GCP ML Engineer (Google Cloud Certifications) cheat sheet

No labeled data and a generic task (OCR, translation, speech, vision labels) maps to a pretrained API - Vision, Natural Language, Speech-to-Text, Translation - not AutoML.

Labeled data with no ML code maps to Vertex AI AutoML; supported tasks include image classification, object detection, tabular regression/classification/forecasting, and text classification, entity extraction, and sentiment.

Labeled data plus custom framework code maps to Vertex AI custom training with a Python package and a prebuilt or custom container.

Vertex AI Studio covers prompt design, model evaluation, and Gemini tuning (supervised fine-tuning, adapter tuning) with no training infrastructure for the user to manage.

Document AI returns structured entities from processors such as Form Parser, OCR, Invoice Parser, and specialized industry parsers; a custom extractor requires roughly ten or more labeled documents per field.

BigQuery ML trains in-warehouse models with CREATE MODEL ... OPTIONS(model_type=...) and scores with ML.PREDICT - the lowest-code path when data already lives in BigQuery.

Dialogflow CX models conversations as flows, pages, and intents; Vertex AI Agent Builder and Conversational Agents ground Gemini on indexed data stores for generative agents.

Recommendations AI and Vertex AI Search for commerce require imported catalog and user-event data and expose no user-facing model training step.

Vertex AI Model Garden supplies deployable foundation models; sustained high-volume Gemini traffic requires provisioned throughput rather than the shared, quota-limited endpoint.

Dataplex structures a lake into zones - a raw zone for ingested assets and a curated zone for cleansed, governed data - with assets attached to zones and discovered automatically.

Data Catalog, now part of Dataplex, provides searchable technical metadata; policy tags drawn from a taxonomy enforce BigQuery column-level access through the fine-grained reader role.

Analytics Hub publishes BigQuery datasets as listings so other teams or organizations query the data in place, with no copy and no duplicated storage cost.

Vertex AI Model Registry stores model versions with aliases such as 'default', evaluation metrics, and lineage back to the training run that produced each version.

Vertex ML Metadata automatically records artifacts, executions, and contexts for Vertex AI Pipelines runs and custom jobs, giving cross-team traceability of data and models.

BigQuery separates roles: BigQuery Data Viewer reads rows, BigQuery Job User runs queries, BigQuery Data Editor writes rows - a collaborator who only queries needs Job User plus Data Viewer.

VPC Service Controls perimeters block data exfiltration from Vertex AI, BigQuery, and Cloud Storage, while Access Transparency logs record Google support access.

Colab Enterprise and Vertex AI Workbench notebooks attach to shared runtime templates and Git repositories, so teams share environment and code instead of duplicating setup.

A Vertex AI custom training job is specified by workerPoolSpecs (machine type, accelerator type and count, replica count), a container image, and the Python package or args to run.

Distributed strategy choice: MirroredStrategy for single-node multi-GPU, MultiWorkerMirroredStrategy for multi-node, and a reduction server to offload all-reduce when GPU counts are large.

Vertex AI hyperparameter tuning (Vizier) requires metricId, goal (MINIMIZE or MAXIMIZE), a parameter spec with scale (linear, log, categorical), and maxTrialCount plus parallelTrialCount.

Spot (preemptible) VMs cut training cost sharply but can be reclaimed mid-run, so write checkpoints to Cloud Storage and configure the job to resume or restart.

Training data can be referenced from BigQuery tables or Cloud Storage URIs, or managed as a Vertex AI Dataset; referencing avoids duplication but couples the job to source schemas.

For temporal data, split train/validation/test by time; random splits leak future information and inflate evaluation metrics relative to production performance.

Push warehouse-scale feature engineering into BigQuery or Dataflow, and use Vertex AI Feature Store when identical feature values must feed both training and online prediction.

Vertex AI Model Evaluation reports metrics for AutoML models; for custom models compute metrics inside a pipeline component and log them to Vertex ML Metadata or the Model Registry so promotion gates can read them.

A deployment sets machine type, accelerator, min/max replica count, and an autoscaling metric with a target value such as 60 percent CPU utilization; one endpoint can host multiple deployed models for traffic splitting.

Set minimum replica count above zero for latency-sensitive endpoints to avoid cold starts; scale-to-zero lowers cost but adds seconds of load latency on the first request.

Batch prediction reads from BigQuery tables or Cloud Storage files and writes predictions to BigQuery or Cloud Storage; it needs no endpoint and fits when latency is not a constraint.

Private endpoints via Private Service Connect expose an internal IP inside the VPC and are mandatory inside VPC Service Controls perimeters; public endpoints are the default otherwise.

A custom prediction container must listen on 0.0.0.0 at the port Vertex AI injects (commonly 8080), implement health and predict routes, and return JSON responses.

Traffic split assigns a percentage of requests to each deployed model on an endpoint, supporting canary rollout and gradual promotion without parallel endpoints.

Model Garden foundation models deploy to dedicated endpoints or run through provisioned throughput; shared public endpoints are constrained by project quota and rate limits.

Latency reduction levers: co-locate the endpoint region with the data and users, shrink or quantize the model, raise the replica floor, and precompute or cache features.

GPU serving on A100, L4, or T4 requires a container with matching CUDA libraries and a model loaded before the health check begins returning healthy.

Vertex AI Pipelines runs Kubeflow Pipelines or TFX specs serverlessly; each component declares typed inputs and outputs and execution lineage is written to Vertex ML Metadata.

With caching enabled, a component whose code and input artifact fingerprints are unchanged reuses its previous outputs; disable caching for components with external side effects.

Recurring runs are submitted through the Pipelines API with a cron schedule, or triggered externally by Cloud Scheduler, Eventarc (for example a Cloud Storage object-finalize event), or Cloud Composer.

Cloud Composer (managed Airflow) fits cross-service DAGs spanning BigQuery, Dataproc, Dataflow, and Vertex AI; Vertex AI Pipelines fits ML-only DAGs that need managed lineage.

Artifact Registry stores the training and serving container images and Python packages, and Cloud Build builds and pushes them as the continuous-integration step.

Parameterize the compiled pipeline template and pass parameter_values at submission so one spec serves development, staging, and production.

Custom training jobs and hyperparameter tuning jobs can be wrapped as pipeline components, so retraining, evaluation, and deployment form a single DAG.

Gate deployment with dsl.Condition on evaluation metrics exceeding a threshold so a model that regresses quality never reaches an endpoint.

Vertex AI Model Monitoring detects training-serving skew (live requests versus the training baseline) and prediction drift (live requests versus an earlier serving window) for tabular models.

A monitoring job requires a baseline dataset, a target dataset sourced from endpoint logging in BigQuery, a sampling rate, a monitoring interval, and per-feature alert thresholds.

Skew is divergence between training and serving feature distributions and usually indicates a preprocessing or pipeline defect; drift is change in serving distributions over time and indicates shifting real-world inputs.

Endpoint request-response logging must be enabled and exported to BigQuery before model monitoring or ground-truth-based evaluation can run.

Cloud Monitoring alerting policies on request count, p95 latency, error rate, and model-monitoring anomaly counts drive on-call paging for a deployed endpoint.

Explainable AI supplies sampled Shapley or integrated-gradients feature attributions; monitoring attribution drift catches models that start keying on the wrong features.

Bias detection in Vertex AI Model Evaluation compares metric values across privileged and unprivileged groups for tabular models.

Log sampling reduces storage cost and PII exposure but lowers monitoring sensitivity, so keep full logging on a canary slice while investigating drift.