Databricks Unity AI Gateway provides a unified way to monitor and understand AI activity through its built-in Workspace AI Gateway Usage Analytics dashboard and the underlying system.ai_gateway.usage system table.
Some useful capabilities include:
- Model usage and performance: Track invocations, token consumption, latency, streaming responsiveness and errors.
- Spend insights: Understand model consumption and available cost information.
- Request attribution: Use request tags to analyse usage across projects and workloads.
- Prompt caching: Examine cached input-token utilization.
- MCP observability: Monitor tool invocations, service activity and performance.
- Coding-agent observability: Gain visibility into supported coding-agent activity, sessions and model consumption.
One interesting observation: MCP usage records include protocol operations such as initialize and tools/list, not just actual tools/call executions. So counting every MCP event as a tool invocation can overstate the number of tools actually executed.
๐ Turns out, not every MCP event means a tool got to do some actual work. Some were just saying hello (initialize) or checking who's available (tools/list)!
The built-in dashboard provides a convenient starting point, while the system table enables deeper SQL analysis and custom reporting.
I've put together an article covering these capabilities, with dashboard screenshots, SQL examples and practical observations.
๐ Exploring Databricks Unity AI Gateway Observability
Would be great to hear how others are leveraging Unity AI Gateway observability for their AI workloads.
