Model persists when endpoint deploys fails at 'createEndpoint'

[evakravi]

I'm a software engineer at AWS that is experiencing this bug in SageMaker Studio (JumpStart). My alias is @evakravi

Describe the bug
If you attempt to deploy a model and are successful in a) uploading the script/deps to s3, (b) creating a model, (c) creating an endpoint config, but d) FAIL to create the endpoint (which could happen if an account does not have GPU instances allocated), then if you attempt to re-deploy the same model with parameters such that the endpoint creation succeeds, the deployment will still fail.

To reproduce
Deploy a model on a disallowed instance type, then deploy the identical model to a permitted instance type.

Expected behavior
The deployment to an allowed instance type should be successful, but it ends up failing.

System information
A description of your system. Please provide:

• SageMaker Python SDK version: 2.21.0
• Python version: Python 3.7.10
• CPU or GPU: SageMaker Studio Jupyter Server
• Custom Docker image (Y/N): SageMaker Studio Jupyter Server

Additional context
While this issue is related to SageMaker JumpStart ModelHub, the sagemaker api issue can be reproduced outside of SageMaker Studio.

[ChoiByungWook]

What is the exact error/stacktrace are you seeing when you deploy for the second time?

From my understanding it sounds like you're running a similar issue shown in: #1470?

[ChoiByungWook]

Something went wrong
We encountered an error while preparing to deploy your endpoint. You can get more details below.
operation deploy failed: An error occurred (ResourceLimitExceeded) when calling the CreateEndpoint operation: The account-level service limit 'ml.g4dn.xlarge for endpoint usage' is 0 Instances, with current utilization of 0 Instances and a request delta of 1 Instances. Please contact AWS support to request an increase for this limit.

[ChoiByungWook]

It looks like we check to see if there is an existing endpoint_config:

sagemaker-python-sdk/src/sagemaker/session.py

Line 3443 in 3a1f0b3

if not _deployment_entity_exists(

Are you using the same name in your code for your deployments?

[marckarp]

There is not enough in the information provided to reproduce/classify this as a bug. Do you have a code snippets/notebook to reproduce the issue?