Azure App Service Autoscale: Why It Might Not Scale In and How to Fix It
Azure App Service autoscale is a powerful feature that automatically adjusts the number of instances of your web app based on predefined metrics. While it excels at scaling out (adding instances), scaling in (reducing instances) can sometimes be problematic. This article explores the reasons behind this issue, offering insights and solutions to ensure your app effectively scales both in and out.
The Problem: Autoscale Doesn't Always Scale In
Imagine you've configured your app to scale out when CPU usage exceeds 70%. As expected, your app scales up to handle the load. However, after the traffic surge subsides, the CPU usage drops below 70%, yet your app doesn't scale back down. This scenario exemplifies the challenge of "stuck" instances, where your app remains at a higher scale than needed, leading to unnecessary costs.
The Code: A Common Autoscale Configuration
{
"profiles": [
{
"name": "Production",
"capacity": {
"default": 1,
"minimum": 1,
"maximum": 10
},
"autoscale": {
"enabled": true,
"rules": [
{
"name": "CPU Rule",
"timeGrain": "PT5M",
"scaleAction": {
"type": "ChangeCount",
"cooldown": "PT5M",
"direction": "Increase",
"count": 1
},
"conditions": [
{
"type": "Metric",
"metricName": "Percentage CPU",
"metricNamespace": "Microsoft.Compute/virtualMachines",
"timeAggregation": "Average",
"operator": "GreaterThan",
"threshold": 70
}
]
},
{
"name": "CPU Rule Scale Down",
"timeGrain": "PT5M",
"scaleAction": {
"type": "ChangeCount",
"cooldown": "PT5M",
"direction": "Decrease",
"count": 1
},
"conditions": [
{
"type": "Metric",
"metricName": "Percentage CPU",
"metricNamespace": "Microsoft.Compute/virtualMachines",
"timeAggregation": "Average",
"operator": "LessThan",
"threshold": 30
}
]
}
]
}
}
]
}
This configuration scales out when CPU usage surpasses 70% and scales in when it drops below 30%. However, the issue often lies in the "cooldown" period or the "threshold" setting.
The Insights: Understanding the Root Causes
- Cooldown Period: The "cooldown" period prevents the autoscaler from rapidly changing the instance count. While this safeguards against overly aggressive scaling, a long cooldown can hinder scaling in.
- Threshold Gap: If the gap between the scale-out threshold (70%) and the scale-in threshold (30%) is too large, it creates a range where CPU usage can fluctuate without triggering scaling in.
- Metric Aggregation: The "timeAggregation" parameter, usually set to "Average", can mask spikes in CPU usage. A high average CPU over a 5-minute interval might not accurately reflect the short-lived spikes that triggered the scale-out initially.
- Warm-up Time: When scaling in, instances need time to warm up before they can be considered for removal. If the cooldown period is too short, the autoscaler might attempt to remove instances before they are ready.
The Solutions: Optimizing for Efficient Scaling In
- Reduce Cooldown: Experiment with shorter cooldown periods for scale-in rules. Start with a value close to the "timeGrain" interval (e.g., 5 minutes) to enable quicker response to decreasing load.
- Narrow Threshold Gap: Bring the scale-in threshold closer to the scale-out threshold. A smaller gap (e.g., 5-10%) can ensure a more responsive scaling behavior.
- Use "Min" Aggregation: Switch the "timeAggregation" to "Min" for the scale-in rule. This will trigger scaling in even if the average CPU usage is above the threshold, but the minimum CPU usage dips below.
- Adjust Warm-up Time: For apps that require significant warm-up, consider using a longer cooldown period for scale-in. This allows the instances to reach a stable state before being removed.
Additional Tips and Best Practices
- Monitor Scaling Events: Use Azure Monitor to track scaling events and identify potential problems.
- Test Thoroughly: Simulate various load scenarios and test your autoscale configuration in a development or staging environment before deploying to production.
- Use a Hybrid Approach: Consider combining autoscale with manual scaling to fine-tune your app's capacity.
By understanding the potential causes of autoscale failing to scale in and implementing the recommended solutions, you can optimize your Azure App Service for efficient resource utilization and cost-effectiveness.
