Optimizing prices for Amazon SageMaker Canvas with automated shutdown of idle apps

Amazon SageMaker Canvas is a wealthy, no-code Machine Studying (ML) and Generative AI workspace that has allowed prospects everywhere in the world to extra simply undertake ML applied sciences to unravel previous and new challenges due to its visible, no-code interface. It does so by protecting the ML workflow end-to-end: whether or not you’re searching for highly effective information preparation and AutoML, managed endpoint deployment, simplified MLOps capabilities, and ready-to-use fashions powered by AWS AI providers and Generative AI, SageMaker Canvas might help you to realize your objectives.

As firms of all sizes undertake SageMaker Canvas, prospects requested for tactics to optimize value. As outlined within the AWS Well-Architected Framework, a cost-optimized workload absolutely makes use of all sources, meets your purposeful necessities, and achieves an end result on the lowest attainable worth level.

In the present day, we’re introducing a brand new solution to additional optimize prices for SageMaker Canvas purposes. SageMaker Canvas now collects Amazon CloudWatch metrics that present perception into app utilization and idleness. Clients can use this info to close down robotically idle SageMaker Canvas purposes to avoiding incurring unintended prices.

On this put up, we’ll present you the way to robotically shut down idle SageMaker Canvas apps to regulate prices through the use of a easy serverless structure. Templates used on this put up are accessible in GitHub.

Understanding and monitoring prices

Training is at all times step one into understanding and controlling prices for any workload, both on-premises or within the cloud. Let’s begin by reviewing the SageMaker Canvas pricing mannequin. In a nutshell, SageMaker Canvas has a pay-as-you-go pricing mannequin, based mostly on two dimensions:

  • Workspace occasion: ­ previously referred to as session time, is the fee related to operating the SageMaker Canvas app
  • AWS service expenses: ­ prices related to coaching the fashions, deploying the endpoints, producing inferences (sources to spin up SageMaker Canvas).

Clients at all times have full management over the sources which might be launched by SageMaker Canvas and might preserve observe of prices related to the SageMaker Canvas app through the use of the AWS Billing and Price Administration service. For extra info, consult with Manage billing and cost in SageMaker Canvas.

To restrict the fee related to the workspace cases, as a finest follow, it’s essential to log off, don’t shut the browser tab. To log off, select the Sign off button on the left panel of the SageMaker Canvas app.

Mechanically shutting down SageMaker Canvas purposes

For IT Directors that wish to present automated controls for shutting down SageMaker Canvas purposes and holding prices underneath management, there are two approaches:

  1. Shutdown purposes on a schedule (day-after-day at 19:00 or each Friday at 18:00)
  2. Shutdown robotically idle purposes (when the appliance hasn’t been used for 2 hours)

Shutdown purposes on a schedule

Canvas Scheduled Shutdown Architecture

Scheduled shutdown of SageMaker Canvas purposes will be achieved with little or no effort through the use of a cron expression (with Amazon EventBridge Cron Rule), a compute part (an AWS Lambda operate) that calls the Amazon SageMaker API DeleteApp. This method has been mentioned within the Provision and manage ML environments with Amazon SageMaker Canvas using AWS CDK and AWS Service Catalog put up, and carried out within the related GitHub repository.

One of many benefits of the above structure is that it is extremely easy to duplicate it to realize scheduled creation of the SageMaker Canvas app. Through the use of a mixture of scheduled creation and scheduled deletion, a cloud administrator can make it possible for the SageMaker Canvas utility is prepared for use each time customers begin their enterprise day (e.g. 9AM on a piece day), and that the app additionally robotically shuts down on the finish of the enterprise day (e.g. 7PM on a piece day, at all times shut down throughout weekends). All that’s wanted to do is change the road of code calling the DeleteApp API into CreateApp, in addition to updating the cron expression to replicate the specified app creation time.

Whereas this method could be very straightforward to implement and check, a disadvantage of the instructed structure is that it doesn’t have in mind whether or not an utility is at present getting used or not, shutting it down no matter its present exercise standing. Based on totally different conditions, this would possibly trigger friction with lively customers, which could all of a sudden see their session terminated.

You may retrieve the template related to this structure from the next GitHub repository:

Shutdown robotically idle purposes

Canvas Shutdown on Idle Architecture

Beginning at this time, Amazon SageMaker Canvas emits CloudWatch metrics that present perception into app utilization and idleness. This enables an administrator to outline an answer that reads the idleness metric, compares it in opposition to a threshold, and defines a selected logic for automated shutdown. A extra detailed overview of the idleness metric emitted by SageMaker Canvas is proven within the following paragraph.

To attain automated shutdown of SageMaker Canvas purposes based mostly on the idleness metrics, we offer an AWS CloudFormation template. This template consists of three fundamental parts:

  1. An Amazon CloudWatch Alarm, which runs a question to test the MAX worth of the TimeSinceLastActive metric. If this worth is bigger than a threshold offered as enter to the CloudFormation template, it triggers the remainder of the automation. This question will be run on a single person profile, on a single area, or throughout all domains. Based on the extent of management that you want, you should use:
    1. the all-domains-all-users template, which checks this throughout all customers and all domains within the area the place the template is deployed
    2. the one-domain-all-users template, which checks this throughout all customers in a single area within the area the place the template is deployed
    3. the one-domain-one-user template, which checks this for one person profile, in a single area, within the area the place the template is deployed
  2. The alarm state change creates an occasion on the default occasion bus in Amazon EventBridge, which has an Amazon EventBridge Rule set as much as set off an AWS Lambda operate
  3. The AWS Lambda operate identifies which SageMaker Canvas app has been operating in idle for greater than the required threshold, and deletes it with the DeleteApp API.

You may retrieve the AWS CloudFormation templates related to this structure from the next GitHub repository:

How SageMaker Canvas idleness metric work

SageMaker Canvas emits a TimeSinceLastActive metric within the /aws/sagemaker/Canvas/AppActivity namespace, which reveals the variety of seconds that the app has been idle with no person exercise. We will use this new metric to set off an automated shutdown of the SageMaker Canvas app when it has been idle for an outlined interval. SageMaker Canvas exposes the TimeSinceLastActive with the next schema:

    "Namespace": "/aws/sagemaker/Canvas/AppActivity",
    "Dimensions": [
    "Metrics": [
            "Name": "TimeSinceLastActive",
            "Unit": "Seconds",
            "Value": 12345

The important thing parts of this metric are as follows:

  • Dimensions, particularly DomainID and UserProfileName, that permit an administrator to pinpoint which purposes are idle throughout all domains and customers
  • Worth of the metric, which signifies the variety of seconds because the final exercise within the SageMaker Canvas purposes. SageMaker Canvas considers the next as exercise:
    • Any motion taken within the SageMaker Canvas utility (clicking a button, reworking a dataset, producing an in-app inference, deploying a mannequin);
    • Utilizing a ready-to-use mannequin or interacting with the Generative AI fashions utilizing chat interface;
    • A batch inference scheduled to run at a selected time; for extra info, consult with  Manage automations.

This metric will be learn by way of Amazon CloudWatch API reminiscent of get_metric_data. For instance, utilizing the AWS SDK for Python (boto3):

import boto3, datetime

cw = boto3.shopper('cloudwatch')
metric_data_results = cw.get_metric_data(
            "Id": "q1",
            "Expression": 'SELECT MAX(TimeSinceLastActive) FROM "/aws/sagemaker/Canvas/AppActivity" GROUP BY DomainId, UserProfileName',
            "Period": 900
    StartTime=datetime.datetime(2023, 1, 1),

The Python question extracts the MAX worth of TimeSinceLastActive from the namespace related to SageMaker Canvas after grouping these values by DomainID and UserProfileName.

Deploying and testing the auto-shutdown resolution

To deploy the auto-shutdown stack, do the next:

  1. Obtain the AWS CloudFormation template that refers back to the resolution you need to implement from the above GitHub repository. Select whether or not you need to implement an answer for all SageMaker Domains, for a single SageMaker Area, or for a single person;
  2. Replace template parameters:
    1. The idle timeout – time (in seconds) that the SageMaker Canvas app is allowed to remain in idle earlier than it will get shutdown; default worth is 2 hours
    2. The alarm interval – aggregation time (in seconds) utilized by CloudWatch Alarm to compute the idle timeout; default worth is 20 minutes
    3. (non-compulsory) SageMaker Area ID and person profile identify
  3. Deploy the CloudFormation stack to create the sources

As soon as deployed (ought to take lower than two minutes), the AWS Lambda operate and Amazon CloudWatch alarm are configured to robotically shut down the Canvas app when idle. To check the auto-shutdown script, do the next:

  1. Guarantee that the SageMaker Canvas app is operating inside the appropriate area and with the appropriate person profile (when you have configured them).
  2. Cease utilizing the SageMaker Canvas app and anticipate the idle timeout interval (default, 2 hours)
  3. Examine that the app is stopped after being idle for the edge time by checking that the CloudWatch alarm has been triggered and, after triggering the automation, it has gone again to the traditional state.

In our check, we have now set the idle timeout interval to 2 hours (7200 seconds). Within the following graph plotted by Amazon CloudWatch Metrics, you’ll be able to see that the SageMaker Canvas app has been emitting the TimeSinceLastActive metric till the edge was met (1), which triggered the alarm. As soon as the alarm was triggered, the AWS Lambda operate was executed, which deleted the app and introduced the metric again beneath the edge (2).

Canvas Auto-shutdown Metrics Plot


On this put up, we carried out an automatic shutdown resolution for idle SageMaker Canvas apps utilizing AWS Lambda and CloudWatch Alarm and the newly emitted metric of idleness from SageMaker Canvas. Because of this resolution, prospects not solely can optimize prices for his or her ML workloads however can even keep away from unintended expenses for purposes that they forgot have been operating of their SageMaker Area.

We’re wanting ahead to seeing what new use instances and workloads prospects can clear up with the peace of thoughts introduced by this resolution. For extra examples of how SageMaker Canvas might help you obtain what you are promoting objectives, consult with the next posts:

To be taught how one can run production-level workloads with Amazon SageMaker Canvas, consult with the next posts:

Concerning the authors

Davide Gallitelli is a Senior Specialist Options Architect for AI/ML. He’s based mostly in Brussels and works intently with prospects throughout the globe that wish to undertake Low-Code/No-Code Machine Studying applied sciences, and Generative AI. He has been a developer since he was very younger, beginning to code on the age of seven. He began studying AI/ML at college, and has fallen in love with it since then.

Huong Nguyen is a Sr. Product Supervisor at AWS. She is main the info ecosystem integration for SageMaker, with 14 years of expertise constructing customer-centric and data-driven merchandise for each enterprise and shopper areas.

Gunjan Garg is a Principal Engineer at Amazon SageMaker group in AWS, offering technical management for the product. She has labored in a number of roles within the AI/ML org for final 5 years and is at present centered on Amazon SageMaker Canvas.

Ziyao Huang is a Software program Growth Engineer with Amazon SageMaker Knowledge Wrangler. He’s enthusiastic about constructing nice product that makes ML straightforward for the shoppers. Outdoors of labor, Ziyao likes to learn, and hang around together with his pals.

Leave a Reply

Your email address will not be published. Required fields are marked *