How to create and manage session pools for fast web service connections in Python

Applies to: Machine Learning Server 9.3 (Python)

Fast connections to a web service are possible when you create sessions and load dependencies in advance. Sessions are available in a pool dedicated to a specific web service, where each session includes an instance of the Python interpreter and a copy of dependencies required by the web service.

For example, if you create ten sessions in advance for a web service that uses numpy, pandas, scikit, revoscalepy, microsoftml, and azureml-model-management-sdk, each session would have its own instance of the Python interpreter plus a copy of each module loaded into memory.

A web service having a dedicated session pool never requests connections from the generic session pool shared resource, not even when maximum sessions are reached. The generic session pool services only those web services that do not have dedicated resources.

For Python script, the MLServer class in the azureml-model-management-sdk function library provides three functions for creating and managing sessions:

Create or modify a dedicated session pool

You can use a Python interactive window to run the following commands on the local server if you configured Machine Learning Server for one-box, or on a compute node if you have a distributed topology.

![Tip] To create and deploy a web service for testing,

 # load modules and classes
 from azureml.deploy import DeployClient
 from azureml.deploy.server import MLServer

 # Set up the connection
 # Be sure to replace the password placeholder
host = "http://localhost:12800"
ctx = ("admin", "password-placeholder")
client = DeployClient(host, use=MLServer, auth=ctx)

 # Return a list of web services to get the service and version 
 # Both service name and version number are required
 svc = client.list_services()

 # Create the session pool using a case-sensitive web service name
 # A status code of 200 is returned upon success
 svc = client.create_or_update_service_pool(name = "myWebservice1234", version = "v1.0.0", initial_pool_size = 5, max_pool_size = 10 )

 # Return status 
 # Pending indicates session creation is in progress. Success indicates sessions are ready.
 svc = client.get_service_pool_status(name = "myWebService1234", version = "v1.0.0")

Currently, there are no commands or functions that return actual session pool usage. The log file is your best resource for analyzing connection and service activity. For more information, see Trace user actions.

Delete a session pool

On the compute node, run the following command to delete the session pool for a given service.

 # Return a list of web services to get the service and version information
 svc = client.list_services()

 # Deletes the dedicated session pool and releases resources
 svc = client.delete_service_pool(name = "myWebService1234", version = "v1.0.0")

This feature is still under development. In rare cases, the delete_service_pool command may fail to actually delete the pool on the computeNode. If you encounter this situation, issue a new delete_service_pool request. Use the get_service_pool_status command to monitor the status of dedicated pools on the computeNode.

# Deletes the dedicated session pool and releases resources
client.delete_service_ool(name = "myWebService1234", version = "v1.0.0")

# Check the realtime status of dedicated pool
client.get_service_pool_status(name = "myWebService1234", version = "v1.0.0")

# make sure the returned status is NotFound on all computeNodes
# if not, issues another delete_service_pool command again

See also