Outlier detection

If you’re watching average CPU across a 20-node cluster:

• One node at 95% barely moves the cluster average, so nothing looks wrong.
• An alert doesn’t fire, so you don’t know until it affects users.

When you’re trying to determine an instance that is behaving differently from its peers, a fixed threshold can’t help.

Outlier detection solves these kinds of issues.

• You define the group of instances that should behave alike.
• Outlier detection flags any that drift from the rest.