GCP ML Engineer (Google Cloud Certifications) key terms

Google Cloud's unified ML platform that brings dataset management, training, model registry, deployment, pipelines, and monitoring under a single API and console, replacing the earlier separate AI Platform services.

A Vertex AI capability that trains models on your own labeled data using Google-managed architectures and hyperparameter search, so no training code is needed. You supply the data and a budget; Vertex AI handles the rest.

A console and API workspace for writing generative AI prompts, comparing model responses, and tuning and evaluating models on your own examples before integrating them into an application.

Google's family of multimodal generative models served through Vertex AI. They accept text, images, audio, and video, and are callable through the Vertex AI API with enterprise controls such as IAM and VPC Service Controls.

A BigQuery feature for creating, training, evaluating, and predicting with models using SQL. Data stays in place, and supported models range from logistic regression to imported or remotely hosted foundation models.

Pre-trained and customizable document parsers that extract text, tables, and structured fields from scans such as invoices, receipts, and forms, returning typed entities rather than raw OCR text alone.

Ready-to-call Google Cloud services such as Vision, Natural Language, Speech-to-Text, and Translation that return predictions for common tasks with no training, suitable when a general-purpose model already meets requirements.

A low-code environment for assembling search and conversational agents over your own documents, combining retrieval, grounding, and Gemini models behind a managed serving API.

Grounding supplies retrieved source content to a model at inference time so answers reflect your data. Retrieval-augmented generation implements this by embedding documents, searching an index, and inserting top results into the prompt.

Adapting a foundation model to a narrow task by training on labeled input-output examples, updating all weights or only a small adapter, which improves consistency and can lower cost per request.

A central catalog of model versions with metadata, lineage, and evaluation results. From it, models can be reviewed, compared, deployed to endpoints, or promoted between environments.

A platform resource that records where training data lives in Cloud Storage or BigQuery along with its schema, enabling splitting, labeling, and statistics without copying the data.

Governance services that discover, catalog, classify, and tag data assets across BigQuery, Cloud Storage, and other sources, giving teams a searchable inventory with lineage and access context.

A non-human identity with scoped permissions that training jobs, pipelines, and endpoints use to read data and call services, keeping access least-privilege and auditable.

A managed JupyterLab environment preintegrated with Google Cloud, used for data exploration, notebook prototyping, and handing work off to training jobs and pipelines.

A structured document describing a model's intended use, training data, evaluation metrics, and known limitations, published with the model version to inform reviewers and downstream consumers.

A metadata service that records the artifacts, executions, and parameters of pipeline runs, producing a queryable lineage graph linking datasets, models, and evaluations.

A catalog of foundation models, open models, and task-specific models that can be used as-is, tuned, or deployed, providing a starting point instead of training from scratch.

Running your own training code on Vertex AI by specifying a container image, machine type, and data location; the service provisions workers, retries failures, and writes outputs to Cloud Storage.

Splitting training across many workers using data or model parallelism. Vertex AI's reduction server aggregates gradients efficiently so synchronous multi-worker jobs scale without network bottlenecks.

Automated search over ranges of training settings such as learning rate, batch size, or tree depth, using Bayesian optimization to find strong configurations within a trial and time budget.

A tracking service that records runs with their parameters, metrics, and artifacts, letting teams compare trials and reproduce the configuration behind a chosen model.

Specialized hardware attached to training or prediction nodes. GPUs suit many deep learning workloads, while TPUs are optimized for the large matrix operations in TensorFlow, JAX, and PyTorch.

A managed repository that serves feature values to both training and online prediction at low latency, with point-in-time lookups that prevent label leakage.

A deployed, addressable resource that hosts one or more model versions behind an HTTP or gRPC API and routes requests to them for online prediction.

Adjusting the number of serving replicas between configured minimum and maximum values based on traffic or utilization, balancing latency under load against cost when idle.

An asynchronous job that scores a large stored dataset and writes results to a destination, trading immediate responses for higher throughput and lower cost per prediction.

Sending specified percentages of prediction requests to different model versions on the same endpoint, enabling canary releases, A/B tests, and gradual rollouts.

A Vertex AI capability that reports how much each input feature contributed to a prediction, using methods such as Sampled Shapley or integrated gradients for tabular, image, and text models.

Deployment and network settings that keep prediction traffic on private IP addresses or within a service perimeter, so requests and data never traverse the public internet.

A managed service that runs containerized pipeline steps defined with KFP or TFX, tracks their artifacts and metadata, and supports scheduled or triggered executions.

An open-source SDK and specification for defining ML workflows as directed acyclic graphs of containerized components; Vertex AI Pipelines compiles and executes those definitions.

A managed Apache Airflow service for orchestrating workflows that span systems, commonly scheduling Vertex AI pipelines alongside BigQuery jobs, data transfers, and other Google Cloud tasks.

A service for storing and versioning container images and language packages, used as the source of component images, training containers, and prediction containers.

Event-driven building blocks: Eventarc routes events from Google Cloud sources to triggers, which invoke Cloud Functions or Cloud Run services to react to new data or pipeline completion.

A service that watches deployed models for training-serving skew and prediction drift, comparing feature and prediction distributions against a baseline and emitting alerts when thresholds are crossed.

Skew is a mismatch between the feature distribution seen in training and the one seen at serving; drift is change in live data over time. Both erode accuracy and signal retraining.

Google Cloud's observability services. They collect endpoint metrics, errors, and logs and support dashboards, log-based metrics, and alerting policies for ML services.

Automated computation of quality metrics such as precision, recall, and RMSE on held-out data, available for AutoML and custom models and stored alongside the model version.