Package Requirements

Skill: databricks-model-serving

What You Can Build

You can get your agent or model running on any Databricks Runtime version with the correct package versions, avoiding the dependency conflicts that cause “works on my cluster, fails on the endpoint” failures. The package landscape for MLflow 3, LangGraph, and the Databricks agent stack moves fast — getting versions right is the difference between a 5-minute deploy and a multi-hour debugging session.

In Action

“Install the full MLflow 3 agent stack on my DBR 16.1+ cluster for building a ResponsesAgent with LangGraph. Use Python.”

%pip install -U mlflow==3.6.0 databricks-langchain langgraph==0.3.4 databricks-agents pydantic
dbutils.library.restartPython()

Key decisions:

• mlflow==3.6.0 — pin exactly. MLflow 3 introduced ResponsesAgent and the file-based logging pattern. Minor versions can introduce breaking changes to the agent interface.
• databricks-langchain without a version pin — this package tracks the runtime and updates frequently. Let it resolve to latest unless you have a known-good version.
• langgraph==0.3.4 — pin exactly. LangGraph’s API surface changes between minor versions, and the serving endpoint must match.
• dbutils.library.restartPython() is mandatory — without it, the running Python process still has old module versions cached.

More Patterns

Memory-Enabled Agent Stack

“Install the agent packages with Lakebase memory support. Use Python.”

%pip install -U mlflow==3.6.0 databricks-langchain[memory] langgraph==0.3.4 databricks-agents
dbutils.library.restartPython()

The [memory] extra installs Lakebase client libraries for agents that maintain conversation state across sessions.

Vector Search Agent Stack

“Install packages for an agent that uses Vector Search as a retriever tool. Use Python.”

%pip install -U mlflow==3.6.0 databricks-langchain databricks-vectorsearch langgraph==0.3.4
dbutils.library.restartPython()

databricks-vectorsearch provides VectorSearchRetrieverTool, which wraps a vector index as a LangChain tool your agent can call.

Capture Versions for pip_requirements

“Build a pip_requirements list from the versions currently installed on my cluster so my serving endpoint matches exactly. Use Python.”

from pkg_resources import get_distribution

pip_requirements = [
    f"mlflow=={get_distribution('mlflow').version}",
    f"databricks-langchain=={get_distribution('databricks-langchain').version}",
    f"langgraph=={get_distribution('langgraph').version}",
]

Use this in your log_model.py to capture exact versions from the cluster where your agent is tested and working. The serving endpoint will install these same versions.

Audit Installed Versions

“Check which versions of the key agent packages are installed on my cluster. Use Python.”

import pkg_resources

packages = ['mlflow', 'langchain', 'langgraph', 'pydantic', 'databricks-langchain']
for pkg in packages:
    try:
        version = pkg_resources.get_distribution(pkg).version
        print(f"{pkg}: {version}")
    except pkg_resources.DistributionNotFound:
        print(f"{pkg}: NOT INSTALLED")

Run this before and after %pip install to confirm the upgrade took effect. Also useful for debugging version mismatch issues between cluster and serving endpoint.

Set Up Local Environment Variables

“Configure my local machine to authenticate with Databricks for local testing. Use bash.”

# Option 1: Host + Token
export DATABRICKS_HOST="https://your-workspace.databricks.com"
export DATABRICKS_TOKEN="your-token"

# Option 2: Profile-based (reads from ~/.databrickscfg)
export DATABRICKS_CONFIG_PROFILE="your-profile"

Local testing with mlflow.models.predict or direct agent imports requires authentication. These environment variables are read by both the Databricks SDK and MLflow.

Watch Out For

• Skipping restartPython() after %pip install — the Python process caches old modules. You will see import errors or stale behavior from the previous version until you restart.
• Pinning databricks-langchain to an old version — this package tracks the Databricks runtime and updates frequently. Unless you have a specific version requirement, let it resolve to latest.
• Mismatched versions between cluster and pip_requirements — the cluster runs mlflow==3.6.0 but your log_model.py says mlflow>=3.0. The serving endpoint resolves >=3.0 differently and breaks. Always capture exact versions.
• Using DBR < 16.1 for agent development — older runtimes ship with MLflow 2.x and require manual upgrades of many packages. DBR 16.1+ has the GenAI stack pre-installed, saving significant setup time.