Study guide
4.4 Deploying Models to Public and Private Endpoints
Google Cloud Professional Machine Learning Engineer Exam
Track your progress as you read
Finished sections get a check mark.
1 Introduction & Exam Overview
0.1 Exam Overview, Blueprint & Certification Logistics 0.2 Study Strategy, Scenario Mechanics & 8-Week Roadmap
2 Section 1: Architecting Low-Code AI Solutions (~13%)
1.1 BigQuery ML: Model Types, SQL Syntax and In-Database Evaluation 1.2 Feature Engineering and Feature Selection in BigQuery ML 1.3 Generating Predictions with BigQuery ML 1.4 AutoML on Agent Platform: Tabular, Image, Text and Video Training 1.5 Fine-Tuning Gemini Models from BigQuery 1.6 Model Garden: Evaluating and Selecting the Right Model 1.7 Industry-Specific AI APIs: Document AI, Vision, Translate and Speech 1.8 Generative Media and Models as a Service: Gemini, Imagen and Veo 1.9 Optimizing Gemini Applications for Cost, Latency and Availability
3 Section 2: Collaborating Within and Across Teams to Manage Data and Models (~16%)
2.1 Organizing and Exploring Tabular, Text, Image and Video Data 2.2 Choosing the Preprocessing Tool: BigQuery, Dataflow, Spark and In-Memory Python 2.3 Creating and Consolidating Features in Agent Platform Feature Store 2.4 Data Privacy, PII Handling and Sensitive Information 2.5 Notebook Environments: Collaboration and Security Best Practices 2.6 Prototyping in Workbench and Colab Enterprise with PyTorch, scikit-learn and JAX 2.7 Prototyping with Model Garden Foundational and Open Models in Notebooks 2.8 Choosing the Development and Experimentation Environment 2.9 Evaluating Generative Solutions: LLM-as-a-Judge and RAG Evaluation 2.10 Experiment Tracking, Artifact Lineage and Model Registry Governance
4 Section 3: Scaling Prototypes into ML Models (~21%)
3.1 Choosing the Model Type: ARIMA, DNN, LLM and Classical Baselines 3.2 Choosing the Product: AutoML, BigQuery ML, Custom Training and Pipelines 3.3 Choosing the Deployment Strategy at Design Time 3.4 Modeling Techniques for Interpretability Requirements 3.5 Organizing Training Data on Cloud Storage and BigQuery 3.6 Ingesting Structured and Unstructured Data into Training Pipelines 3.7 Training SDKs and Services: Custom Training, Kubeflow on GKE, AutoML and Tabular Workflows 3.8 Troubleshooting ML Model Training Failures 3.9 Hyperparameter Tuning, Bayesian Optimization and Vizier 3.10 Model Evaluation Metrics, Transfer Learning and Fine-Tuning 3.11 Evaluating Compute and Accelerator Options: CPU, GPU and TPU 3.12 Distributed Training Strategies, Hardware Acceleration and Performance Optimization
5 Section 4: Serving and Scaling Models (~20%)
4.1 Batch and Online Inference: Agent Platform, Model Garden, Cloud Run and GKE 4.2 Rollout Strategies: Canary, Traffic Splitting, Shadow and Blue-Green 4.3 Tuning Models for Production: Vector Search and Model Compression 4.4 Deploying Models to Public and Private Endpoints 4.5 Choosing Serving Hardware: CPU, GPU, TPU and Edge 4.6 Scaling the Serving Backend for Throughput
6 Section 5: Automating and Orchestrating ML Pipelines (~18%)
5.1 Validating Data and Models Inside Pipelines 5.2 Agent Platform Pipelines: Kubeflow Pipelines (KFP) and TFX Architecture 5.3 Reusable Pipeline Components, Artifacts and ML Metadata Lineage 5.4 Ray on Agent Platform for Distributed Python Workloads 5.5 MLOps CI/CD Automation, Retraining Triggers and Skew Prevention
7 Section 6: Monitoring AI Solutions (~13%)
6.1 Securing AI Systems: Exfiltration, Malicious Prompting and Model Armor 6.2 Responsible AI, Explainable AI, Fairness Metrics and Model Governance 6.3 Model Monitoring: Feature Skew and Drift Detection 6.4 Operational Troubleshooting, Performance Profiling and Cost Optimization 6.5 Monitoring, Testing and Evaluating Generative AI in Production
4.4 Deploying Models to Public and Private Endpoints
Key Takeaways Explain This Topic AI
A public endpoint is reachable over the internet at a Google-managed address and is still authenticated and authorized by IAM; "public" describes the network path, not the access control.
Private Service Connect endpoints expose the model through a private IP inside the consumer VPC, so traffic never traverses the internet.
VPC Service Controls add an exfiltration perimeter that IAM alone cannot provide, blocking data movement to projects outside the perimeter even for authorized principals.
Private endpoints require DNS and connectivity design — private DNS zones, and Cloud VPN or Interconnect for on-premises callers.
Regional endpoints must match the data residency requirement; a private path does not by itself satisfy a residency obligation.
Last updated: September 2026
- 4.4 Deploying Models to Public and Private Endpoints
- Blueprint reference: Section 4.2, "Deploying models to public and private endpoints."
The word "public" causes more confusion than any other term in this area, so start there.
Public Does Not Mean Unauthenticated
A public endpoint is reachable at a Google-managed regional address over the internet. Every request is still authenticated with Google Cloud credentials and authorized by IAM. A caller without the appropriate role gets a 403 regardless of where it sits on the network.
So "public endpoint" is a statement about the network path , not about access control. When a scenario objects that an endpoint is "exposed", the real requirement is almost always one of:
Traffic must not traverse the public internet (a network path requirement).
Data must not be able to leave a defined boundary (an exfiltration requirement).
Callers live on-premises or in a peered VPC and must reach the model privately (a connectivity requirement).
Each maps to a different control, and conflating them is how wrong answers get chosen.
The Three Controls and What Each Actually Solves
- Requirement Control What it does
- Who may call the model IAM Grants or denies the caller's identity
Traffic must stay off the public internet Private Service Connect endpoint Exposes the service through a private IP in the consumer VPC
Data must not leave a boundary, even for authorized users VPC Service Controls Defines a service perimeter; blocks reads and copies across it
On-premises callers must reach a private endpoint Cloud VPN / Cloud Interconnect Provides the network route into the VPC
Name resolution for the private address Cloud DNS private zone Resolves the service name to the private IP
The exam's favourite pairing is IAM versus VPC Service Controls . IAM answers may this identity access this resource . VPC Service Controls answers may this data move to that place . An insider with legitimate read access is exactly the case IAM cannot address and VPC-SC can.
Private Service Connect for Inference
Private Service Connect (PSC) publishes the inference service into the consumer's VPC behind a private IP address. Requests originate inside the VPC, resolve to that private address, and reach the model without egressing to the internet.
Design elements that accompany it:
A reserved internal IP in the consumer VPC for the endpoint.
A private DNS zone so application code can use a hostname rather than a hard-coded IP.
Firewall rules permitting the application subnets to reach the endpoint address.
Routes from on-premises , if applicable, through Cloud VPN or Interconnect, plus DNS forwarding so on-premises resolvers can resolve the private zone.
The most common operational failure is DNS: connectivity is correct, the firewall allows the traffic, and the application still fails because the hostname resolves to the public address. When a scenario says "we configured a private endpoint but calls still leave the VPC," DNS resolution is the first thing to check.
Private Endpoints and Latency
There is a secondary benefit worth knowing: a private path can reduce and stabilize latency by avoiding internet routing variability, particularly for callers already inside Google Cloud. It is not the primary reason to choose PSC, and a scenario that only mentions latency is usually pointing at hardware, autoscaling, or regional placement instead.
Residency Is a Separate Axis
Making a path private does not make it compliant with a data residency rule. Residency is satisfied by deploying the endpoint in a region permitted by the requirement, and by ensuring the model, its logs, and its monitoring data also stay in that region. A private endpoint in the wrong region still violates residency; a public endpoint in the right region may satisfy it while failing a separate network-path requirement.
Related controls that show up in the same scenarios:
CMEK for organization-controlled encryption of the model and its data.
Request-response logging destination — prediction logs land in BigQuery, and that dataset's location is part of the residency story.
Access Transparency and audit logs for evidencing who called what.
- Choosing, in Practice Callers are Google Cloud services in the same org, no network-path restriction
- → public endpoint + IAM (simplest, fully authenticated)
- Security policy forbids inference traffic on the public internet
- → Private Service Connect endpoint + private DNS + firewall rules
- Regulators require that data cannot be copied out of a defined boundary
- → add VPC Service Controls perimeter (with or without PSC)
- Callers are on-premises
- → PSC + Cloud VPN/Interconnect + DNS forwarding
- Data must remain in a specific country
- → choose the region; verify logs and monitoring destinations too
- Exam Traps
Believing a public endpoint is unauthenticated. IAM always applies.
Using IAM to stop exfiltration. That is VPC Service Controls' job.
Assuming a private endpoint satisfies residency. Region does that.
Forgetting DNS when configuring Private Service Connect.
Overlooking log destinations when reasoning about where data lives.