All releases
MLflowv2.17
MLflow 2.17 — Native LLM Tracing and Prompt Registry
MLflow 2.17 makes LLM observability a first-class feature: native tracing for OpenAI, Anthropic, LangChain, and LlamaIndex with zero-code instrumentation, plus a new Prompt Registry for versioning and promoting prompts alongside models.
What's New
- 1Native LLM tracing — automatic span capture for OpenAI, Anthropic, LangChain, LlamaIndex, and Bedrock with a single mlflow.tracing.enable() call
- 2Prompt Registry — version, tag, and promote prompts through dev → staging → production lifecycle, linked to model versions
- 3MLflow Deployments Server supports OpenAI-compatible routing — proxy any model behind a standard API for cost control and fallback
- 4Evaluate improvements — new LLM judge metrics (faithfulness, toxicity, answer relevance) backed by Claude or GPT-4 as the evaluator
- 5Gateway rate limiting and cost tracking per team or project
- 6Unity Catalog integration (Databricks) — models, datasets, and prompts discoverable in the same catalog
- 7GPU profiling in autologging — VRAM usage, utilization captured automatically during training runs
Breaking Changes
- mlflow.gateway module renamed to mlflow.deployments — update all imports
- Python 3.8 support dropped — requires Python 3.9+
Upgrade Notes
The gateway rename is a one-line change per import. LLM tracing is opt-in and additive — existing experiment tracking code is unaffected. Prompt Registry requires MLflow Tracking Server ≥ 2.17; local file-based backends are not supported.