v3.9.0

✨ New Features

• Introduction of MLflow Assistant, an in-product chatbot backed by Claude Code for debugging ML apps and agents.
• Trace Overview Dashboard added to GenAI experiments providing pre-built statistics like latency, request count, and quality metrics.
• Revamped AI Gateway with the server now integrated directly into the tracking server, supporting passthrough endpoints, traffic splits, and fallback models.
• Online Monitoring with LLM Judges allows automatic scoring of traces without custom code, supporting pre-defined or custom judges.
• Judge Builder UI introduced for defining and iterating on custom LLM judge prompts directly within the UI.
• Support for Distributed Tracing, propagating trace context across different services and processes via APIs in `mlflow.tracing.distributed`.
• Introduction of `MemAlignOptimizer`, a new algorithm for judge optimization that learns from past feedback.
• Addition of `GePaAlignmentOptimizer` for judge instruction optimization.
• Addition of `Fluency` scorer for evaluating text quality.
• Addition of `KnowledgeRetention` built-in scorer.
• Integration of Phoenix (Arize) third-party scorer.
• Support for conversational datasets (with persona, goal, context) in evaluation.
• Introduction of conversational guidelines scorer.
• Support for Jinja2 template in prompt registry.
• Support for metaprompting in `mlflow.genai.optimize_prompts()`.
• Addition of `log_stream` API for logging binary streams as artifacts.
• Addition of `import_checkpoints` API for Databricks SGC Checkpointing with MLflow.