Enforce Request-Response Logging for GCP Vertex AI Endpoints

GCP Vertex AI request-response logging captures a sample of the prediction requests and responses served by an endpoint to a BigQuery table. Enabling it gives you an audit trail of how deployed models are being used, which is valuable for monitoring model behavior, troubleshooting, abuse detection, and compliance evidence.

This policy pack can help you configure the following settings for GCP Vertex AI endpoints:

• Check that request-response logging is enabled

Review policy settings →

Getting Started

Requirements

• Terraform
• Guardrails mods:
  • @turbot/gcp-vertexai

Credentials

To create a policy pack through Terraform:

• Ensure you have Turbot/Admin permissions (or higher) in Guardrails
• Create access keys in Guardrails

And then set your credentials:

export TURBOT_WORKSPACE=myworkspace.acme.com
export TURBOT_ACCESS_KEY=acce6ac5-access-key-here
export TURBOT_SECRET_KEY=a8af61ec-secret-key-here

Please see Turbot Guardrails Provider authentication for additional authentication methods.

Usage

Install Policy Pack

Note

By default, installed policy packs are not attached to any resources.

Policy packs must be attached to resources in order for their policy settings to take effect.

Clone:

git clone https://github.com/turbot/guardrails-samples.git
cd guardrails-samples/policy_packs/gcp/vertexai/enforce_request_response_logging_for_endpoints

Run the Terraform to create the policy pack in your workspace:

terraform init
terraform plan

Then apply the changes:

terraform apply

Apply Policy Pack

Log into your Guardrails workspace and attach the policy pack to a resource.

If this policy pack is attached to a Guardrails folder, its policies will be applied to all accounts and resources in that folder. The policy pack can also be attached to multiple resources.

For more information, please see Policy Packs.

Enable Enforcement

Tip

You can also update the policy settings in this policy pack directly in the Guardrails console.

Please note your Terraform state file will then become out of sync and the policy settings should then only be managed in the console.

This control checks that request-response logging is enabled and raises an alarm when it is not. To disable the check entirely, you can set the policy to Skip:

resource "turbot_policy_setting" "gcp_vertexai_endpoint_active_request_response_logging" {
  resource = turbot_policy_pack.main.id
  type     = "tmod:@turbot/gcp-vertexai#/policy/types/endpointActiveRequestResponseLogging"
  value    = "Skip"
}

Then re-apply the changes:

terraform plan
terraform apply