Databricks AI Functions usage¶
Databricks contributes system-table usage rows for model-serving and foundation model workloads, plus cluster and Lakeflow context for attribution.
Release state
This guide documents the read-only databricks_ai_functions connector
implemented in the platform connector surface.
Required access¶
Grant read access to these Databricks system tables:
| Table | Purpose |
|---|---|
system.billing.usage |
Read AI billing usage rows. |
system.compute.clusters |
Correlate usage to cluster metadata. |
system.lakeflow.job_run_timeline |
Correlate usage to Lakeflow job runs. |
Do not grant workspace mutation, job mutation, or cluster administration permissions.
Filter system.billing.usage.billing_origin_product to documented values. Model
serving uses MODEL_SERVING; foundation-model fine-tuning uses
FOUNDATION_MODEL_TRAINING. Unknown values remain visible as unknown rather than
being silently dropped.
Databricks documents the billing table and product origins in
model-serving cost monitoring.
“AI Functions” in this guide is the connector package name; the usage rows are
model-serving and foundation-model-training charges from system.billing.usage,
not a claim that every row came from an SQL AI Function.
Setup¶
- Create or select a service principal for Venturi.
- Grant read access to the system tables above.
- Store the credential reference in Venturi.
- The default lookback is 24 hours. Set
VENTURI_DATABRICKS_AI_LOOKBACK_HOURSandVENTURI_DATABRICKS_AI_HIGH_VOLUME_THRESHOLDonly when you need to override the default poll window or high-volume flag. - In Venturi, open Integrations and select Databricks AI Functions and run Test connection.
Verification¶
- Only
MODEL_SERVINGandFOUNDATION_MODEL_TRAININGusage rows survive the connector filter. - Cluster and Lakeflow pipeline context appears when available.
- High-volume AI usage is flagged without blocking ingestion.
- The connector remains read-only and continues operating, skipping upstream read errors without interrupting ingestion.
Rotation and offboarding¶
Rotate the service-principal credential on your Databricks schedule. Removing the connector stops new Databricks AI usage correlation.