GCP ML Engineer (Google Cloud Certifications) study plan
GCP ML Engineer · Give each week a clear focus.
8 weeks · About 7 hours per week
Architect low-code AI solutions
Map the Vertex AI surface (Workbench, Datasets, Training, Model Registry, Endpoints, Pipelines, Model Monitoring) onto each stage of the ML lifecycle and write the mapping down from memory
Compare prebuilt Vision, Natural Language, Translation, and Speech-to-Text APIs against AutoML and custom training for three given business scenarios, and justify each pick in one paragraph
Build an AutoML tabular model end to end from a BigQuery table and deploy it to an endpoint
Draft a decision matrix that distinguishes BigQuery ML, AutoML, and custom container training by data type, team skill, latency, and cost
Your checkpoint
A one-page decision matrix plus a live endpoint that returns predictions for 5 held-out rows, with a written record of every configuration choice made
Scale prototypes into ML models
Configure and run a custom training job on Vertex AI from a packaged Python application, selecting machine type, worker count, and output location deliberately
Containerize a training script, push it to Artifact Registry, and run it as a custom container job with a custom prediction routine
Run a hyperparameter tuning job with an explicit search space, optimization metric, and early stopping, then read the trial results
Select between data-parallel training with a reduction server and model-parallel training for a stated model size, and write the reasoning
Your checkpoint
A training run summary showing the best hyperparameter trial, its objective value, and a measured comparison against the single-configuration baseline
Scale prototypes into ML models & Collaborate within and across teams to manage data and models
Cut training cost by switching to spot VMs and a smaller accelerator, then document the failure-handling and restart tradeoffs
Register two model versions in Model Registry, attach aliases, and promote the newer version to the production alias
Track and compare experiment runs so that metrics, parameters, and artifacts of two variants can be diffed side by side
Version a managed dataset and serve the same features from an online and offline feature store
Your checkpoint
A Model Registry holding two versions with a production alias pointing to the newer one, plus a two-run experiment comparison table
Collaborate within and across teams to manage data and models
Design and apply IAM bindings for three personas (data scientist, ML engineer, pipeline service account) using least privilege, then verify them with an actual permission test
Tag and catalog a dataset with a data governance layer so that lineage from source table to deployed model is traceable
Specify a train/validation/test split and leakage-prevention rule set for a time-ordered dataset
Decide when a project needs VPC Service Controls and customer-managed encryption keys, and state the residual risk if omitted
Your checkpoint
A written access-and-governance design document backed by applied IAM bindings in a sandbox project, evidenced by one successful and one correctly denied permission check
Serve and scale models
Deploy a model to an endpoint with autoscaling, minimum and maximum replica bounds, and a 90/10 traffic split between two versions
Run a batch prediction job over a large dataset and justify the chosen job shape against an online endpoint alternative
Configure a private endpoint with VPC peering and explain the scenario in which it is required
Estimate serving cost per 1000 predictions and tune machine type and accelerator to hit a target latency
Your checkpoint
A measured A/B deployment at a 90/10 split with observed latency percentiles and a per-1000-prediction cost estimate table
Automate and orchestrate ML pipelines
Author a Vertex AI Pipeline whose components cover data preparation, training, evaluation, and a conditional deployment step gated on a metric threshold
Trigger the pipeline on a schedule and on a data-arrival event, passing parameters through each path
Enable step caching and inspect metadata lineage to prove that an unchanged step did not re-execute
Choose between Vertex AI Pipelines and Cloud Composer for two stated orchestration scenarios and justify each
Your checkpoint
A scheduled pipeline that blocks deployment when the evaluation metric falls below threshold, with the execution lineage recorded and the cache hit demonstrated
Monitor AI solutions
Configure model monitoring against a training baseline with explicit skew and drift thresholds and an alerting channel
Build a Cloud Monitoring dashboard and alerting policy on prediction latency, error rate, and request volume
Enable feature attributions on an endpoint and interpret the top contributing features for a sample prediction
Define a retraining trigger policy that names the drift metric, threshold, window, and owner of the retrain
Your checkpoint
A simulated drift event that fires the configured alert, plus a written retraining trigger specification naming metric, threshold, window, and responsible role
Milestones
Week 2: Can train, containerize, and tune a model on Vertex AI without notes, and explain the cost and latency consequences of each configuration choice
Week 4: Score 75% or better on a self-written 25-question set covering the low-code solution design, data/model collaboration, and prototyping domains, with every miss converted into a one-line correction note
Week 6: Can whiteboard an end-to-end architecture from source data to a monitored endpoint, naming the GCP service at every hop and the reason for it
Week 8: Complete a full timed end-to-end build (data to monitored endpoint) inside one sitting with no unplanned blockers, and hold a domain-by-domain accuracy scorecard above 80% on the two initially weakest domains