google__adk-docs/docs/integrations/mlflow-tracing.md

catalog_title, catalog_description, catalog_icon, catalog_tags

catalog_title	catalog_description	catalog_icon	catalog_tags
MLflow Tracing	Ingest OpenTelemetry traces for agent runs, tool calls, and model requests	/integrations/assets/mlflow.png	observability

MLflow observability for ADK

Supported in ADKPython

MLflow Tracing provides first-class support for ingesting OpenTelemetry (OTel) traces. Google ADK emits OTel spans for agent runs, tool calls, and model requests, which you can send directly to an MLflow Tracking Server for analysis and debugging.

Prerequisites

• MLflow version 3.6.0 or newer. OpenTelemetry ingestion is only supported in MLflow 3.6.0+.
• A SQL-based backend store (e.g., SQLite, PostgreSQL, MySQL). File-based stores do not support OTLP ingestion.
• Google ADK installed in your environment.

Install dependencies

pip install "mlflow>=3.6.0" google-adk opentelemetry-sdk opentelemetry-exporter-otlp-proto-http

Start the MLflow Tracking Server

Start MLflow with a SQL backend and a port (5000 in this example):

mlflow server --backend-store-uri sqlite:///mlflow.db --port 5000

You can point --backend-store-uri to other SQL backends (PostgreSQL, MySQL, MSSQL). OTLP ingestion is not supported with file-based backends.

Configure OpenTelemetry (required)

You must configure an OTLP exporter and set a global tracer provider before using any ADK components so that spans are emitted to MLflow.

Initialize the OTLP exporter and global tracer provider in code before importing or constructing ADK agents/tools:

# my_agent/agent.py
from opentelemetry import trace
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor

exporter = OTLPSpanExporter(
    endpoint="http://localhost:5000/v1/traces",
    headers={"x-mlflow-experiment-id": "123"}  # replace with your experiment id
)

provider = TracerProvider()
provider.add_span_processor(SimpleSpanProcessor(exporter))
trace.set_tracer_provider(provider)  # set BEFORE importing/using ADK

This configures the OpenTelemetry pipeline and sends ADK spans to the MLflow server on each run.

Example: Trace an ADK agent

Now you can add the agent code for a simple math agent, after the code that sets up the OTLP exporter and tracer provider:

# my_agent/agent.py
from google.adk.agents import LlmAgent
from google.adk.tools import FunctionTool


def calculator(a: float, b: float) -> str:
    """Add two numbers and return the result."""
    return str(a + b)


calculator_tool = FunctionTool(func=calculator)

root_agent = LlmAgent(
    name="MathAgent",
    model="gemini-flash-latest",
    instruction=(
        "You are a helpful assistant that can do math. "
        "When asked a math problem, use the calculator tool to solve it."
    ),
    tools=[calculator_tool],
)

Run the agent with:

adk run my_agent

And ask it a math problem:

What is 12 + 34?

You should then see output similar to:

[MathAgent]: The answer is 46.

View traces in MLflow

Open the MLflow UI at http://localhost:5000, select your experiment, and inspect the trace tree and spans generated by your ADK agent.

Tips

• Set the tracer provider before importing or initializing ADK objects so all spans are captured.
• Behind a proxy or on a remote host, replace localhost:5000 with your server address.

Resources

• MLflow Tracing Documentation: Official documentation for MLflow Tracing that covers other library integrations and downstream usage of traces, such as evaluation, monitoring, searching, and more.
• OpenTelemetry in MLflow: Detailed guide on how to use OpenTelemetry with MLflow.
• MLflow for Agents: Comprehensive guide on how to use MLflow for building production-ready agents.