GCP ML Engineer (Google Cloud Certifications) cheat sheet

Vertex AI AutoML trains tabular, image, text, and video models from labeled data without custom training code; the console or API produces a managed model artifact.

BigQuery ML creates, trains, evaluates, and predicts with models using SQL statements such as CREATE MODEL; supported types include linear regression, logistic regression, XGBoost, k-means, and imported TensorFlow models.

Generative AI Studio in Vertex AI provides prompt design, model tuning, and evaluation for Gemini models without requiring serving infrastructure.

Vertex AI Agent Builder combines Vertex AI Search and Conversation to build enterprise search and chat agents grounded in private data with low-code configuration.

Document AI extracts structured entities from documents using prebuilt or custom processors for OCR, forms, invoices, and identity documents.

Dialogflow CX builds conversational agents using flows, pages, and intents and integrates with telephony, chat, and messaging channels.

Vertex AI Model Garden provides discoverable foundation models and task-specific models that can be deployed or tuned within Vertex AI.

Decision rule: if the task is tabular prediction and data resides in BigQuery, BigQuery ML is the lowest-code path; if custom preprocessing or an unsupported framework is required, use Vertex AI custom training.

Dataplex provides centralized metadata discovery, data quality checks, lineage, and lifecycle management across BigQuery, Cloud Storage, and BigLake.

Data Catalog, now part of Dataplex, stores searchable metadata entries and tags for datasets, tables, and models across projects.

Vertex AI Model Registry is a central repository for model versions, artifacts, and deployment state and supports model lineage tracking.

Vertex AI Experiments records parameters, metrics, and artifacts for each run so teams can compare experiments and reproduce results.

IAM roles such as Vertex AI User, BigQuery Data Viewer, and Storage Object Viewer control access; permissions are scoped per project and resource.

Vertex AI Feature Store shares feature values across teams for online and offline serving and reduces duplicate feature engineering.

Cloud Storage buckets and BigQuery datasets are common shared data boundaries; IAM policies and dataset-level access controls govern cross-team use.

Artifact Registry stores container images and Python packages used by training and prediction; versioned tags support reproducible deployments.

Vertex AI custom training runs a containerized training application on managed compute; you specify a prebuilt or custom container and a machine type.

Distributed training on Vertex AI supports reduction server for multi-worker GPU training and TPU pods for TensorFlow and JAX workloads.

Vertex AI Hyperparameter Tuning runs parallel trials using Bayesian optimization or grid and random search and reports the best trial metrics.

Vertex AI Pipelines orchestrates training steps as a directed acyclic graph; each step runs in a container and artifacts flow through Vertex ML Metadata.

BigQuery ML can train models at scale using SQL, supports hyperparameter tuning through SQL, and can export models for online serving.

Feature Store ingestion from BigQuery or Cloud Storage provides point-in-time correct features for training and low-latency online serving.

Vertex AI model evaluation reports metrics such as confusion matrix, precision, recall, ROC/AUC, and feature attributions depending on model type.

Decision rule: use custom training when a framework or preprocessing step is unsupported by AutoML; use AutoML when data is standard and time-to-model matters.

Vertex AI Endpoints serve online predictions with autoscaling based on traffic; each endpoint can host multiple model versions through traffic splits.

Batch prediction reads input from BigQuery or Cloud Storage and writes predictions without an endpoint, which suits large offline scoring jobs.

Vertex AI prediction containers expose an HTTP contract; prebuilt containers support frameworks including TensorFlow, PyTorch, scikit-learn, and XGBoost.

Traffic splitting on an endpoint enables canary releases and A/B testing by assigning percentage shares to model versions.

Private endpoints use VPC peering and do not expose a public IP; access requires configured network connectivity and IAM permissions.

Model monitoring on endpoints detects training-serving skew and prediction drift using configurable thresholds and baseline statistics.

GPUs and accelerators on endpoints accelerate inference for large models; machine type and accelerator are selected at deployment time.

Decision rule: use online endpoints for low-latency per-request predictions; use batch prediction for throughput-oriented offline scoring.

Vertex AI Pipelines runs Kubeflow Pipelines or TFX pipelines on managed infrastructure; pipelines are compiled to JSON and submitted for execution.

Cloud Composer is managed Apache Airflow for scheduling workflows that span BigQuery, Vertex AI, Cloud Storage, and other Google Cloud services.

Cloud Workflows orchestrates HTTP and Google Cloud API calls using YAML or JSON definitions and suits lightweight service orchestration.

Eventarc routes events from Cloud Storage, Pub/Sub, and Audit Logs to triggers such as Cloud Run services and Workflows.

Cloud Scheduler creates cron jobs that invoke HTTP endpoints, Pub/Sub topics, or Workflows on a defined schedule.

Cloud Build runs CI/CD steps from a build configuration; it can build containers, push images to Artifact Registry, and deploy models.

Vertex AI Pipelines integrates with Vertex ML Metadata to record artifacts, parameters, and lineage for every pipeline run.

Decision rule: use Vertex AI Pipelines for ML DAGs with artifacts and lineage; use Cloud Composer when the workflow spans many non-ML systems and needs Airflow operators.

Vertex AI Model Monitoring compares training data distributions with serving data to detect training-serving skew and prediction drift.

Skew is a mismatch between training and serving feature distributions; drift is a change in the serving distribution over time.

Monitoring jobs write results to BigQuery and can emit alerts through Cloud Monitoring notification channels.

Explainable AI on Vertex AI provides feature attributions using methods such as Sampled Shapley, Integrated Gradients, and XRAI for supported models.

Cloud Monitoring collects metrics, logs, and uptime checks; alerting policies notify through email, SMS, Pub/Sub, or PagerDuty.

Cloud Logging stores prediction request logs, while audit logs record admin activity and data access for compliance.

Model evaluation metrics should be monitored alongside operational metrics such as latency, error rate, and CPU or GPU utilization.

Decision rule: configure skew detection when serving preprocessing may differ from training; configure drift detection when input populations change over time.