MLflow Hybrid Runner

This section provides guides and references to use the MLflow connector. Configure and schedule MLflow metadata and profiler workflows from the Collate UI:

Requirements
Metadata Ingestion

Requirements

To extract metadata, Collate needs two elements:

Tracking URI: Address of local or remote tracking server. More information on the MLflow documentation here
Registry URI: Address of local or remote model registry server.

ML Features

Collate builds ML Features from the input columns of each model’s signature. Columns of type string become categorical features, and every other type becomes numerical. Inputs without a name, such as unnamed tensors, produce no ML Features. MLflow 2.x records the signature on the run that logged the model. MLflow 3.x doesn’t, so for MLflow 3.x models, Collate reads the signature from the model’s MLmodel file in the artifact store. Collate downloads only that file, never the model weights. The ingestion therefore needs read access to the model artifacts. If your tracking server doesn’t proxy artifacts, for example because it runs with --no-serve-artifacts, give the Hybrid Runner its own credentials for the artifact store. If Collate can’t read the file, it still ingests the model without ML Features and logs a warning.

Databricks Unity Catalog

Ingesting models from Databricks Unity Catalog requires the Hybrid Runner, because the connector reads the Databricks credentials from environment variables rather than from the connection form. Use these connection values:

trackingUri: databricks
registryUri: databricks-uc

Set these environment variables on the runner’s ingestion pods:

DATABRICKS_HOST: Your Databricks workspace URL, for example https://<workspace>.cloud.databricks.com
DATABRICKS_TOKEN: A personal access token for a user or service principal that can access the registered models

Store the token in a Kubernetes Secret rather than in your Helm values. Don’t put it in config.ingestionPods.extraEnvs. Helm saves those values in the release and writes them in plain text into the runner’s Deployment. The runner then copies them into every ingestion pod it creates. These steps update a Hybrid Runner that’s already installed. If you haven’t installed one yet, follow Set Up Hybrid Ingestion Runner first, then return to these steps.

Create the Secret in the namespace where the ingestion pods run, which is the chart’s namespace by default:

kubectl create secret generic databricks-mlflow \
  --namespace argo-workflows \
  --from-literal=token='<databricks-token>'

In the runner’s values.yaml, add both variables to the ingestion pods’ container config. The token entry references the Secret, so the token itself never appears in your values:

config:
  ingestionPods:
    customConfig:
      enabled: true
      containerParams:
        env:
          - name: DATABRICKS_HOST
            value: "https://<workspace>.cloud.databricks.com"
          - name: DATABRICKS_TOKEN
            valueFrom:
              secretKeyRef:
                name: databricks-mlflow
                key: token

If the runner uses the Simple Kubernetes executor, which is the default when Argo Workflows isn’t installed, also add the following to values.yaml. The chart applies the container config automatically only for the Argo Workflows executor:
extraEnvs: - name: SIMPLEK8S_CONTAINER_SPEC_FILE_PATH value: /custom-config/container-params.yaml

Apply the updated values to the runner with the Helm CLI, using your own release name and namespace:

helm upgrade collate-prod collate-hybrid/hybrid-ingestion-runner \
  --namespace argo-workflows \
  --values values.yaml

Unity Catalog lists only the models that the token’s user or service principal can access. If the principal can’t access any models, the listing comes back empty instead of failing. To ingest ML Features from MLflow 3.x models, the same principal also needs read access to the model artifacts. Unity Catalog model names have three parts, <catalog>.<schema>.<model>, and filter patterns match against the full name. Databricks also lists its built-in foundation models under system.ai. These models aren’t backed by runs in your workspace, so Collate can’t ingest them and reports each one as failed. To skip them, add an Always exclude rule to the ML Model section that matches regex ^system\.ai\..*.

Metadata Ingestion

To ingest metadata from MLflow, you need to create a service connection. The service connects MLflow with Collate. Once you create a service, Collate automatically starts ingesting metadata.

Step 1: Add New Service

In the left navigation, click Connections.
On the Connections page, click Add New Service.

Step 2: Select a Service and Connector

From the service type dropdown, select Mlmodel Services, then click the MLflow connector tile.

Step 3: Add Service Name and Description

Enter a unique, descriptive Service Name. Collate identifies services by their service name. Enter a name that distinguishes this deployment from other MLflow services you are ingesting metadata from.
Optional: Enter a Description for the service.

Note: The service name cannot be changed after it is set.

Step 4: Configure Connection Options

Specify where ingestion runs, provide your source credentials, and verify the connection.

Select Ingestion Runner

Select an Ingestion Runner: the runner where the ingestion pipeline will execute.

Enter Connection Details

Enter the connection details for MLflow. The right-hand panel in the UI displays inline help for each field.

trackingUri: Mlflow Experiment tracking URI. E.g., http://localhost:5000
registryUri: Mlflow Model registry backend. E.g., mysql+pymysql://mlflow:password@localhost:3307/experiments

For Databricks Unity Catalog, enter databricks as the trackingUri and databricks-uc as the registryUri. For the credentials this needs, see Databricks Unity Catalog.

Tip: When using a Hybrid Ingestion Runner, any sensitive credential fields—such as passwords, API keys, or private keys—must reference secrets using the following format:

password: secret:/my/database/password

This applies only to fields marked as secrets in the connection form (these typically mask input and show a visibility toggle icon). For more information about managing secrets in hybrid setups, see the Hybrid Ingestion Runner Secret Management Guide

Test Connection

Once the credentials have been added, click on Test Connection and Save the changes.

Step 5: Configure Ingestion Options

In the What to Ingest step, use filter patterns to control which assets Collate ingests from your ML model service. Filter patterns use regular expressions applied to model names.

How Filter Patterns Work

Include: Add one or more comma-separated regular expressions. Collate ingests only assets whose names match at least one expression. Leave blank to include all assets.
Exclude: Add one or more comma-separated regular expressions. Collate skips any asset whose name matches an expression. Leave blank to exclude nothing.

Rules match asset names using one of five expressions:

contains: matches any name containing the value. For example, sales matches my_sales_data and sales_2024.
starts with: matches names beginning with the value. For example, prod_ matches prod_db and prod_schema.
ends with: matches names ending with the value. For example, _raw matches events_raw and logs_raw.
is exactly: matches the exact name only. For example, analytics matches only analytics.
matches regex: matches names using a regular expression. For example, ^prod_.*_v\d+$ matches prod_events_v1.

When both Include and Exclude are set, Exclude takes priority.

Tip: Leave the filter pattern empty to ingest all ML models available in the source.

Filter Options The ML Model section includes the following filter options:

ML Model: Controls which ML models Collate ingests from the source.

Each section provides the following controls:

Scan Mode: You can choose between the following scan modes:
- Scan all: Ingests every asset of that type the connector can access. This is the default.
- Only specific: Enables include rules so only assets matching at least one rule are ingested.
Exclude system toggle: Use this toggle to automatically filter out system-reserved names defined by the connector — for example, Exclude system databases for the Databases section.
Always exclude: Add permanent exclusion rules (shown in red). Assets matching these rules are never ingested, regardless of include rules.
Preview: Shows a real-time summary of what will be in scope based on your current rules.
Include rules (available only in Only specific mode): Click + Add to define a rule. Added rules appear as chips; an asset is included if it matches any rule.

Tip: If AutoPilot is enabled, usage tracking, data lineage, and other downstream workflows start automatically after the first metadata ingestion completes.

Step 6: Create & Deploy

Click Create & Deploy to deploy the agent and start the first metadata ingestion run. Collate saves the service configuration and immediately begins pulling metadata from the source. To monitor ingestion progress or view the service you just added, go to Connections in the left navigation and select your service.

Configure Metadata Agent and Schedule Ingestion

The Metadata Agent extracts ML models, features, and other structural metadata from your source and keeps your Collate catalog in sync. It powers discovery, lineage, and governance across your data assets. When you click Create & Deploy, Collate automatically deploys a Metadata Agent for this service and triggers the first ingestion run. View its status and run history from the Agents tab on the service detail page. To configure the additional Metadata Agent and schedule ingestion, follow these steps:

In the left navigation, click Connections and select your service.
Click the Agents tab.
Click Add Agent and select Metadata from the dropdown. For some services, the dropdown is not available and clicking Add Agent takes you directly to the agent configuration page.

On the Configure Ingestion page, do the following and click Next.

Name this Ingestion: Enter a unique recognizable name for this ingestion pipeline.

Agent Setup: Configure the core parameters for this agent. The following fields are available:

Field	Default	Description
Db Service Prefixes	—	List of database service name prefixes used to resolve lineage between ML models and database assets. Accepted formats: `DBService`, `DBService.Database`, `DBService.Database.Schema`, `DBService.Database.Schema.Table`.
Mark Deleted ML Models	On	Soft-delete ML models in Collate when they are removed from the source. Associated entities like lineage are also deleted.

Filter Patterns: Apply include or exclude rules to scope which ML models this agent ingests. These follow the same filter options described in Step 5.

Scope & Behaviour: Control what metadata to include and how to handle deletions. Toggle each option on or off based on your needs:

Toggle	Default	Description
Enable Debug Log	Off	Sets the ingestion log level to DEBUG. Useful for troubleshooting.
Override Metadata	Off	When on, source values overwrite existing descriptions, tags, owners, and display names in Collate. When off, Collate only updates fields that have no existing value.
Override Lineage	Off	When on, existing lineage is replaced with newly extracted lineage on each run.

On the Schedule Interval page, set when the agent runs:
- Schedule: Choose a preset interval (Hourly, Daily, Weekly, Monthly) or enter a custom cron expression.
- On-Demand: No automatic schedule; trigger the agent manually when needed.
Click Add to deploy the agent.

Troubleshooting

MLflow Troubleshooting

Learn more about how to troubleshoot common MLflow connector issues and resolve configuration or ingestion errors.

Connectors

Connectors

Requirements

ML Features

Databricks Unity Catalog

Metadata Ingestion

Step 1: Add New Service

Step 2: Select a Service and Connector

Step 3: Add Service Name and Description

Step 4: Configure Connection Options

Select Ingestion Runner

Enter Connection Details

Test Connection

Step 5: Configure Ingestion Options

How Filter Patterns Work

Step 6: Create & Deploy

Configure Metadata Agent and Schedule Ingestion

Troubleshooting

MLflow Troubleshooting

​Requirements

​ML Features

​Databricks Unity Catalog

​Metadata Ingestion

​Step 1: Add New Service

​Step 2: Select a Service and Connector

​Step 3: Add Service Name and Description

​Step 4: Configure Connection Options

​Select Ingestion Runner

​Enter Connection Details

​Test Connection

​Step 5: Configure Ingestion Options

​How Filter Patterns Work

​Step 6: Create & Deploy

​Configure Metadata Agent and Schedule Ingestion

​Troubleshooting

MLflow Troubleshooting

Requirements

ML Features

Databricks Unity Catalog

Metadata Ingestion

Step 1: Add New Service

Step 2: Select a Service and Connector

Step 3: Add Service Name and Description

Step 4: Configure Connection Options

Select Ingestion Runner

Enter Connection Details

Test Connection

Step 5: Configure Ingestion Options

How Filter Patterns Work

Step 6: Create & Deploy

Configure Metadata Agent and Schedule Ingestion

Troubleshooting