PMLE Tests the Least-Effort Path from Data to Monitored Model — Not Algorithm Trivia
For engineers who ship and run ML on Google Cloud, using the June 2026 guide's Agent Platform names. The shift: honour every stated constraint (skills, latency, explainability, budget) before reaching for custom training or more hardware.
Check Your Readiness →Most candidates understand Google Cloud Professional Machine Learning Engineer concepts — and still fail. This exam tests how you apply knowledge under pressure.
Scenarios give a goal plus constraints, and several options would technically work. The scored answer meets every constraint with the most managed, lowest-effort service. Agent Platform was formerly Vertex AI and the Airflow service was formerly Cloud Composer. Published weights round to 101%.
Vision API text detection or training an AutoML model
A Document AI processor returns structured key-value fields rather than raw OCR text and needs no ML expertise. Prebuilt processors and task APIs come before AutoML, which still requires labelled training data.
Switch to the largest Gemini model and lengthen the prompt
Prompt design has been tried, so supervised fine-tuning, preferably parameter-efficient LoRA, is the next step for fixed format and tone. Tuning changes behaviour; it should never be used to inject frequently changing facts such as catalogues.
Deploy an online endpoint and call it in a loop
Batch prediction reads from BigQuery or Cloud Storage and writes results back with no always-on endpoint billing. Online endpoints suit real-time requests, where minimum replicas prevent cold starts during predictable peaks.
Use integrated gradients feature attributions
Integrated gradients need a differentiable model such as a neural network. Sampled Shapley suits tree ensembles, and example-based explanations surface similar training cases to spot mislabelled or missing data.
When data already sits in BigQuery and the team writes SQL, BigQuery ML trains in place with CREATE MODEL and ML.PREDICT. Its TRANSFORM clause keeps preprocessing identical at prediction time, and ARIMA_PLUS handles many time series.
Low accelerator utilisation with a busy CPU means the input pipeline is the bottleneck. Fix it with sharded files, parallel reads and prefetching first; extra GPUs would sit idle. OOM errors need a smaller batch, not a bigger one.
Continuous training can produce a worse model. Add data validation before training and an evaluation gate that compares the candidate with production, then register the version in Model Registry and release it through a canary traffic split.
Skew right after launch is usually a preprocessing mismatch between training and serving code; retraining will not fix it. Data drift compares recent serving windows, while concept drift needs ground-truth labels because inputs may look unchanged.
Safety filters block harmful content categories. Model Armor screens prompts and responses for injection, jailbreaks and sensitive data such as account numbers, and VPC Service Controls limit exfiltration, which encryption keys alone cannot prevent.
Strengthen weak areas with exam-style practice questions and detailed explanations.
Simulate the real exam experience and assess your readiness under timed conditions.
Review key concepts, objectives, and exam topics in one place.
Get personalized explanations, learning recommendations, and instant answers.
Follow a structured learning path designed to help you prepare efficiently.
See whether you pick the managed ML option that meets every constraint in each PMLE scenario.