AzureFeeds

We’re excited to announce the public preview of Predictive Pooling for Azure Virtual Machine Scale Sets (VMSS) and Azure Container Instances (ACI)—a new capability that brings intelligent, data-driven sizing information to your standby pools.

What is Predictive Pooling?

Predictive pooling uses historical usage data and machine learning to forecast demand and recommend the optimal number of pre-provisioned instances to maintain in your standby pool.

 This helps you:

• Right-size your standby pools to balance cost and readiness.


• Reduce latency by ensuring the right number of instances are always ready.


• Improve operational efficiency by automating pool sizing decisions.

Prediction results are available 2-3 weeks after creating the standby pool.

How it Works

For both Virtual Machine Scale Sets and Azure Container Instances, you can now query the runtime view of your standby pool to retrieve:

• Hourly forecasts of the number of instances to maintain over a 12-hour window.


• Recent request history for context.


• Forecast accuracy metrics to guide confidence in the predictions.

Here’s a sample response from the API:

{ “prediction”: {
“forecastValues”: {
“instancesRequestedCount”: “[5, 3, 2, 8, 8, 2, 3, 6, 5, 3, 9, 8]”
},
“forecastStartTime”: “2024-11-12T16:30”,
“info”: {
“forecastAccuracy”: 85,
“seriesUnitIntervalInMins”: 60,
“instancesRequestedCount_recentHistory”: “[9, 4, 2, 8, 8, 2, 3, 6, 5, 3, 2, 6]”
}
}
}

This data can be retrieved using various APIs such as CLI, PowerShell, and REST.

Why it Matters

Predictive pooling is a major step forward in making Azure’s standby pool infrastructure smarter and more cost-effective. Whether you’re running latency-sensitive services on Virtual Machine Scale Sets or bursty container workloads on Azure Container Instances, this feature helps ensure your resources are ready when you need them.

Learn More

• Get Prediction Results for Standby pools for Virtual Machine Scale Sets (Preview) 


• Get prediction results for standby pools for Azure Container Instances (Preview)