v3.7.0rc0
Breaking Changes📦 mlflowView on GitHub →
⚠ 2 breaking✨ 4 features🔧 3 symbols
Summary
MLflow 3.7.0rc0 introduces trace comparison, multi‑turn evaluation, full‑text UI search, and a Gemini TypeScript SDK, while changing the default backend to SQLite and removing the deprecated `diviner` and `promptflow` flavors.
⚠️ Breaking Changes
- SQLite is now the default backend instead of file‑based storage. Existing setups that relied on the file store may see data written to a new SQLite database. To keep using file‑based storage, set the backend store URI to a file path (e.g., `mlflow server --backend-store-uri file:/path/to/mlruns`) or migrate existing data to SQLite.
- The `diviner` and `promptflow` flavors have been removed. Code that references these flavors will raise errors. Migrate to supported alternatives (e.g., use other MLflow model flavors) to restore functionality.
Migration Steps
- If you need to continue using file‑based storage, explicitly set `--backend-store-uri` to a file path or configure the `MLFLOW_BACKEND_STORE_URI` environment variable.
- Replace any usage of the removed `diviner` or `promptflow` flavors with supported alternatives (e.g., use other MLflow model flavor APIs).
- To disable the new telemetry installation ID, set `MLFLOW_DISABLE_TELEMETRY=true` or `DO_NOT_TRACK=true` in your environment.
✨ New Features
- Trace Comparison UI: side‑by‑side comparison of traces for debugging LLM applications.
- Multi‑turn conversation support in `mlflow.genai.evaluate` for evaluating conversational AI.
- Full Text Trace Search from the UI, allowing keyword search across inputs, outputs, and span details.
- Gemini TypeScript SDK with auto‑tracing support for Gemini models in JavaScript/TypeScript.