GCP GenAI Engineer flashcards
GCP GenAI Engineer · Build recall with a focused review session.
32 cards. Read the term and the definition here. The deck at the top of this page is optional.
- Vertex AI Pipelines
- A managed service that runs containerized ML workflow DAGs defined with KFP or TFX, tracking artifacts and parameters in ML Metadata for reproducibility.
- Vertex AI Model Registry
- A central catalog for model versions with aliases and metadata, used to promote, compare, and roll back models deployed to endpoints.
- Batch prediction
- An offline Vertex AI job that scores a large dataset from BigQuery or Cloud Storage and writes outputs without keeping an endpoint running.
- Online endpoint
- A deployed Vertex AI endpoint that serves low-latency predictions and autoscales replicas based on traffic and configured min/max counts.
- Traffic split
- An endpoint setting that routes a percentage of prediction requests to different deployed model versions, enabling canary and A/B testing.
- Model monitoring skew
- A Vertex AI check comparing the live serving feature distribution against the training baseline to detect training-serving skew.
- Model monitoring drift
- A Vertex AI check comparing recent live prediction distributions against earlier live data to detect changes in input or output behavior over time.
- Feature Store
- A Vertex AI service that stores feature values by entity ID and timestamp, enabling point-in-time correct feature retrieval for training and serving.
- Point-in-time correctness
- The practice of retrieving feature values as they existed at the prediction timestamp, preventing future information from leaking into training labels.
- Dataflow
- A managed Apache Beam runner on GCP for batch and streaming data processing, commonly used to build preprocessing and feature engineering pipelines.
- BigQuery ML
- A BigQuery capability to train and evaluate models such as linear regression, boosted trees, and AutoML directly with SQL against warehouse data.
- AutoML
- A Vertex AI service that trains custom models on your labeled data without writing training code, selecting architecture and hyperparameters automatically.
- Custom training
- A Vertex AI job that runs your own training container or Python package on managed machines with optional GPUs or TPUs.
- Hyperparameter tuning
- A Vertex AI training job that searches parameter ranges using Bayesian optimization to maximize or minimize a chosen metric.
- Vertex AI Experiments
- A tracking service that records parameters, metrics, and artifacts across training runs so results can be compared and reproduced.
- Model Garden
- A Vertex AI catalog of foundation and open models that can be used via publisher endpoints or deployed to dedicated endpoints for customization.
- Publisher endpoint
- A Google-managed Vertex AI endpoint for foundation models that bills per token or request, requiring no deployment or capacity management.
- Dedicated endpoint
- A Vertex AI endpoint where you deploy a model to your own machine resources, giving control over throughput, scaling, and customization.
- VPC Service Controls
- A GCP security perimeter that restricts access to services such as Vertex AI and BigQuery, helping prevent data exfiltration.
- CMEK
- Customer-managed encryption keys let you control the keys protecting data at rest in GCP services including Vertex AI and Cloud Storage.
- TFRecord
- A binary serialization format for TensorFlow training data that stores records efficiently and avoids CSV parsing overhead at scale.
- Pub/Sub to Dataflow
- A streaming ingestion pattern where Pub/Sub receives events and Dataflow applies windowed transformations before writing to BigQuery or Feature Store.
- Dataplex
- A GCP data governance service that provides discovery, lineage, and quality metadata across BigQuery, Cloud Storage, and other data sources.
- Cloud Scheduler trigger
- A cron service used with Cloud Functions or Pub/Sub to start Vertex AI Pipelines on a schedule, since pipelines do not self-schedule.
- Component caching
- A Vertex AI Pipelines setting that skips a component execution when its inputs and code are unchanged, reducing cost and runtime.
- ML Metadata
- A Vertex AI service that records executions, artifacts, and lineage for pipeline runs, enabling traceability from model back to data.
- Cloud Build for ML
- A CI/CD service used to build and test training and serving container images before a pipeline deploys them to staging or production.
- Autoscaling replicas
- An endpoint setting that adjusts the number of serving replicas between configured minimum and maximum values based on incoming traffic.
- Private endpoint
- A Vertex AI endpoint reachable only through VPC peering from your network, reducing exposure compared with public internet endpoints.
- Prediction container contract
- The routes a custom serving container must expose, including health and predict, so Vertex AI can route requests and check readiness.
- Spot VMs for training
- Discounted, preemptible compute used for fault-tolerant custom training jobs to reduce cost, with checkpointing to survive interruptions.
- Model rollback
- Reassigning a serving alias in Model Registry to a previous model version to restore behavior after a bad deployment.