Databricks pruned Model Serving API configurations as MLflow expanded tracing and judge integration.

Databricks pushed breaking changes to its client libraries throughout September, cleaning up configuration schemas across Model Serving endpoints. SDK release v0.135.0 removed routing and rate limiting configuration fields from the model serving APIs [7]. Later in the month, release v0.178.0 removed further fields from model serving and catalog configurations [5]. Teams maintaining automated infrastructure or CI/CD pipelines had to update their deployment scripts to avoid errors against these stripped parameters [5, 7].

Model observability and evaluation capabilities matured across two MLflow releases. MLflow 3.16.0 introduced Unity Catalog model service support for built-in evaluation judges, alongside per-user budget policies within the AI Gateway [6]. MLflow 3.16.1 followed with an operational fix for trace location resolution specifically within Databricks Model Serving, while also adding span links in Unity Catalog traces to improve diagnostic visibility across distributed pipelines [4].

Practitioners also saw updates around inference workflows and multi-region deployment strategies. Guides showed how to deploy specialized open weights such as SemIf-OpenJev on managed Model Serving using serverless GPUs, allowing teams to query custom endpoints directly from SQL and Lakeflow jobs via ai_query [1]. Concurrently, community members shared architectures for coordinating multi-region model serving on Databricks using OpenSharing to distribute workloads across regional boundaries [3].