Use endpoints for inference

APPLIES TO: Azure CLI ml extension v2 (current) Python SDK azure-ai-ml v2 (current)

After you train a machine learning model or a machine learning pipeline, you need to deploy them so others can consume their predictions. Such execution mode is called inference. Azure Machine Learning uses the concept of endpoints and deployments for machine learning inference.

Endpoints and deployments are two constructs that allow you to decouple the interface of your production workload from the implementation that serves it.


Let's imagine you are working on an application that needs to predict the type and color of a car given its photo. The application only needs to know that they make an HTTP request to a URL using some sort of credentials, provide a picture of a car, and they get the type and color of the car back as string values. This thing we have just described is an endpoint.

A diagram showing the concept of an endpoint.

Now, let's imagine that one data scientist, Alice, is working on its implementation. Alice is well versed on TensorFlow so she decided to implement the model using a Keras sequential classifier with a RestNet architecture she consumed from TensorFlow Hub. She tested the model and she is happy with the results. She decides to use that model to solve the car prediction problem. Her model is large in size, it would require 8GB of memory with 4 cores to run it. This thing we have just described is a deployment.

A diagram showing the concept of a deployment.

Finally, let's imagine that after running for a couple of months, the organization discovers that the application performs poorly on images with no ideal illumination conditions. Bob, another data scientist, knows a lot about data argumentation techniques that can be used to help the model build robustness on that factor. However, he feels more comfortable using Torch rather than TensorFlow. He trained another model then using those techniques and he's happy with the results. He would like to try this model on production gradually until the organization is ready to retire the old one. His model shows better performance when deployed to GPU, so he needs one to the deployment. We have just described another deployment under the same endpoint.

A diagram showing the concept of an endpoint with multiple deployments.

Endpoints and deployments

An endpoint, is a stable and durable URL that can be used to request or invoke the model, provide the required inputs, and get the outputs back. An endpoint provides:

  • a stable and durable URL (like
  • An authentication and authorization mechanism.

A deployment is a set of resources required for hosting the model or component that does the actual inferencing. A single endpoint can contain multiple deployments which can host independent assets and consume different resources based on what the actual assets require. Endpoints have a routing mechanism that can route the request generated for the clients to specific deployments under the endpoint.

To function properly, each endpoint needs to have at least one deployment. Endpoints and deployments are independent Azure Resource Manager resources that appear in the Azure portal.

Online and batch endpoints

Azure Machine Learning allows you to implement online endpoints and batch endpoints. Online endpoints are designed for real-time inference so the results are returned in the response of the invocation. Batch endpoints, on the other hand, are designed for long-running batch inference so each time you invoke the endpoint you generate a batch job that performs the actual work.

When to use what

Use online endpoints to operationalize models for real-time inference in synchronous low-latency requests. We recommend using them when:

  • You have low-latency requirements.
  • Your model can answer the request in a relatively short amount of time.
  • Your model's inputs fit on the HTTP payload of the request.
  • You need to scale up in term of number of request.

Use batch endpoints to operationalize models or pipelines (preview) for long-running asynchronous inference. We recommend using them when:

  • You have expensive models or pipelines that require a longer time to run.
  • You want to operationalize machine learning pipelines and reuse components.
  • You need to perform inference over large amounts of data, distributed in multiple files.
  • You don't have low latency requirements.
  • Your model's inputs are stored in a Storage Account or in an Azure Machine Learning data asset.
  • You can take advantage of parallelization.


Both online and batch endpoints are based on the idea of endpoints and deployments, which help you transition easily from one to the other. However, when moving from one to another, there are some differences that are important to take into account. Some of these differences are due to the nature of the work:


The following table shows a summary of the different features in Online and Batch endpoints.

Feature Online Endpoints Batch endpoints
Stable invocation URL Yes Yes
Multiple deployments support Yes Yes
Deployment's routing Traffic split Switch to default
Mirror traffic to all deployment Yes No
Swagger support Yes No
Authentication Key and token Azure AD
Private network support Yes Yes
Managed network isolation1 Yes No
Customer-managed keys Yes No

1 Managed network isolation allows managing the networking configuration of the endpoint independently from the Azure Machine Learning workspace configuration.


The following table shows a summary of the different features in Online and Batch endpoints at the deployment level. These concepts apply per each deployment under the endpoint.

Feature Online Endpoints Batch endpoints
Deployment's types Models Models and Pipeline components (preview)
MLflow model's deployment Yes (requires public networking) Yes
Custom model's deployment Yes, with scoring script Yes, with scoring script
Inference server 1 - Azure Machine Learning Inferencing Server
- Triton
- Custom (using BYOC)
Batch Inference
Compute resource consumed Instances or granular resources Cluster instances
Compute type Managed compute and Kubernetes Managed compute and Kubernetes
Low-priority compute No Yes
Scales compute to zero No Yes
Autoscale compute2 Yes, based on resources' load Yes, based on jobs count
Overcapacity management Throttling Queuing
Test deployments locally Yes No

1 Inference server makes reference to the serving technology that takes request, process them, and creates a response. The inference server also dictates the format of the input and the expected outputs.

2 Autoscale makes reference to the ability of dynamically scaling up or down the deployment's allocated resources based on its load. Online and Batch Deployments use different strategies. While online deployments scale up and down based on the resource utilization (like CPU, memory, requests, etc.), batch endpoints scale up or down based on the number of jobs created.

Developer interfaces

Endpoints are designed to help organization operationalize production level workloads in Azure Machine Learning. They're robust, and scalable resources and they provide the best of the capabilities to implement MLOps workflows.

Create and manage batch and online endpoints with multiple developer tools:

  • The Azure CLI and the Python SDK
  • Azure Resource Manager/REST API
  • Azure Machine Learning studio web portal
  • Azure portal (IT/Admin)
  • Support for CI/CD MLOps pipelines using the Azure CLI interface & REST/ARM interfaces

Next steps