Issue to deploy a model and create the endpoint keys

Hello Tasmia Tun Nazmun,

Welcome to Microsoft Q&A .Thank you for reaching out.

The issue described commonly occurs during the first few deployments in Azure Machine Learning, especially when using Designer and a free or trial subscription. When endpoint creation or key generation fails multiple times, it usually indicates a configuration, permission, or quota limitation, rather than a temporary system issue. Repeating the same action without changes typically results in the same outcome.

The most common and practical root causes for endpoint deployment or key creation to fail are listed below.

1. Insufficient permissions (RBAC) - Endpoint creation and key retrieval require write-level access on the Machine Learning workspace. If the assigned role does not allow deployment actions, the operation is blocked and fails consistently across retries.
2. Free or trial subscription limits - Free subscriptions have limited credits and compute quotas. When available quota is exhausted or a required compute SKU is unavailable in the selected region, endpoint deployment cannot proceed. Retrying does not resolve quota-based failures.
3. Compute not available or not suitable for inference - Designer deployments require an inference-capable compute. If no valid compute is available, stopped, or unsupported for real-time inference, deployment fails deterministically.
4. Region capacity or availability constraints - Some regions may not have sufficient capacity or supported compute SKUs at the time of deployment. This results in repeated provisioning failures until the region or compute selection changes.
5. Deployment configuration issues - Failures can occur if the deployment configuration cannot complete container build or startup. In such cases, endpoint keys are not generated because the endpoint never reaches a healthy state.
6. Endpoint not successfully deployed - Endpoint keys are created only after a deployment succeeds. If deployment status is Failed, keys are not available, and key-related errors are a downstream symptom rather than the root cause.

When the same operation fails two or more times with repeated failures, it often indicates that

• The issue is almost always deterministic
• Retrying without changes does not resolve:
  • Permission issues
  • Quota exhaustion
  • Missing or invalid compute
  • Deployment configuration problems
• Resolution requires a change in configuration, quota, permissions or region

Please check if the below troubleshooting steps help:

Please review the steps below in order:

1. Check deployment status and logs
   • Open Azure Machine Learning Studio
   • Navigate to Endpoints
   • Select the endpoint and review Deployment logs
   • Logs provide the exact reason for failure (quota, permission, compute, or configuration)
2. Verify available quota and credits
   • Review remaining subscription credits
   • Check compute quota for the selected region
   • Confirm required VM families are available
3. Confirm compute readiness
   • Ensure an inference-capable compute is selected
   • Confirm compute is running and supported for real-time endpoints
4. Verify permissions
   • Confirm the assigned role allows endpoint creation and management
   • Deployment actions require contributor-level access or equivalent
5. Access endpoint keys only after successful deployment
   • Keys become available only when deployment status is Succeeded
   • Keys are visible under Endpoints >  Keys & tokens

References:

• Troubleshoot online endpoint deployment - Azure Machine Learning | Microsoft Learn
• Online endpoints for real-time inference - Azure Machine Learning | Microsoft Learn
• Manage roles in your workspace - Azure Machine Learning | Microsoft Learn
• Authenticate Clients for Online Endpoints - Azure Machine Learning | Microsoft Learn
• Authentication and authorization for online endpoints - Azure Machine Learning | Microsoft Learn
• Manage resources and quotas - Azure Machine Learning | Microsoft Learn

Thank you!

Please 'Upvote'(Thumbs-up) and 'Accept' as answer if the reply was helpful. This will be benefitting other community members who face the same issue.