Feat: Add MLflow Analytics provider

[WeichenXu123]

Summary

Adds MLflow as an analytics/tracing provider in Flowise, allowing users to send LLM traces to an MLflow tracking server. Traces are logged as MLflow spans with proper parent-child hierarchy (CHAIN → LLM/TOOL), token usage, and model metadata.

Uses the official https://www.npmjs.com/package/@mlflow/core TypeScript SDK (^0.2.0).

Motivation

Flowise already supports analytics providers like LangSmith, Langfuse, LunaryAI, and Opik, but lacked support for the popular MLflow tracing provider.

Provider setup

[gemini-code-assist]

Code Review

This pull request integrates MLflow as an analytic provider, adding credential definitions, UI assets, and tracing logic within the AnalyticHandler for chains, LLMs, and tools. The review feedback highlights a critical concurrency issue where the global initialization of MLflow could lead to configuration leaks between requests. Additionally, several bugs were identified where span maps are overwritten rather than merged, potentially causing data loss in complex execution flows, and a recommendation was made to improve type safety for configuration options.

[gemini-code-assist]

The mlflow.init() function configures a global singleton in the MLflow SDK. Since AnalyticHandler instances are created per chatId, concurrent requests with different MLflow configurations (e.g., different tracking URIs or credentials) will overwrite each other's settings. This can lead to traces being sent to the wrong tracking server in a multi-user environment. If the MLflow SDK supports it, consider using a client-based approach to create independent instances instead of global initialization.

[gemini-code-assist]

The current implementation overwrites the chainSpan map every time a new chain starts. This will cause parent spans or concurrent spans to be lost, preventing them from being correctly retrieved (e.g., for nesting) or ended. Use object spreading to preserve existing spans in the map.

Suggested change

	this.handlers['mlflow'].chainSpan = { [spanId]: chainSpan }
	this.handlers['mlflow'].chainSpan = { ...this.handlers['mlflow'].chainSpan, [spanId]: chainSpan }

References

1. Spreading null or undefined within an object literal is safe and does not require a ?? {} fallback, as it evaluates to an empty object.

[gemini-code-assist]

The llmSpan map is being overwritten, which will cause issues if multiple LLM calls occur within the same trace (e.g., in a loop or multi-agent setup). Use object spreading to maintain all active spans.

Suggested change

	this.handlers['mlflow'].llmSpan = { [spanId]: llmSpan }
	this.handlers['mlflow'].llmSpan = { ...this.handlers['mlflow'].llmSpan, [spanId]: llmSpan }

References

1. Spreading null or undefined within an object literal is safe and does not require a ?? {} fallback, as it evaluates to an empty object.

[gemini-code-assist]

The toolSpan map is being overwritten. This will lead to data loss if multiple tools are executed within the same request context. Use object spreading to preserve the spans.

Suggested change

	this.handlers['mlflow'].toolSpan = { [spanId]: toolSpan }
	this.handlers['mlflow'].toolSpan = { ...this.handlers['mlflow'].toolSpan, [spanId]: toolSpan }

References

1. Spreading null or undefined within an object literal is safe and does not require a ?? {} fallback, as it evaluates to an empty object.

[gemini-code-assist]

Avoid using any for mlflowOptions. It is better to use a specific type from the @mlflow/core package or at least Record<string, any> to maintain better type safety and clarity.

Suggested change

	const mlflowOptions: any = {
	const mlflowOptions: Record<string, any> = {

[Copilot]

Pull request overview

Adds MLflow as a new analytics/tracing provider alongside existing providers (LangSmith/Langfuse/Lunary/LangWatch/Arize/Phoenix/Opik), enabling Flowise to emit chain/LLM/tool traces to an MLflow tracking server via the @mlflow/core SDK.

Changes:

• Add @mlflow/core dependency and lockfile updates for MLflow + transitive OpenTelemetry/Databricks SDK packages.
• Add MLflow provider configuration to the UI analytics settings (including optional experiment ID).
• Implement MLflow span lifecycle wiring in AnalyticHandler (CHAIN/LLM/TOOL spans + flush on shutdown) and add an MLflow credential definition.

Reviewed changes

Copilot reviewed 4 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
pnpm-lock.yaml	Adds @mlflow/core and transitive dependencies needed for MLflow tracing
packages/components/package.json	Declares @mlflow/core dependency for components package
packages/components/src/handler.ts	Implements MLflow provider init + CHAIN/LLM/TOOL span start/end/error handling
packages/components/credentials/MLflowApi.credential.ts	Adds credential schema for MLflow tracking server connection/auth
packages/ui/src/ui-component/extended/AnalyseFlow.jsx	Adds MLflow as an analytics provider option in the UI
packages/ui/src/assets/images/mlflow.svg	Adds MLflow icon asset

Files not reviewed (1)

• pnpm-lock.yaml: Language not supported

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

@mlflow/core pulls in @databricks/sdk-experimental@0.15.0, which declares engines: { node: '>=22.0' }. This repo targets Node 20 (.nvmrc v20.19.2; root package.json allows ^20 and even 18.x), so enabling MLflow/Databricks tracing may break at runtime on supported Node versions. Consider using an @mlflow/core (or Databricks SDK) version compatible with Node 20/18, or add a pnpm override to a Node-20-compatible Databricks SDK (and verify MLflow still works), rather than introducing a Node >=22-only transitive dependency.