This browser is no longer supported.

Upgrade to Microsoft Edge to take advantage of the latest features, security updates, and technical support.

Learn how to resolve common issues in the deployment and scoring of Azure Machine Learning online endpoints.

This document is structured in the way you should approach troubleshooting:

The section HTTP status codes explains how invocation and prediction errors map to HTTP status codes when scoring endpoints with REST requests.

Prerequisites

Deploy locally

Local deployment is deploying a model to a local Docker environment. Local deployment is useful for testing and debugging before deployment to the cloud.

You can also use Azure Machine Learning inference HTTP server Python package to debug your scoring script locally. Debugging with the inference server helps you to debug the scoring script before deploying to local endpoints so that you can debug without being affected by the deployment container configurations.

Local deployment supports creation, update, and deletion of a local endpoint. It also allows you to invoke and get logs from the endpoint.

To use local deployment, add local=True parameter in the command:

ml_client.begin_create_or_update(online_deployment, local=True)
ml_client is the instance for MLCLient class, and online_deployment is the instance for either ManagedOnlineDeployment class or KubernetesOnlineDeployment class.
As a part of local deployment the following steps take place:
Docker either builds a new container image or pulls an existing image from the local Docker cache. An existing image is used if there's one that matches the environment part of the specification file.
Docker starts a new container with mounted local artifacts such as model and code files.
For more, see Deploy locally in Deploy and score a machine learning model.
Use Visual Studio Code to test and debug your endpoints locally. For more information, see debug online endpoints locally in Visual Studio Code.
Conda installation
Generally, issues with MLflow deployment stem from issues with the installation of the user environment specified in the conda.yaml file.
To debug conda installation problems, try the following steps:
Check the logs for conda installation. If the container crashed or taking too long to start up, it is likely that conda environment update has failed to resolve correctly.
Install the mlflow conda file locally with the command conda env create -n userenv -f <CONDA_ENV_FILENAME>.
If there are errors locally, try resolving the conda environment and creating a functional one before redeploying.
If the container crashes even if it resolves locally, the SKU size used for deployment may be too small.
Conda package installation occurs at runtime, so if the SKU size is too small to accommodate all of the packages detailed in the conda.yaml environment file, then the container may crash.
A Standard_F4s_v2 VM is a good starting SKU size, but larger ones may be needed depending on which dependencies are specified in the conda file.
For Kubernetes online endpoint, the Kubernetes cluster must have minimum of 4 vCPU cores and 8-GB memory.
Get container logs
You can't get direct access to the VM where the model is deployed. However, you can get logs from some of the containers that are running on the VM. The amount of information you get depends on the provisioning status of the deployment. If the specified container is up and running, you see its console output; otherwise, you get a message to try again later.
There are two types of containers that you can get the logs from:
Inference server: Logs include the console log (from the inference server) which contains the output of print/logging functions from your scoring script (score.py code).
Storage initializer: Logs contain information on whether code and model data were successfully downloaded to the container. The container runs before the inference server container starts to run.
To see log output from a container, use the following CLI command:
az ml online-deployment get-logs -e <endpoint-name> -n <deployment-name> -l 100
az ml online-deployment get-logs --endpoint-name <endpoint-name> --name <deployment-name> --lines 100
Add --resource-group and --workspace-name to these commands if you have not already set these parameters via az configure.
To see information about how to set these parameters, and if you have already set current values, run:
az ml online-deployment get-logs -h
By default the logs are pulled from the inference server.
If you use Python logging, ensure you use the correct logging level order for the messages to be published to logs. For example, INFO.
You can also get logs from the storage initializer container by passing –-container storage-initializer.
Add --help and/or --debug to commands to see more information.
To see log output from container, use the get_logs method as follows:
ml_client.online_deployments.get_logs(
    name="<deployment-name>", endpoint_name="<endpoint-name>", lines=100
To see information about how to set these parameters, see
reference for get-logs
By default the logs are pulled from the inference server.
If you use Python logging, ensure you use the correct logging level order for the messages to be published to logs. For example, INFO.
You can also get logs from the storage initializer container by adding container_type="storage-initializer" option.
ml_client.online_deployments.get_logs(
    name="<deployment-name>", endpoint_name="<endpoint-name>", lines=100, container_type="storage-initializer"
To see log output from a container, use the Endpoints in the studio:
In the left navigation bar, select Endpoints.
(Optional) Create a filter on compute type to show only managed compute types.
Select an endpoint's name to view the endpoint's details page.
Select the Deployment logs tab in the endpoint's details page.
Use the dropdown to select the deployment whose log you want to see.
For Kubernetes online endpoint, the administrators are able to directly access the cluster where you deploy the model, which is more flexible for them to check the log in Kubernetes. For example:
kubectl -n <compute-namespace> logs <container-name>
Request tracing
There are two supported tracing headers:
x-request-id is reserved for server tracing. We override this header to ensure it's a valid GUID.
When you create a support ticket for a failed request, attach the failed request ID to expedite the investigation.
x-ms-client-request-id is available for client tracing scenarios. This header is sanitized to only accept alphanumeric characters, hyphens and underscores, and is truncated to a maximum of 40 characters.
Common deployment errors
This is a list of common deployment errors that are reported as part of the deployment operation status.
ImageBuildFailure
OutOfQuota
BadArgument
ResourceNotReady
ResourceNotFound
OperationCanceled
If you are creating or updating a Kubernetes online deployment, you can see Common errors specific to Kubernetes deployments.
ERROR: ImageBuildFailure
This error is returned when the environment (docker image) is being built. You can check the build log for more information on the failure(s). The build log is located in the default storage for your Azure Machine Learning workspace. The exact location may be returned as part of the error. For example, "The build log is available in the workspace blob store '[storage-account-name]' under the path '/azureml/ImageLogs/your-image-id/build.log'". In this case, "azureml" is the name of the blob container in the storage account.
This is a list of common image build failure scenarios:
Azure Container Registry (ACR) authorization failure
Generic or unknown failure
We also recommend reviewing the default probe settings in case of ImageBuild timeouts.
Container registry authorization failure
If the error message mentions "container registry authorization failure" that means you cannot access the container registry with the current credentials.
The desynchronization of a workspace resource's keys can cause this error and it takes some time to automatically synchronize.
However, you can manually call for a synchronization of keys, which may resolve the authorization failure.
Container registries that are behind a virtual network may also encounter this error if set up incorrectly. You must verify that the virtual network that you have set up properly.
Generic image build failure
As stated above, you can check the build log for more information on the failure.
If no obvious error is found in the build log and the last line is Installing pip dependencies: ...working..., then a dependency may cause the error. Pinning version dependencies in your conda file can fix this problem.
We also recommend deploying locally to test and debug your models locally before deploying to the cloud.
ERROR: OutOfQuota
This is a list of common resources that might run out of quota when using Azure services:
Memory
Role assignments
Endpoints
Region-wide VM capacity
Other
Additionally, this is a list of common resources that might run out of quota only for Kubernetes online endpoint:
Kubernetes
CPU Quota
Before deploying a model, you need to have enough compute quota. This quota defines how much virtual cores are available per subscription, per workspace, per SKU, and per region. Each deployment subtracts from available quota and adds it back after deletion, based on type of the SKU.
A possible mitigation is to check if there are unused deployments that you can delete. Or you can submit a request for a quota increase.
Disk quota
This issue happens when the size of the model is larger than the available disk space and the model is not able to be downloaded. Try a SKU with more disk space or reducing the image and model size.
This issue happens when the size of the model is larger than the available disk space and the model is not able to be downloaded. Try a SKU with more disk space or reducing the image and model size.
Memory quota
This issue happens when the memory footprint of the model is larger than the available memory. Try a SKU with more memory.
Role assignment quota
When you are creating a managed online endpoint, role assignment is required for the managed identity to access workspace resources. If you've reached the role assignment limit, try to delete some unused role assignments in this subscription. You can check all role assignments in the Azure portal by navigating to the Access Control menu.
Endpoint quota
Try to delete some unused endpoints in this subscription. If all of your endpoints are actively in use, you can try requesting an endpoint quota increase.
For Kubernetes online endpoints, there is the endpoint quota boundary at the cluster level as well, you can check the Kubernetes online endpoint quota section for more details.
Kubernetes quota
This issue happens when the requested CPU or memory couldn't be satisfied due to all nodes are unschedulable for this deployment, such as nodes are cordoned or nodes are unavailable.
The error message will typically indicate the resource insufficient in cluster, for example, OutOfQuota: Kubernetes unschedulable. Details:0/1 nodes are available: 1 Too many pods..., which means that there are too many pods in the cluster and not enough resources to deploy the new model based on your request.
You can try the following mitigation to address this issue:
For IT ops who maintain the Kubernetes cluster, you can try to add more nodes or clear some unused pods in the cluster to release some resources.
For machine learning engineers who deploy models, you can try to reduce the resource request of your deployment:
If you directly define the resource request in the deployment configuration via resource section, you can try to reduce the resource request.
If you use instance type to define resource for model deployment, you can contact the IT ops to adjust the instance type resource configuration, more detail you can refer to How to manage Kubernetes instance type.
Region-wide VM capacity
Due to a lack of Azure Machine Learning capacity in the region, the service has failed to provision the specified VM size. Retry later or try deploying to a different region.




    

Other quota
To run the score.py provided as part of the deployment, Azure creates a container that includes all the resources that the score.py needs, and runs the scoring script on that container.
If your container could not start, this means scoring could not happen. It might be that the container is requesting more resources than what instance_type can support. If so, consider updating the instance_type of the online deployment.
To get the exact reason for an error, run:
Azure CLI
Python SDK
Studio
ml_client.online_deployments.get_logs(
    name="<deployment-name>", endpoint_name="<endpoint-name>", lines=100
In the left navigation bar, select Endpoints.
(Optional) Create a filter on compute type to show only managed compute types.
Select an endpoint name to view the endpoint's details page.
Select the Deployment logs tab in the endpoint's details page.
Use the dropdown to select the deployment whose log you want to see.
ERROR: BadArgument
This is a list of reasons you might run into this error when using either managed online endpoint or Kubernetes online endpoint:
Subscription does not exist
Startup task failed due to authorization error
Startup task failed due to incorrect role assignments on resource
Unable to download user container image
Unable to download user model
Additionally, this is a list of reasons you might run into this error only when using Kubernetes online endpoint:
Resource request was greater than limits
azureml-fe for kubernetes online endpoint is not ready
Subscription does not exist
The Azure subscription that is entered must be existing. This error occurs when we cannot find the Azure subscription that was referenced. This is likely due to a typo in the subscription ID. Please double-check that the subscription ID was correctly typed and that it is currently active.
For more information about Azure subscriptions, you can see the prerequisites section.
Authorization error
After you've provisioned the compute resource (while creating a deployment), Azure tries to pull the user container image from the workspace Azure Container Registry (ACR) and mount the user model and code artifacts into the user container from the workspace storage account.
To do these, Azure uses managed identities to access the storage account and the container registry.
If you created the associated endpoint with System Assigned Identity, Azure role-based access control (RBAC) permission is automatically granted, and no further permissions are needed.
If you created the associated endpoint with User Assigned Identity, the user's managed identity must have Storage blob data reader permission on the storage account for the workspace, and AcrPull permission on the Azure Container Registry (ACR) for the workspace. Make sure your User Assigned Identity has the right permission.
For more information, please see Container Registry Authorization Error.
Unable to download user container image
It's possible that the user container couldn't be found. Check container logs to get more details.
Make sure container image is available in workspace ACR.
For example, if image is testacr.azurecr.io/azureml/azureml_92a029f831ce58d2ed011c3c42d35acb:latest check the repository with
az acr repository show-tags -n testacr --repository azureml/azureml_92a029f831ce58d2ed011c3c42d35acb --orderby time_desc --output table.
Unable to download user model
It is possible that the user's model can't be found. Check container logs to get more details.
Make sure whether you have registered the model to the same workspace as the deployment. To show details for a model in a workspace:
Azure CLI
Python SDK
Studio
You must specify either version or label to get the model's information.
You can also check if the blobs are present in the workspace storage account.
For example, if the blob is https://foobar.blob.core.windows.net/210212154504-1517266419/WebUpload/210212154504-1517266419/GaussianNB.pkl, you can use this command to check if it exists:
az storage blob exists --account-name foobar --container-name 210212154504-1517266419 --name WebUpload/210212154504-1517266419/GaussianNB.pkl --subscription <sub-name>`
If the blob is present, you can use this command to obtain the logs from the storage initializer:
Azure CLI
Python SDK
Studio
ml_client.online_deployments.get_logs(
  name="<deployment-name>", endpoint_name="<endpoint-name>", lines=100, container_type="storage-initializer"
Resource requests greater than limits
Requests for resources must be less than or equal to limits. If you don't set limits, we set default values when you attach your compute to an Azure Machine Learning workspace. You can check limits in the Azure portal or by using the az ml compute show command.
azureml-fe not ready
The front-end component (azureml-fe) that routes incoming inference requests to deployed services automatically scales as needed. It's installed during your k8s-extension installation.
This component should be healthy on cluster, at least one healthy replica. You will get this error message if it's not available when you trigger kubernetes online endpoint and deployment creation/update request.
Please check the pod status and logs to fix this issue, you can also try to update the k8s-extension installed on the cluster.
ERROR: ResourceNotReady
To run the score.py provided as part of the deployment, Azure creates a container that includes all the resources that the score.py needs, and runs the scoring script on that container. The error in this scenario is that this container is crashing when running, which means scoring can't happen. This error happens when:
There's an error in score.py. Use get-logs to diagnose common problems:
A package that score.py tries to import isn't included in the conda environment.
A syntax error.
A failure in the init() method.
If get-logs isn't producing any logs, it usually means that the container has failed to start. To debug this issue, try deploying locally instead.
Readiness or liveness probes aren't set up correctly.
Container initialization is taking too long so that readiness or liveness probe fails beyond failure threshold. In this case, adjust probe settings to allow longer time to initialize the container, or try a bigger VM SKU among supported VM SKUs which will accelerate the initialization.
There's an error in the environment set up of the container, such as a missing dependency.
When you face TypeError: register() takes 3 positional arguments but 4 were given error, the error may be caused by the dependency between flask v2 and azureml-inference-server-http. See FAQs for inference HTTP server for more details.
ERROR: ResourceNotFound
This is a list of reasons you might run into this error only when using either managed online endpoint or Kubernetes online endpoint:
Azure Resource Manager cannot find a required resource
Azure Container Registry is private or otherwise inaccessible
Resource Manager cannot find a resource
This error occurs when Azure Resource Manager can't find a required resource. For example, you can receive this error if a storage account was referred to but can't be found at the path on which it was specified. Be sure to double check resources that might have been supplied by exact path or the spelling of their names.
For more information, see Resolve Resource Not Found Errors.
Container registry authorization error
This error occurs when an image belonging to a private or otherwise inaccessible container registry was supplied for deployment.
At this time, our APIs cannot accept private registry credentials.
To mitigate this error, either ensure that the container registry is not private or follow the following steps:
Grant your private registry's acrPull role to the system identity of your online endpoint.
In your environment definition, specify the address of your private image as well as the additional instruction to not modify (build) the image.
If the mitigation is successful, the image will not require any building and the final image address will simply be the given image address.
At deployment time, your online endpoint's system identity will pull the image from the private registry.
For more diagnostic information, see How To Use the Workspace Diagnostic API.
ERROR: OperationCanceled
This is a list of reasons you might run into this error when using either managed online endpoint or Kubernetes online endpoint:
Operation was canceled by another operation that has a higher priority
Operation was canceled due to a previous operation waiting for lock confirmation
Operation canceled by another higher priority operation
Azure operations have a certain priority level and are executed from highest to lowest. This error happens when your operation happened to be overridden by another operation that has a higher priority.
Retrying the operation might allow it to be performed without cancellation.
Operation canceled waiting for lock confirmation
Azure operations have a brief waiting period after being submitted during which they retrieve a lock to ensure that we don't run into race conditions. This error happens when the operation you submitted is the same as another operation that is currently still waiting for confirmation that it has received the lock to proceed. It may indicate that you've submitted a very similar request too soon after the initial request.
Retrying the operation after waiting several seconds up to a minute may allow it to be performed without cancellation.
ERROR: InternalServerError
Although we do our best to provide a stable and reliable service, sometimes things don't go according to plan. If you get this error, it means that something isn't right on our side, and we need to fix it. Submit a customer support ticket with all related information and we can address the issue.
Common errors specific to Kubernetes deployments
ACRSecretError
ImagePullLoopBackOff
DeploymentCrashLoopBackOff
KubernetesCrashLoopBackOff
NamespaceNotFound
UserScriptInitFailed
UserScriptImportError
UserScriptFunctionNotFound
EndpointAlreadyExists
ScoringFeUnhealthy
ValidateScoringFailed
InvalidDeploymentSpec
PodUnschedulable
PodOutOfMemory
InferencingClientCallFailed
ERROR: ACRSecretError
This is a list of reasons you might run into this error when creating/updating the Kubernetes online deployments:
Role assignment has not yet been completed. In this case, please wait for a few seconds and try again later.
The Azure ARC (For Azure Arc Kubernetes cluster) or Azure Machine Learning extension (For AKS) is not properly installed or configured. Please try to check the Azure ARC or Azure Machine Learning extension configuration and status.
The Kubernetes cluster has improper network configuration, please check the proxy, network policy or certificate.
If you are using a private AKS cluster, it is necessary to set up private endpoints for ACR, storage account, workspace in the AKS vnet.
ERROR: ImagePullLoopBackOff
The reason you might run into this error when creating/updating Kubernetes online deployments is because you can't download the images from the container registry, resulting in the images pull failure.
In this case, you can check the cluster network policy and the workspace container registry if cluster can pull image from the container registry.
ERROR: DeploymentCrashLoopBackOff
The reason you might run into this error when creating/updating Kubernetes online deployments is the user container crashed initializing. There are two possible reasons for this error:
User script score.py has syntax error or import error then raise exceptions in initializing.
Or the deployment pod needs more memory than its limit.
To mitigate this error, first you can check the deployment logs for any exceptions in user scripts. If error persists, try to extend resources/instance type memory limit.
ERROR: KubernetesCrashLoopBackOff
This is a list of reasons you might run into this error when creating/updating the Kubernetes online endpoints/deployments:




    

One or more pod(s) stuck in CrashLoopBackoff status, you can check if the deployment log exists, and check if there are error messages in the log.
There is an error in score.py and the container crashed when init your score code, you can follow ERROR: ResourceNotReady part.
Your scoring process needs more memory that your deployment config limit is insufficient, you can try to update the deployment with a larger memory limit.
ERROR: NamespaceNotFound
The reason you might run into this error when creating/updating the Kubernetes online endpoints is because the namespace your Kubernetes compute used is unavailable in your cluster.
You can check the Kubernetes compute in your workspace portal and check the namespace in your Kubernetes cluster. If the namespace is not available, you can detach the legacy compute and reattach to create a new one, specifying a namespace that already exists in your cluster.
ERROR: UserScriptInitFailed
The reason you might run into this error when creating/updating the Kubernetes online deployments is because the init function in your uploaded score.py file raised exception.
You can check the deployment logs to see the exception message in detail and fix the exception.
ERROR: UserScriptImportError
The reason you might run into this error when creating/updating the Kubernetes online deployments is because the score.py file you uploaded has imported unavailable packages.
You can check the deployment logs to see the exception message in detail and fix the exception.
ERROR: UserScriptFunctionNotFound
The reason you might run into this error when creating/updating the Kubernetes online deployments is because the score.py file you uploaded does not have a function named init() or run(). You can check your code and add the function.
ERROR: EndpointNotFound
The reason you might run into this error when creating/updating Kubernetes online deployments is because the system can't find the endpoint resource for the deployment in the cluster. You should create the deployment in an exist endpoint or create this endpoint first in your cluster.
ERROR: EndpointAlreadyExists
The reason you might run into this error when creating a Kubernetes online endpoint is because the creating endpoint already exists in your cluster.
The endpoint name should be unique per workspace and per cluster, so in this case, you should create endpoint with another name.
ERROR: ScoringFeUnhealthy
The reason you might run into this error when creating/updating a Kubernetes online endpoint/deployment is because the Azureml-fe that is the system service running in the cluster is not found or unhealthy.
To trouble shoot this issue, you can reinstall or update the Azure Machine Learning extension in your cluster.
ERROR: ValidateScoringFailed
The reason you might run into this error when creating/updating Kubernetes online deployments is because the scoring request URL validation failed when processing the model deploying.
In this case, you can first check the endpoint URL and then try to redeploy the deployment.
ERROR: InvalidDeploymentSpec
The reason you might run into this error when creating/updating Kubernetes online deployments is because the deployment spec is invalid.
In this case, you can check the error message.
Make sure the instance count is valid.
If you have enabled auto scaling, make sure the minimum instance count and maximum instance count are both valid.
ERROR: PodUnschedulable
This is a list of reasons you might run into this error when creating/updating the Kubernetes online endpoints/deployments:
Unable to schedule pod to nodes, due to insufficient resources in your cluster.
No node match node affinity/selector.
To mitigate this error, you can follow these steps:
Check the node selector definition of the instance type you used, and node label configuration of your cluster nodes.
Check instance type and the node SKU size for AKS cluster or the node resource for Arc-Kubernetes cluster.
If the cluster is under-resourced, you can reduce the instance type resource requirement or use another instance type with smaller resource required.
If the cluster has no more resource to meet the requirement of the deployment, delete some deployment to release resources.
ERROR: PodOutOfMemory
The reason you might run into this error when you creating/updating online deployment is the memory limit you give for deployment is insufficient. You can set the memory limit to a larger value or use a bigger instance type to mitigate this error.
ERROR: InferencingClientCallFailed
The reason you might run into this error when creating/updating Kubernetes online endpoints/deployments is because the k8s-extension of the Kubernetes cluster is not connectable.
In this case, you can detach and then re-attach your compute.
To troubleshoot errors by reattaching, please guarantee to reattach with the exact same configuration as previously detached compute, such as the same compute name and namespace, otherwise you may encounter other errors.
If it is still not working, you can ask the administrator who can access the cluster to use kubectl get po -n azureml to check whether the relay server pods are running.
Autoscaling issues
If you're having trouble with autoscaling, see Troubleshooting Azure autoscale.
For Kubernetes online endpoint, there is Azure Machine Learning inference router which is a front-end component to handle autoscaling for all model deployments on the Kubernetes cluster, you can find more information in Autoscaling of Kubernetes inference routing
Common model consumption errors
This is a list of common model consumption errors resulting from the endpoint invoke operation status.
Bandwidth limit issues
HTTP status codes
Blocked by CORS policy
Bandwidth limit issues
Managed online endpoints have bandwidth limits for each endpoint. You find the limit configuration in Manage and increase quotas for resources with Azure Machine Learning. If your bandwidth usage exceeds the limit, your request will be delayed. To monitor the bandwidth delay:
Use metric “Network bytes” to understand the current bandwidth usage. For more information, see Monitor managed online endpoints.
There are two response trailers will be returned if the bandwidth limit enforced:
ms-azureml-bandwidth-request-delay-ms: delay time in milliseconds it took for the request stream transfer.
ms-azureml-bandwidth-response-delay-ms: delay time in milliseconds it took for the response stream transfer.
HTTP status codes
When you access online endpoints with REST requests, the returned status codes adhere to the standards for HTTP status codes. These are details about how endpoint invocation and prediction errors map to HTTP status codes.
Common error codes for managed online endpoints
These are common error codes when consuming managed online endpoints with REST requests:
Status code
Reason phrase
Why this code might get returned
Unauthorized
You don't have permission to do the requested action, such as score, or your token is expired.
Not found
The endpoint doesn't have any valid deployment with positive weight.
Request timeout
The model execution took longer than the timeout supplied in request_timeout_ms under request_settings of your model deployment config.
Model Error
If your model container returns a non-200 response, Azure returns a 424. Check the Model Status Code dimension under the Requests Per Minute metric on your endpoint's Azure Monitor Metric Explorer. Or check response headers ms-azureml-model-error-statuscode and ms-azureml-model-error-reason for more information. If 424 comes with liveness or readiness probe failing, consider adjusting probe settings to allow longer time to probe liveness or readiness of the container.
Too many pending requests
Your model is currently getting more requests than it can handle. Azure Machine Learning has implemented a system that permits a maximum of 2 * max_concurrent_requests_per_instance * instance_count requests to be processed in parallel at any given moment to guarantee smooth operation. Additional requests that exceed this maximum will be rejected. You can review your model deployment configuration under the request_settings and scale_settings sections to verify and adjust these settings. Additionally, as outlined in the YAML definition for RequestSettings, it is important to ensure that the environment variable WORKER_COUNT is correctly passed. 

 If you're using auto-scaling and get this error, it means your model is getting requests quicker than the system can scale up. In this situation, consider resending requests with an exponential backoff to give the system the time it needs to adjust. You could also increase the number of instances by using code to calculate instance count. These steps, combined with setting auto-scaling, will help ensure that your model is ready to handle the influx of requests.
Rate-limiting
The number of requests per second reached the limit of managed online endpoints.
Internal server error
Azure Machine Learning-provisioned infrastructure is failing.
Common error codes for kubernetes online endpoints
These are common error codes when consuming Kubernetes online endpoints with REST requests:
Status code
Reason phrase
Why this code might get returned
Conflict error
When an operation is already in progress, any new operation on that same online endpoint responds with 409 conflict error. For example, If create or update online endpoint operation is in progress and if you trigger a new Delete operation it throws an error.
Has thrown an exception or crashed in the run() method of the score.py file
When there's an error in score.py, for example an imported package does not exist in the conda environment, a syntax error, or a failure in the init() method. You can follow here to debug the file.
Receive large spikes in requests per second
The autoscaler is designed to handle gradual changes in load. If you receive large spikes in requests per second, clients may receive an HTTP status code 503. Even though the autoscaler reacts quickly, it takes AKS a significant amount of time to create more containers. You can follow here to prevent 503 status codes.
Request has timed out
A 504 status code indicates that the request has timed out. The default timeout setting is 5 seconds. You can increase the timeout or try to speed up the endpoint by modifying the score.py to remove unnecessary calls. If these actions don't correct the problem, you can follow here to debug the score.py file. The code may be in a non-responsive state or an infinite loop.
Internal server error
Azure Machine Learning-provisioned infrastructure is failing.
How to prevent 503 status codes
Kubernetes online deployments support autoscaling, which allows replicas to be added to support extra load, more information you can find in Azure Machine Learning inference router. Decisions to scale up/down is based off of utilization of the current container replicas.
There are two things that can help prevent 503 status codes:
These two approaches can be used individually or in combination.
Change the utilization level at which autoscaling creates new replicas. You can adjust the utilization target by setting the autoscale_target_utilization to a lower value.
Important
This change does not cause replicas to be created faster. Instead, they are created at a lower utilization threshold. Instead of waiting until the service is 70% utilized, changing the value to 30% causes replicas to be created when 30% utilization occurs.
If the Kubernetes online endpoint is already using the current max replicas and you're still seeing 503 status codes, increase the autoscale_max_replicas value to increase the maximum number of replicas.
Change the minimum number of replicas. Increasing the minimum replicas provides a larger pool to handle the incoming spikes.
To increase the number of instances, you could calculate the required replicas following these codes.
from math import ceil
# target requests per second
target_rps = 20
# time to process the request (in seconds, choose appropriate percentile)
request_process_time = 10
# Maximum concurrent requests per instance
max_concurrent_requests_per_instance = 1
# The target CPU usage of the model container. 70% in this example
target_utilization = .7
concurrent_requests = target_rps * request_process_time / target_utilization
# Number of instance count
instance_count = ceil(concurrent_requests / max_concurrent_requests_per_instance)
If you receive request spikes larger than the new minimum replicas can handle, you may receive 503 again. For example, as traffic to your endpoint increases, you may need to increase the minimum replicas.
How to calculate instance count
To increase the number of instances, you can calculate the required replicas by using the following code:
from math import ceil
# target requests per second
target_rps = 20
# time to process the request (in seconds, choose appropriate percentile)
request_process_time = 10
# Maximum concurrent requests per instance
max_concurrent_requests_per_instance = 1
# The target CPU usage of the model container. 70% in this example
target_utilization = .7
concurrent_requests = target_rps * request_process_time / target_utilization
# Number of instance count
instance_count = ceil(concurrent_requests / max_concurrent_requests_per_instance)
Blocked by CORS policy
Online endpoints (v2) currently do not support Cross-Origin Resource Sharing (CORS) natively. If your web application tries to invoke the endpoint without proper handling of the CORS preflight requests, you can see the following error message:
Access to fetch at 'https://{your-endpoint-name}.{your-region}.inference.ml.azure.com/score' from origin http://{your-url} has been blocked by CORS policy: Response to preflight request doesn't pass access control check. No 'Access-control-allow-origin' header is present on the request resource. If an opaque response serves your needs, set the request's mode to 'no-cors' to fetch the resource with the CORS disabled.
We recommend that you use Azure Functions, Azure Application Gateway, or any service as an interim layer to handle CORS preflight requests.
Common network isolation issues
Online endpoint creation fails with a V1LegacyMode == true message
The Azure Machine Learning workspace can be configured for v1_legacy_mode, which disables v2 APIs. Managed online endpoints are a feature of the v2 API platform, and won't work if v1_legacy_mode is enabled for the workspace.
Important
Check with your network security team before disabling v1_legacy_mode. It may have been enabled by your network security team for a reason.
For information on how to disable v1_legacy_mode, see Network isolation with v2.
Online endpoint creation with key-based authentication fails
Use the following command to list the network rules of the Azure Key Vault for your workspace. Replace <keyvault-name> with the name of your key vault:
az keyvault network-rule list -n <keyvault-name>
The response for this command is similar to the following JSON document:
    "bypass": "AzureServices",
    "defaultAction": "Deny",
    "ipRules": [],
    "virtualNetworkRules": []
If the value of bypass isn't AzureServices, use the guidance in the Configure key vault network settings to set it to AzureServices.
Online deployments fail with an image download error
Check if the egress-public-network-access flag is disabled for the deployment. If this flag is enabled, and the visibility of the container registry is private, then this failure is expected.
Use the following command to check the status of the private endpoint connection. Replace <registry-name> with the name of the Azure Container Registry for your workspace:
az acr private-endpoint-connection list -r <registry-name> --query "[?privateLinkServiceConnectionState.description=='Egress for Microsoft.MachineLearningServices/workspaces/onlineEndpoints'].{Name:name, status:privateLinkServiceConnectionState.status}"
In the response document, verify that the status field is set to Approved. If it isn't approved, use the following command to approve it. Replace <private-endpoint-name> with the name returned from the previous command:
az network private-endpoint-connection approve -n <private-endpoint-name>
Scoring endpoint can't be resolved
Verify that the client issuing the scoring request is a virtual network that can access the Azure Machine Learning workspace.
Use the nslookup command on the endpoint hostname to retrieve the IP address information:
nslookup endpointname.westcentralus.inference.ml.azure.com
The response contains an address. This address should be in the range provided by the virtual network
For Kubernetes online endpoint, the endpoint hostname should be the CName (domain name) which has been specified in your Kubernetes cluster.
If it is an HTTP endpoint, the IP address will be contained in the endpoint URI which you can get directly in the Studio UI.
More ways to get the IP address of the endpoint can be found in Secure Kubernetes online endpoint.
If the host name isn't resolved by the nslookup command:
For Managed online endpoint,
Check if an A record exists in the private DNS zone for the virtual network.
To check the records, use the following command:
az network private-dns record-set list -z privatelink.api.azureml.ms -o tsv --query [].name
The results should contain an entry that is similar to *.<GUID>.inference.<region>.
If no inference value is returned, delete the private endpoint for the workspace and then recreate it. For more information, see How to configure a private endpoint.
If the workspace with a private endpoint is setup using a custom DNS How to use your workspace with a custom DNS server, use following command to verify if resolution works correctly from custom DNS.
dig endpointname.westcentralus.inference.ml.azure.com
For Kubernetes online endpoint,
Check the DNS configuration in Kubernetes cluster.
Additionally, you can check if the azureml-fe works as expected, use the following command:
kubectl exec -it deploy/azureml-fe -- /bin/bash
(Run in azureml-fe pod)
curl -vi -k https://localhost:<port>/api/v1/endpoint/<endpoint-name>/swagger.json
"Swagger not found"
For HTTP, use
curl https://localhost:<port>/api/v1/endpoint/<endpoint-name>/swagger.json
"Swagger not found"
If curl HTTPs fails (e.g. timeout) but HTTP works, please check that certificate is valid.
If this fails to resolve to A record, verify if the resolution works from Azure DNS(168.63.129.16).
dig @168.63.129.16 endpointname.westcentralus.inference.ml.azure.com
If this succeeds then you can troubleshoot conditional forwarder for private link on custom DNS.
Online deployments can't be scored
Use the following command to see if the deployment was successfully deployed:
az ml online-deployment show -e <endpointname> -n <deploymentname> --query '{name:name,state:provisioning_state}' 
If the deployment completed successfully, the value of state will be Succeeded.
If the deployment was successful, use the following command to check that traffic is assigned to the deployment. Replace <endpointname> with the name of your endpoint:
az ml online-endpoint show -n <endpointname>  --query traffic
This step isn't needed if you are using the azureml-model-deployment header in your request to target this deployment.
The response from this command should list percentage of traffic assigned to deployments.
If the traffic assignments (or deployment header) are set correctly, use the following command to get the logs for the endpoint. Replace <endpointname> with the name of the endpoint, and <deploymentname> with the deployment:
az ml online-deployment get-logs  -e <endpointname> -n <deploymentname> 
Look through the logs to see if there's a problem running the scoring code when you submit a request to the deployment.
Troubleshoot inference server
In this section, we provide basic troubleshooting tips for Azure Machine Learning inference HTTP server.
Basic steps
The basic steps for troubleshooting are:
Gather version information for your Python environment.
Make sure the azureml-inference-server-http python package version that specified in the environment file matches the AzureML Inferencing HTTP server version that displayed in the startup log. Sometimes pip's dependency resolver leads to unexpected versions of packages installed.
If you specify Flask (and or its dependencies) in your environment, remove them. The dependencies include Flask, Jinja2, itsdangerous, Werkzeug, MarkupSafe, and click. Flask is listed as a dependency in the server package and it's best to let our server install it. This way when the server supports new versions of Flask, you'll automatically get them.
Server version
The server package azureml-inference-server-http is published to PyPI. You can find our changelog and all previous versions on our PyPI page. Update to the latest version if you're using an earlier version.
0.4.x: The version that is bundled in training images ≤ 20220601 and in azureml-defaults>=1.34,<=1.43. 0.4.13 is the last stable version. If you use the server before version 0.4.11, you may see Flask dependency issues like can't import name Markup from jinja2. You're recommended to upgrade to 0.4.13 or 0.8.x (the latest version), if possible.
0.6.x: The version that is preinstalled in inferencing images ≤ 20220516. The latest stable version is 0.6.1.
0.7.x: The first version that supports Flask 2. The latest stable version is 0.7.7.
0.8.x: The log format has changed and Python 3.6 support has dropped.
Package dependencies
The most relevant packages for the server azureml-inference-server-http are following packages:
flask
opencensus-ext-azure
inference-schema
If you specified azureml-defaults in your Python environment, the azureml-inference-server-http package is depended on, and will be installed automatically.
If you're using Python SDK v1 and don't explicitly specify azureml-defaults in your Python environment, the SDK may add the package for you. However, it will lock it to the version the SDK is on. For example, if the SDK version is 1.38.0, it will add azureml-defaults==1.38.0 to the environment's pip requirements.
Frequently asked questions
1. I encountered the following error during server startup:
TypeError: register() takes 3 positional arguments but 4 were given
  File "/var/azureml-server/aml_blueprint.py", line 251, in register
    super(AMLBlueprint, self).register(app, options, first_registration)
TypeError: register() takes 3 positional arguments but 4 were given
You have Flask 2 installed in your python environment but are running a version of azureml-inference-server-http that doesn't support Flask 2. Support for Flask 2 is added in azureml-inference-server-http>=0.7.0, which is also in azureml-defaults>=1.44.
If you're not using this package in an AzureML docker image, use the latest version of
azureml-inference-server-http or azureml-defaults.
If you're using this package with an AzureML docker image, make sure you're using an image built in or after July,
2022. The image version is available in the container logs. You should be able to find a log similar to the following:
2022-08-22T17:05:02,147738763+00:00 | gunicorn/run | AzureML Container Runtime Information
2022-08-22T17:05:02,161963207+00:00 | gunicorn/run | ###############################################
2022-08-22T17:05:02,168970479+00:00 | gunicorn/run | 
2022-08-22T17:05:02,174364834+00:00 | gunicorn/run | 
2022-08-22T17:05:02,187280665+00:00 | gunicorn/run | AzureML image information: openmpi4.1.0-ubuntu20.04, Materializaton Build:20220708.v2
2022-08-22T17:05:02,188930082+00:00 | gunicorn/run | 
2022-08-22T17:05:02,190557998+00:00 | gunicorn/run | 
The build date of the image appears after "Materialization Build", which in the above example is 20220708, or July 8, 2022. This image is compatible with Flask 2. If you don't see a banner like this in your container log, your image is out-of-date, and should be updated. If you're using a CUDA image, and are unable to find a newer image, check if your image is deprecated in AzureML-Containers. If it's, you should be able to find replacements.
If you're using the server with an online endpoint, you can also find the logs under "Deployment logs" in the online endpoint page in Azure Machine Learning studio. If you deploy with SDK v1 and don't explicitly specify an image in your deployment configuration, it will default to using a version of openmpi4.1.0-ubuntu20.04 that matches your local SDK toolset, which may not be the latest version of the image. For example, SDK 1.43 will default to using openmpi4.1.0-ubuntu20.04:20220616, which is incompatible. Make sure you use the latest SDK for your deployment.
If for some reason you're unable to update the image, you can temporarily avoid the issue by pinning azureml-defaults==1.43 or azureml-inference-server-http~=0.4.13, which will install the older version server with Flask 1.0.x.
2. I encountered an ImportError or ModuleNotFoundError on modules opencensus, jinja2, MarkupSafe, or click during startup like the following message:
ImportError: cannot import name 'Markup' from 'jinja2'
Older versions (<= 0.4.10) of the server didn't pin Flask's dependency to compatible versions. This problem is fixed in the latest version of the server.
Next steps
Deploy and score a machine learning model by using an online endpoint
Safe rollout for online endpoints
Online endpoint YAML reference
Troubleshoot kubernetes compute