September 7 — the enroll button applies it automatically.
Will this prepare me for the real exam?
The questions mirror the real exam's style and are mapped to the official domains. This is exam-focused preparation — combine the free test with the full course's timed simulations to gauge your readiness.