GCP GenAI Engineer key terms

A managed service that runs ML workflows defined as directed acyclic graphs, where each step is a containerized component and artifacts pass between steps through documented inputs and outputs.

An open-source platform for building and running portable ML workflows on Kubernetes, using Python-defined components that Vertex AI Pipelines can compile and execute.

A self-contained, reusable step in an ML workflow with a declared interface of typed inputs and outputs, packaged so it can run in isolation and be shared across pipelines.

The recorded graph of which executions produced which artifacts and from which inputs, enabling reproducibility, debugging, and impact analysis when data or code changes.

A managed metadata store that records artifacts, executions, and contexts from ML workflows so teams can trace model provenance and compare runs across experiments.

A tracking service that logs parameters, metrics, and artifacts for training runs so candidates can compare trials and choose the best configuration systematically.

An automated search over training configuration values such as learning rate or tree depth, guided by a chosen metric, to find settings that improve model quality without manual trial and error.

The defined ranges or discrete candidate values for each hyperparameter that a tuning job may explore, including scaling choices such as linear or logarithmic sampling.

A Vertex AI training run that executes your own packaged application in a managed container on selected machine types, with optional distributed workers and hyperparameter tuning.

A container image bundling your training code and dependencies that Vertex AI runs on managed compute; it must expose an entrypoint that reads arguments and writes outputs to expected paths.

A Google-maintained image with common ML frameworks already installed, usable for training or prediction so teams avoid building and patching their own base images.

A user-built image for serving predictions when the prebuilt runtimes do not meet dependency, framework, or preprocessing needs; it must implement the required HTTP contract.

The HTTP process inside a serving container that loads the model and answers health, predict, and explain requests according to the platform's request and response contract.

Synchronous inference against a deployed endpoint that returns results within the request, suited to interactive use where low latency matters more than throughput cost.

Asynchronous inference over a large stored dataset that writes predictions to a destination such as BigQuery or Cloud Storage, trading latency for throughput and cost efficiency.

A deployed, addressable resource that routes prediction traffic to one or more model versions on allocated compute, with configurable traffic splitting and autoscaling.

An immutable snapshot of a trained model registered under a model resource, allowing several versions to coexist and be deployed or rolled back independently.

A percentage allocation of endpoint requests across deployed model versions, used for canary releases, A/B tests, and gradual rollouts with quick rollback.

A central catalog of models and their versions with metadata such as training source and evaluation results, supporting governance and deployment workflows.

Dynamic adjustment of serving replicas based on load or utilization so an endpoint meets latency targets during peaks while releasing capacity when demand falls.

A regional or project-level limit on resources such as accelerators, endpoints, or requests per minute that must be planned for before scaling a deployment.

A capability to create, train, and evaluate models using SQL directly on BigQuery data, useful for quick baselines and keeping training close to large datasets.

A managed repository for feature values that serves consistent features to training and online inference, reducing skew and duplicated preprocessing logic across teams.

A mismatch between how features are computed during training and during inference, often from duplicated code paths, which degrades live model quality silently.

A change in input feature distributions over time relative to training data, which can erode model accuracy even when the underlying relationship remains stable.

A change over time in the relationship between inputs and the target, so the same features now imply different outcomes and retraining on recent data becomes necessary.

Continuous tracking of a deployed model's inputs, outputs, and performance signals to detect drift, skew, or degradation and trigger alerts or retraining.

A managed feature that watches deployed endpoints for training-serving skew and drift against configured baselines and thresholds, emitting alerts when violations occur.

A managed Apache Beam runner for batch and streaming pipelines, commonly used to transform raw data into features and to preprocess inputs for training or prediction.

An open-source programming model for unified batch and streaming pipelines, where the same transform code runs on different runners such as Dataflow.

A packaged, parameterized pipeline that can be launched without writing code, useful for recurring ingestion and preprocessing jobs triggered by orchestration.

Grouping unbounded streaming data into finite chunks by time or count, such as fixed or sliding windows, so aggregations can be computed on continuous input.

A managed Apache Airflow service for authoring, scheduling, and monitoring workflows that coordinate services such as Dataflow, BigQuery, and Vertex AI training jobs.

A managed cron service that triggers HTTP, Pub/Sub, or App Engine targets on a schedule, often used to start pipelines or batch jobs periodically.

A managed asynchronous messaging service with topics and subscriptions that decouples producers from consumers, commonly feeding streaming pipelines and event-driven retraining.

The optimization service behind Vertex AI hyperparameter tuning that searches parameter spaces efficiently, including black-box optimization for custom objective functions.

Vertex AI features that attribute prediction outputs to input features, such as Sampled Shapley or integrated gradients, helping validate model behavior and satisfy review requirements.