GCP Professional Machine Learning Engineer Practice

Vertex AI combines training, tuning, pipelines, and prediction services into one managed ML platform. Data warehouses, messaging, and function runners are different services.

Vertex AI Pipelines runs compiled pipeline steps on managed infrastructure so training, evaluation, and deployment can be reproduced. Storage, notebooks, and dashboards are not orchestrators.

Experiments record runs, parameters, and metrics so teams can compare attempts. VMs, deployment targets, and databases do not track experiments.

The Feature Store centralizes feature definitions and online or offline serving. Key management, scheduling, and DNS are unrelated.

Model Garden provides access to Google and partner foundation models plus ready-to-use solutions. Billing, firewalls, and data lakes are not catalogs.

Endpoints host model versions and route online prediction traffic. SQL instances, topics, and functions serve different purposes.

Vertex AI TensorBoard visualizes scalars, images, graphs, and other training data. Load balancers, warehouses, and policies are not visualization tools.

Vector Search indexes embeddings and returns nearest neighbors at scale. Relational databases, job runners, and resolvers do not perform similarity search.

Traffic split supports gradual rollouts by routing a configurable share of requests to a model version. Deleting, compressing, and retraining are separate actions.

Pipelines orchestrate workflows, Experiments track runs, the Registry manages model versions, and endpoints serve predictions.

BigQuery ML trains and evaluates models with SQL so analytics teams can run ML without moving data. Notebooks, key stores, and load balancers are not SQL ML services.

Dataflow provides unified stream and batch processing with autoscaling. Notebooks, secrets, and load balancers serve other roles.

Pub/Sub ingests and delivers real-time messages so streaming pipelines can react to new data. Models, SQL, and DNS are handled elsewhere.

Cloud Storage holds datasets, checkpoints, and model artifacts that training and prediction jobs read. SQL, messages, and keys use other services.

Dataproc runs managed Spark, Hadoop, Flink, and related clusters. Warehouses, event buses, and registries are different services.

Cloud Composer schedules and orchestrates data and ML workflows with Airflow. Vector databases, firewalls, and key management are not orchestrators.

Slots are units of query compute capacity that reservations allocate to projects and teams. Keys, storage, and roles are separate concepts.

Vertex AI Datasets organize and version data for AutoML and custom training. Topics, registries, and images are not datasets.

Templates package pipeline configuration so teams can launch repeatable jobs. Image formats, frameworks, and DNS are unrelated.

BigQuery runs SQL analytics, Dataflow processes streams and batches, Pub/Sub delivers real-time events, and Cloud Storage stores artifacts.

Custom training jobs execute your containerized training code on managed compute. AutoML, warehouse queries, and registry entries are different.

AutoML handles architecture search, training, and tuning for common data types. Storage, load balancing, and roles are unrelated.

Hyperparameter tuning runs trials with different values to maximize a target metric. Encryption, versioning, and logging are separate capabilities.

Vizier performs black-box optimization over parameters defined by the user. Views, runtimes, and alert channels are not optimizers.

Explainable AI computes attributions such as integrated gradients for predictions. Queues, buckets, and endpoints do not explain predictions.

Vertex AI Forecast builds time-series predictions using historical data. Databases, DNS, and key management are not forecasting services.

Tabular Workflows package automated pipelines that train and evaluate models on tabular data. Spreadsheets, SQL builders, and firewall rules are unrelated.

Custom training containers bundle code and dependencies so Vertex AI can run them consistently. Registry policies, empty images, and queues do not train models.

Custom prediction containers let you control dependencies and runtime for model serving. AutoML models, tables, and alerts are different resources.

Custom jobs run your code, AutoML automates training, tuning searches parameters, and Explainable AI provides prediction explanations.

minReplicaCount keeps a baseline number of replicas warm so latency stays low. Network, epochs, and encryption are separate settings.

maxReplicaCount caps autoscaling so cost stays controlled during traffic spikes. Batch size, accuracy, and log storage are different.

GPU quota limits accelerator usage per region and project. Encryption, key rotation, and billing are separate controls.

TPU quota caps tensor processing unit capacity for training jobs. Version labels, retention, and DNS are unrelated.

Batch prediction runs offline jobs over large inputs without per-request latency guarantees. Real-time serving, resharding, and compression are different.

Online prediction responds to individual requests quickly. Training, exporting, and backups are not real-time serving.

Slot autoscaling automatically adds query capacity when demand rises, up to a configured maximum. Buckets, keys, and functions are separate.

Autoscaling adjusts endpoint replicas in response to traffic while respecting min and max limits. Retraining, formats, and deletion are different.

Streaming Engine decouples stream job compute from state, allowing efficient scaling and updates. SQL optimization, serving, and firewalls are unrelated.

minReplicaCount keeps warm standby, maxReplicaCount sets the ceiling, batch prediction handles bulk inference, and Streaming Engine scales stream processing.

VPC Service Controls applies perimeters to Google Cloud services so data cannot move outside allowed contexts. GPUs, registries, and warehouses are not perimeters.

CMEK gives customers control over the keys protecting data in supported Google Cloud services. Passwords, firewall rules, and aliases are different.

Cloud KMS manages cryptographic keys and their lifecycle. Pipelines, networks, and billing are unrelated.

Confidential VMs use AMD SEV to encrypt memory in use. Private IPs, batch use, and perimeters address different concerns.

Cloud DLP inspects data for sensitive content and applies redaction or masking before use in ML. Scaling, storage, and SQL are separate.

IAM principals are identities such as users, groups, service accounts, and workloads. Keys, storage classes, and DNS records are not identities.

Custom service accounts let workloads use only the permissions required. User accounts, firewalls, and versions are not workload identities.

Private Service Connect lets clients reach Vertex AI through private IP addresses inside their VPC. Encryption, deletion, and projects are unrelated.

Vertex AI Workbench offers managed notebooks with preinstalled ML tools. Queues, routers, and key management are not notebooks.

VPC Service Controls builds perimeters, CMEK gives key control, Cloud DLP de-identifies data, and IAM roles control access.

Model Monitoring watches input distributions and compares them to training data to detect skew and drift. Billing, routes, and vulnerabilities are separate.

Training-serving skew occurs when prediction inputs differ in distribution or processing from training data. Scheduling, policies, and regions are different issues.

Data drift means the live data distribution has moved away from what the model was trained on. Corruption, quotas, and networks are unrelated.

The Model Registry tracks model versions, aliases, and deployment state. Labeling, GPUs, and logging are different resources.

Aliases let deployments follow a stable name while the underlying version changes. Keys, accounts, and schemas are not model aliases.

Continuous training triggers retraining on a schedule or on new data so models stay current. Manual runs, tuning, and copies are not continuous.

Cloud Logging centralizes audit and operational logs from ML resources. Training, routing, and scaling are handled by other services.

Together they capture experiment runs and lineage so teams can reproduce and audit ML workflows. Network, billing, and login data are outside their scope.

Caching reuses outputs for unchanged steps, saving time and cost. Prediction caching, encryption, and VM sizing are different.

Model Monitoring detects skew and drift, the Registry manages versions, caching skips unchanged steps, and Cloud Logging provides an audit trail.