Skip to main content

​
Troubleshooting

​
Workflow Deployment Error

If there were any errors during the workflow deployment process, the Ingestion Pipeline Entity will still be created, but no workflow will be present in the Ingestion container.
  • You can then Edit the Ingestion Pipeline and Deploy it again.
  • From the Connection tab, you can also Edit the Service if needed.

​
Connector Debug Troubleshooting

This section provides instructions to help resolve common issues encountered during connector setup and metadata ingestion in Collate. Below are some of the most frequently observed troubleshooting scenarios.

​
How to Enable Debug Logging for Any Ingestion

To enable debug logging for any ingestion workflow in Collate:
  1. Navigate to Connections Go to Connections > Service Type (e.g., Database) in the Collate UI.
  2. Select a Service Choose the specific service for which you want to enable debug logging.
  3. Access Agents Tab Go to the Agents tab and click the three-dot menu on the right-hand side of the ingestion type, and select Edit.
  4. Enable Debug Logging In the configuration dialog, enable the Debug Log option and click Next.
  5. Schedule and Submit Configure the schedule if needed and click Submit to apply the changes.

​
Permission Issues

If you encounter permission-related errors during connector setup or metadata ingestion, ensure that all the prerequisites and access configurations specified for each connector are properly implemented. Refer to the connector-specific documentation to verify the required permissions.

​
MLflow Ingestion Issues

These issues are specific to the MLflow connector. The messages quoted in each section appear in the ingestion logs.

​
No ML Models Are Ingested

The ingestion run succeeds, but no ML models appear in Collate. The logs contain The MLflow registry returned no registered models, which means the registry listing came back empty. Check the following:
  • Registry URI: Make sure registryUri points at the registry that holds your models. For Databricks, databricks-uc lists the models in Unity Catalog, while databricks lists the legacy Workspace Model Registry. For the Unity Catalog setup, see Databricks Unity Catalog.
  • Permissions: Make sure the credentials can see the models. Access-controlled registries list only the models the caller can read and return an empty list rather than an error.
While the listing is empty, Collate doesn’t mark any existing ML models as deleted.

​
Some ML Models Are Missing from a Large Registry

The logs contain Stopped listing registered models after 100 pages. Collate pages through the registry automatically but stops after 100 pages. It ingests the models it listed and doesn’t mark any ML models as deleted in that run. To ingest a registry this large, contact Collate support.

​
ML Models Are Ingested Without ML Features

An ML model appears in Collate with its hyperparameters and ML Store, but without ML Features. Collate builds ML Features from the input columns of the model’s signature, so check the following:
  • No signature: The model was logged without a signature. Log it with one, for example by passing signature to log_model.
  • Unnamed inputs: The signature’s inputs have no names, as with unnamed tensors. Collate can’t map these to ML Features.
  • Unreadable artifacts: The logs contain Cannot read the model signature of <model> from its artifacts. MLflow 3.x doesn’t record the signature on the run, so Collate reads it from the model’s MLmodel file in the artifact store. Make sure the ingestion can read the model artifacts. If your tracking server doesn’t proxy artifacts, the ingestion needs its own credentials for the artifact store.

​
ML Models Fail with Version Not Found or Run ID Not Found

The ingestion run reports an ML model as failed with one of these errors:
  • Version not found: Collate couldn’t resolve a numeric version for the model. Either the model has no versions, or the version search failed. Look for Error searching for versions of model or Gave up paginating versions in the logs.
  • Run ID not found: The model’s latest version isn’t linked to an MLflow run, for example because it was registered from files outside a run. Collate needs the run to ingest the model.
Databricks’ built-in system.ai foundation models fail this way because they aren’t backed by runs in your workspace. To skip them, add an Always exclude rule to the ML Model section that matches regex ^system\.ai\..*.

​
Removed ML Models Aren’t Marked as Deleted

With Mark Deleted ML Models turned on, Collate soft-deletes ML models that are no longer in the registry. It skips this step when the registry listing didn’t complete or came back empty, so a configuration or permissions problem can’t delete your ML models in Collate. When this happens, the logs contain Skipping deleted-model reconciliation. Deletion resumes on the next run that lists the whole registry. If you remove every model from the registry, delete the corresponding ML models in Collate manually.