v3.6.0rc0
Breaking Changes📦 mlflowView on GitHub →
⚠ 1 breaking✨ 6 features⚡ 3 deprecations🔧 6 symbols
Summary
MLflow 3.6.0rc0 adds full OpenTelemetry support, session‑level trace UI, a new experiment tab layout, Vercel AI TypeScript tracing, cost tracking for LLM judges, and an Agent Server, while deprecating the filesystem backend and removing span name deduplication.
⚠️ Breaking Changes
- [Tracing] Span name deduplication has been removed, which may affect code that relied on automatic deduplication of span names. Update custom tracing logic to handle unique span names manually.
Migration Steps
- Switch tracking store configuration from `filesystem` to `sqlite` in `mlflow.set_tracking_uri` or `mlflow.tracking` settings.
- Run `mlflow db upgrade` to migrate existing tracking data to SQLite if needed.
- Review any custom tracing code that relied on automatic span name deduplication and ensure unique span names are provided.
- Test your pipelines against the new Agent Server if you were using the previous agent deployment mechanism.
✨ New Features
- Full OpenTelemetry Support in OSS Server: comprehensive OpenTelemetry integration with MLflow SDK for unified traces and OTLP span ingestion.
- Session-level View in Trace UI: new chat sessions tab for organizing and analyzing related traces at the session level.
- New experiment tab bar: experiment navigation moved from top to left side for better layout.
- Vercel AI Support in TypeScript Tracing SDK: auto‑tracing support for Vercel AI SDK in TypeScript.
- Tracking Judge Cost and Traces: automatic tracking and rendering of LLM judge evaluation costs and traces.
- Agent Server: new infrastructure for managing and deploying scoring agents with enhanced orchestration.
Affected Symbols
⚡ Deprecations
- [Tracking] Filesystem backend is deprecated in favor of SQLite. Migrate to the SQLite backend for artifact storage.
- [Flavors] Promptflow flavor is deprecated.
- [Flavors] pmdarima and diviner flavors are deprecated.