What changed
AWS has introduced a centralized observability data lake for Amazon Bedrock, allowing users to track model invocations across multiple accounts. This solution aggregates data from linked accounts into a central location, providing detailed insights into each invocation.
What this means in practice
The observability data lake pulls raw model invocation logs from linked accounts into Amazon CloudWatch. Subscription log filters then forward these records to a central account, where they are processed through Amazon Data Firehose and exposed via Amazon Athena as a queryable table. This enables users to see prompts, model responses, token counts, and other details for each invocation. The data lake is designed to help platform and machine learning engineers debug spikes, identify wasteful prompts, and track usage across teams.
Limitations and availability
The data lake is available through Infrastructure as Code using AWS CloudFormation templates, which can be deployed across all accounts in an organization. Users need AWS CLI configured with credentials for each account, CloudFormation permissions, and a cloned repository with all resources. Deployment involves several stages, including setting up the central data lake stack, deploying logging stacks in source accounts, enabling model invocation logging, verifying data arrival, and creating an Athena view for querying. The solution is currently in preview and requires users to follow the outlined deployment steps.
