GCP ML Engineer (Google Cloud Certifications) key terms

Google Cloud's unified managed platform for building, training, deploying, and monitoring machine learning models, offering both code-based and point-and-click interfaces on shared infrastructure.

A no-code training option on Vertex AI where you supply labeled data and Google selects the model architecture and tunes it automatically, useful for teams without dedicated ML engineers.

A console and API workspace for prototyping generative AI: writing and testing prompts, comparing foundation models, and exporting working code into applications.

A curated catalog in Vertex AI of Google, partner, and open foundation models that can be deployed directly or customized, with sample notebooks and deployment options.

Ready-made Google Cloud services such as Vision, Speech-to-Text, and Translation that return predictions through a simple API call without any training or model hosting.

A BigQuery feature that lets analysts create, train, and evaluate models using SQL statements, keeping data in place and lowering the barrier to producing a first model.

Adapting a foundation model to a specific task by training it further on a curated set of labeled input-output examples, typically through managed tuning jobs rather than full retraining.

A low-code environment for assembling conversational and task agents that combine foundation models, search over your own data, and calls to tools or APIs.

Google Cloud's serverless, columnar data warehouse for analytics at scale, commonly the system of record that feeds training datasets and downstream models.

Object storage in Google Cloud, addressed by buckets and URIs, used to hold datasets, training code, and model artifacts that services read from.

A central catalog that versions and tracks ML models with their artifacts and metadata, acting as the source for deployment and governance decisions across teams.

A metadata management service that discovers, catalogs, and governs data assets across projects, helping teams locate datasets and trace their lineage.

Identity and Access Management defines who can act on which Google Cloud resources through roles and policies, controlling access to data, models, and endpoints.

A managed repository that serves feature values to both training and prediction from a single definition, reducing duplicated feature engineering across teams.

A Vertex AI job that runs your own training code in a container on managed compute, with configurable machine types, accelerators, and output locations.

Google-maintained container images for frameworks such as TensorFlow, PyTorch, and XGBoost, so you can run training code without building or patching images yourself.

Splitting training across multiple machines or accelerators using strategies such as data parallelism or all-reduce, shortening wall-clock time on large datasets.

Automatically running many training trials with different hyperparameter values and selecting the best configuration according to an objective metric you specify.

Specialized hardware attached to training or prediction machines; GPUs suit general deep learning while TPUs are Google-designed chips for very large workloads.

A tracking service that records runs, parameters, metrics, and artifacts so teams can compare model iterations and reproduce the results they choose to promote.

A managed, addressable resource that hosts one or more deployed models and answers prediction requests, with configurable machines, scaling, and traffic rules.

Synchronous request-response serving where the caller waits for a prediction, suited to interactive use cases that demand low latency.

Asynchronous scoring of a large stored dataset in a single job, writing results to a destination such as BigQuery or Cloud Storage without keeping a live endpoint.

Automatically adjusting the number of serving replicas to match incoming traffic, within minimum and maximum bounds, balancing latency against cost.

Sending specified percentages of prediction requests to different model versions behind one endpoint, enabling canary releases and A/B comparisons.

A prediction endpoint reachable only from your VPC network rather than the public internet, used when request data must not traverse public addresses.

A managed service that runs ML workflows defined as containerized steps in a directed acyclic graph, tracking artifacts and supporting scheduled or triggered execution.

A reusable, containerized step with typed inputs and outputs that can be composed into a pipeline definition and shared across teams and workflows.

A managed Apache Airflow service for orchestrating workflows that span multiple systems, often used to move data and trigger ML pipeline runs on a schedule.

A store that records artifacts, executions, and their lineage within ML workflows, so you can trace which data and code produced a given model.

A managed cron service that emits timed triggers such as HTTP calls or Pub/Sub messages, commonly used to start recurring pipeline or batch prediction jobs.

A managed repository for container images and language packages, used to store the custom training and pipeline component images that jobs pull at runtime.

A service that watches a deployed model's inputs and outputs against baselines and alerts when skew, drift, or out-of-range values exceed thresholds you configure.

A mismatch between how features are computed for training and for live requests, which silently degrades prediction quality after deployment.

A change over time in the distribution of live input data or model outputs relative to the training baseline, signaling the model may need retraining.

An explanation technique that estimates how much each input feature contributed to a single prediction, available through Explainable AI on Vertex AI.

Google Cloud's observability service for metrics, dashboards, and alert policies, used to notify teams when serving latency, error rates, or custom ML metrics breach thresholds.