mlflow

MLflow provides dual-API experiment tracking through a fluent interface (mlflow.log_param, mlflow.log_metric) and a client-based API (MlflowClient) that both persist to pluggable storage backends (file system, SQL databases, cloud storage). The tracking system uses a hierarchical run context model where experiments contain runs, and runs store parameters, metrics, artifacts, and tags with automatic timestamp tracking and run lifecycle management (active, finished, deleted states).

MLflow's Model Registry provides a centralized catalog for registered models with version control, stage management (Staging, Production, Archived), and metadata tracking. Models are registered from logged artifacts via the fluent API (mlflow.register_model) or client API, with each version immutably linked to a run artifact. The registry supports stage transitions with optional descriptions and user annotations, enabling governance workflows where models progress through validation stages before production deployment.

MLflow provides a REST API server (mlflow.server) that exposes tracking, model registry, and gateway functionality over HTTP, enabling remote access from different machines and languages. The server implements REST handlers for all MLflow operations (log metrics, register models, search runs) and supports authentication via HTTP headers or Databricks tokens. The server can be deployed standalone or integrated with Databricks workspaces.

MLflow provides native LangChain integration through MlflowLangchainTracer that automatically instruments LangChain chains and agents, capturing execution traces with inputs, outputs, and latency for each step. The integration also enables dynamic prompt loading from MLflow's Prompt Registry and automatic logging of LangChain runs to MLflow experiments. The tracer uses LangChain's callback system to intercept chain execution without modifying application code.

MLflow's environment packaging system captures Python dependencies (via conda or pip) and serializes them with models, ensuring reproducible inference across different machines and environments. The system uses conda.yaml or requirements.txt files to specify exact package versions and can automatically infer dependencies from the training environment. PyFunc models include environment specifications that are activated at inference time, guaranteeing consistent behavior.

MLflow integrates with Databricks workspaces to provide multi-tenant experiment and model management, where experiments and models are scoped to workspace users and can be shared with teams. The integration uses Databricks authentication and authorization to control access, and stores artifacts in Databricks Unity Catalog for governance. Workspace management enables role-based access control (RBAC) and audit logging for compliance.

MLflow's Prompt Registry enables version-controlled storage and retrieval of LLM prompts with metadata tracking, similar to model versioning. Prompts are registered with templates, variables, and provider-specific configurations (OpenAI, Anthropic, etc.), and versions are immutably linked to registry entries. The system supports prompt caching, variable substitution, and integration with LangChain for dynamic prompt loading during inference.

MLflow's evaluation framework provides a unified interface for assessing LLM and GenAI model quality through built-in metrics (ROUGE, BLEU, token-level accuracy) and LLM-as-judge evaluation using external models (GPT-4, Claude) as evaluators. The system uses a metric plugin architecture where custom metrics implement a standard interface, and evaluation results are logged as artifacts with detailed per-sample scores and aggregated statistics. GenAI metrics support multi-turn conversations and structured output evaluation.

MLflow's tracing system captures execution traces of LLM applications (chains, agents, tool calls) with OpenTelemetry integration, recording spans for each operation (LLM calls, tool invocations, retrieval steps) with inputs, outputs, latency, and error information. The MlflowLangchainTracer automatically instruments LangChain applications, and traces are stored in a dedicated trace backend with UI visualization for debugging and issue detection. Traces link to experiment runs for end-to-end observability.

MLflow's PyFunc system provides a unified model interface (mlflow.pyfunc.PythonModel) that abstracts framework-specific details, allowing models from any framework (scikit-learn, TensorFlow, PyTorch, custom Python code) to be loaded and invoked through a consistent API. Models are serialized with their environment (dependencies, Python version) using conda or virtualenv, and PyFunc handles environment activation and model loading at inference time. The system supports batch prediction and Spark UDF integration for distributed inference.

MLflow's autologging system automatically captures training metrics, hyperparameters, and model artifacts without explicit logging code by instrumenting ML frameworks (scikit-learn, TensorFlow, PyTorch, XGBoost, LightGBM, Transformers). The system uses framework-specific hooks (e.g., callbacks for TensorFlow, patching for scikit-learn) to intercept training events and log relevant data. Autologging is enabled per-framework via mlflow.autolog() and can be customized with include/exclude filters.

MLflow's artifact system provides a pluggable storage abstraction (ArtifactRepository) supporting local filesystem, cloud storage (S3, Azure Blob Storage, GCS), and Databricks Unity Catalog. Artifacts are stored with run-based paths (runs:// URIs) and can be logged as files, directories, or models. The system handles authentication, credential management, and cloud-specific optimizations (multipart uploads, parallel transfers) transparently. Artifact repositories are configured via tracking URI or environment variables.

MLflow's Gateway system provides a unified API for accessing multiple LLM providers (OpenAI, Anthropic, Cohere, etc.) through a single endpoint, abstracting provider-specific APIs and handling authentication. The gateway uses a configuration file (gateway.yaml) to define routes for different models and providers, with built-in secret management for API keys and credentials. The system supports request routing, rate limiting, and cost tracking across providers.

MLflow provides a search API (mlflow.search_runs, mlflow.search_experiments) that enables querying experiments and runs using a filter syntax supporting parameter ranges, metric thresholds, tags, and timestamps. The search system uses a backend-specific query engine (SQL for database backends, in-memory filtering for file-based backends) and returns results as pandas DataFrames. Advanced filtering supports logical operators (AND, OR) and comparison operators (>, <, ==, !=).