Debugging guide for Model Serving

This article demonstrates debugging steps for common issues that users might encounter when working with model serving endpoints. Common issues could include errors users encounter when the endpoint fails to initialize or start, build failures related to the container, or problems during the operation or running of the model on the endpoint.

:::tip Validate before debugging Having deployment issues? Start with pre-deployment validation to catch common problems before they occur. :::

Debug your container build

Databricks recommends reviewing logs for debugging and troubleshooting errors in your model serving workloads. See Monitor model quality and endpoint health for information about logs and how to view them.

The event logs (click the Events tab) in the workspace UI contain information about the progress of a container build. A successful container build is highlighted by a SERVED_ENTITY_CONTAINER_EVENT event type along with a message Container image creation finished successfully. If you do not see any build event or message after an hour of creating the endpoint, reach out to Databricks support for assistance.

If your build is successful, but you encounter other errors see Debug after container build succeeds. If your build fails, see Debug after container build failure.

Debug after container build succeeds

Even if the container builds successfully, there might be issues when you run the model or during the operation of the endpoint itself. The following subsections detail common issues and how to troubleshoot them.

Missing dependency

You might get an error like An error occurred while loading the model. No module named <module-name>., which might indicate that a dependency is missing from the container. Verify that you properly denoted all the dependencies that should be included in the build of the container. Pay special attention to custom libraries and ensure that the .whl files are included as artifacts.

Model fails or times out when requests are sent to the endpoint

You might receive an error like Encountered an unexpected error while evaluating the model. Verify that the input is compatible with the model for inference. when predict() is called on your model.

This error can indicate a code issue in the predict() function. Databricks recommends that you load the model from MLflow in a notebook and call it. Doing so highlights the issues in the predict() function, and you can see where the failure is happening within the method.

Root cause analysis of failed requests

If a request to an endpoint fails, you can perform root cause analysis by using inference tables. If enabled, inference tables automatically log all requests and responses to your endpoint in a Unity Catalog table for you to query.

• For external models, provisioned throughput endpoints, and AI agents, see Monitor served models using AI Gateway-enabled inference tables.

• For custom models, see Inference tables for monitoring and debugging models.

To query inference tables:

1. In your workspace, go to the Serving tab and select your endpoint name.
2. In the Inference tables section, find the inference table's fully-qualified name. For example, my-catalog.my-schema.my-table.
3. Run the following in a Databricks notebook:

   %sql
   SELECT * FROM my-catalog.my-schema.my-table
   

4. View and filter on columns such as request, response, request_time and status_code to understand the requests and narrow down results.

   %sql
   SELECT * FROM my-catalog.my-schema.my-table
   WHERE status_code != 200
   

5. If you enabled agent tracing for AI agents, see the Response column to view detailed traces. See Enable inference tables for AI agents.

Workspace exceeds provisioned concurrency

You might receive a Workspace exceeded provisioned concurrency quota error. This indicates that you have reached your workspace quota for provisioned concurrency. See Model Serving limits and regions for more information about concurrency limits.

You can free up this quota by deleting or stopping unused endpoints.

This limit can be increased depending on region availability. Reach out to your Databricks account team and provide your workspace ID to request a concurrency increase.

Workspace exceeds parallel requests limit

You might receive the following 429 error: Exceeded max number of parallel requests. Please contact your Databricks representative to increase the limit. This limit indicates that you have reached the workspace limit on the maximum number of requests that can be sent in parallel. See Model Serving limits and regions for more information about this limit.

Databricks recommends moving to route optimized endpoints, where this limit has been removed. If you cannot move to route optimized endpoints, you can either reduce the number of clients sending inference requests or contact your Databricks representative for a quota increase.

Too many concurrent requests

You might receive the following 429 error: Too many concurrent requests. Consider increasing the provisioned concurrency of the served entity. This error indicates that your endpoint's current provisioned concurrency cannot handle the incoming traffic volume. If you have enabled autoscaling for your endpoint, the system will automatically provision additional concurrency up to the endpoint's configured limit to handle the increased load. If autoscaling is not enabled, consider manually increasing the provisioned concurrency or enabling autoscaling to handle traffic spikes.

Debug after container build failure

This section details issues that might occur when your build fails.

OSError: [Errno 28] No space left on device

The No space left error can be due to too many large artifacts being logged alongside the model unnecessarily. Check in MLflow that extraneous artifacts are not logged alongside the model and try to redeploy the slimmed down package.

Azure Firewall issues with serving models from Unity Catalog

You might see the error: Build could not start due to an internal error. If you are serving a model from UC and Azure Firewall is enabled, this is not supported by default..

Reach out to your Databricks account team to help resolve.

Build failure due to lack of GPU availability

Due to restrictions in GPU supply and availability, your GPU build may fail with this error: Build could not start due to an internal error - please contact your Databricks representative..

Reach out to your Databricks account team to help resolve. Depending on region availability, the team can provision more GPU resources.

Installed library package versions

Databricks recommends that you define all important libraries as model dependencies to ensure consistent and reproducible model behavior across environments. In the build logs you can confirm the package versions that are installed correctly.

• For MLflow versions, if you do not have a version specified, Model Serving uses the latest version.
• For custom GPU serving, Model Serving installs the recommended versions of cuda and cuDNN according to public PyTorch and Tensorflow documentation.

Log models that require flash-attn

If you are logging a model that requires flash-attn, Databricks recommends using a custom wheel version of flash-attn. Otherwise, build errors such as ModuleNotFoundError: No module named 'torch' can result.

To use a custom wheel version of flash-attn, specify all pip requirements as a list and pass it as a parameter into your mlflow.transformers.log_model function. You must also specify the pytorch, torch, and torchvision versions that are compatible with the CUDA version specified in your flash attn wheel.

For example, Databricks recommends using the following versions and wheels for CUDA 11.8:

• Pytorch
• Torch 2.0.1+cu118
• Torchvision 0.15.2+cu118
• Flash-Attn

logged_model=mlflow.transformers.log_model(
transformers_model=test_pipeline,
       artifact_path="artifact_path",
       pip_requirements=["--extra-index-url https://download.pytorch.org/whl/cu118", "mlflow==2.13.1", "setuptools<70.0.0", "torch==2.0.1+cu118", "accelerate==0.31.0", "astunparse==1.6.3", "bcrypt==3.2.0", "boto3==1.34.39", "configparser==5.2.0", "defusedxml==0.7.1", "dill==0.3.6", "google-cloud-storage==2.10.0", "ipython==8.15.0", "lz4==4.3.2", "nvidia-ml-py==12.555.43", "optree==0.12.1", "pandas==1.5.3", "pyopenssl==23.2.0", "pytesseract==0.3.10", "scikit-learn==1.3.0", "sentencepiece==0.1.99", "torchvision==0.15.2+cu118", "transformers==4.41.2", "https://github.com/Dao-AILab/flash-attention/releases/download/v2.5.8/flash_attn-2.5.8+cu118torch2.0cxx11abiFALSE-cp311-cp311-linux_x86_64.whl"],
       input_example=input_example,
       registered_model_name=registered_model_name)