GCP ML Engineer (Google Cloud Certifications) flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

AutoML versus pretrained API

Use a pretrained API such as Vision or Natural Language when the task is generic and no labeled data exists; use AutoML when labeled domain data exists but ML code does not.

Vertex AI Agent Builder

A low-code Google Cloud product for building search and conversational agents by grounding Gemini models on data stores indexed from websites, documents, or BigQuery tables.

Document AI custom extractor

A Document AI processor trained on your own labeled documents to pull custom fields; it typically needs ten or more labeled examples per field and returns structured entities.

BigQuery ML

A BigQuery capability that trains models with a SQL CREATE MODEL statement and scores with ML.PREDICT, so analysts can build and use models without exporting data.

Vertex AI Studio

A console and API for prompt design, model evaluation, and Gemini tuning, including supervised fine-tuning and adapter tuning, without managing any training infrastructure.

Dataplex zones

Within a Dataplex lake, zones group data by processing stage: a raw zone holds unprocessed assets and a curated zone holds cleansed, governed data that other teams can discover.

Analytics Hub

A Google Cloud service that publishes BigQuery datasets as listings which subscribers in other projects or organizations can query in place without copying the underlying data.

Model Registry alias

A mutable name, such as the alias default, that points to one model version in Vertex AI Model Registry so serving code resolves a version without hardcoded resource IDs.

Vertex ML Metadata

A managed metadata store that automatically records artifacts, executions, and lineage for Vertex AI Pipelines runs and custom jobs, enabling reproducibility and cross-team traceability of models.

BigQuery policy tags

Taxonomy tags from Data Catalog applied to BigQuery columns; a user needs the fine-grained reader role on the tag to read that column, giving column-level access control.

Vizier hyperparameter tuning

The Vertex AI tuning service where you define the metric, goal, search space, max trial count, and parallel trial count, and it proposes trials using Bayesian optimization.

Reduction server

A Vertex AI training option that offloads all-reduce gradient aggregation to dedicated workers, speeding up large multi-node distributed training jobs that use many GPUs.

Spot VM training

Preemptible-capacity VMs used for Vertex AI custom training at a steep discount; because jobs can be reclaimed, write checkpoints to Cloud Storage so training resumes.

Custom training container

A Docker image you supply to Vertex AI training that packages your training code and dependencies, must run the training application, and must exit when training finishes.

Time-based data split

Splitting temporal data so training uses earlier periods and evaluation uses later periods, which prevents leakage and mirrors how the model will be used in production.

Endpoint autoscaling target

The utilization level, such as 60 percent CPU, that Vertex AI compares against to add or remove replicas between an endpoint's minimum and maximum replica counts.

Traffic split

Assigning percentage routes to multiple deployed models on one Vertex AI endpoint, which enables canary releases and A/B tests without standing up separate endpoints.

Batch prediction

A Vertex AI job that scores a large dataset read from BigQuery or Cloud Storage and writes results back to storage, requiring no endpoint or online serving path.

Private Service Connect endpoint

A Vertex AI endpoint configuration that serves prediction traffic over a private IP inside your VPC, avoiding the public internet and working within VPC Service Controls perimeters.

Custom container routes

A Vertex AI custom prediction container must serve HTTP on 0.0.0.0 at the port Vertex AI assigns, expose health and predict routes, and return JSON responses.

Minimum replicas

The floor on deployed replicas for an endpoint; setting it above zero keeps a warm model in memory and reduces cold-start latency at the cost of idle capacity.

Pipeline caching

Vertex AI Pipelines reuses the output of a component whose code and input artifact fingerprints are unchanged when caching is enabled, skipping redundant component executions.

Cloud Composer

A managed Apache Airflow service used to orchestrate multi-service workflows, including Vertex AI, BigQuery, Dataproc, and Cloud Storage tasks, on a defined schedule.

Artifact Registry

The Google Cloud repository for container images and language packages; Vertex AI training and pipeline jobs pull custom images from it and Cloud Build pushes to it.

Eventarc trigger

An event-driven trigger, for example on a Cloud Storage object-finalize event, that invokes a service or workflow such as a Vertex AI pipeline run.

Recurring pipeline runs

Vertex AI Pipelines schedules that submit a pipeline job on a cron cadence with supplied parameters, used for periodic retraining without an external orchestrator.

Cloud Build CI

A serverless build service that compiles pipeline templates, runs tests, builds container images, and can deploy pipeline specs as part of continuous integration.

Training-serving skew

A model monitoring signal comparing feature distributions in live prediction requests against the training data baseline; large divergence suggests preprocessing or data pipeline problems.

Prediction drift

A model monitoring signal measuring how the distribution of live request features changes over time relative to an earlier serving window, flagging a possibly stale model.

Monitoring sampling rate

The fraction of prediction requests that Vertex AI Model Monitoring logs and analyzes; lowering it reduces cost and data exposure but may miss rare drift events.

Explainable AI attribution

Vertex AI feature attributions computed with sampled Shapley or integrated gradients that quantify each feature's contribution to a prediction, useful for debugging and importance drift.

Endpoint request logging

Enabling request-response logging on a Vertex AI endpoint writes prediction payloads to BigQuery, which model monitoring and ground-truth evaluation both require.