How does the metric cluster widget calculate health level?

6000110162

2015-08-21T13:01:20-04:00

6000171361

false

Dashboards 5

2015-10-30T11:00:31-04:00

2015-09-24T19:10:39-04:00

The Metric Cluster Widget allows you to visualize the health level of each member of a cluster - giving you a quick way to check the health of a whole group, or "cluster," of machines. Select the Cluster Widget: Choose the Cluster (and set the threshold): That's it - You get a grid of all the cluster members with color coded health levels. When you view your Dashboard, you can hover over a cluster member to see the actual values, then click on it to jump to the check and begin your analysis. The cluster health is calculated by taking 2 hours of time series data from each of the metrics and running a linear correlation against them. We then walk the matrix to determine the covariance and use a voting algorithm to allow those that are highly correlated to vote that others metrics are “bad.” Basically, the health measurement is determined by how much the current signal looks like the majority of signals. We are continuing to improve, and once we refine this procedure a bit, the new procedure will be documented through the UI.

The Metric Cluster Widget allows you to visualize the health level of each member of a cluster - giving you a quick way to check the health of a whole group, or "cluster," of machines. Select the Cluster Widget: <img src="https://s3.amazonaws.com/cdn.freshdesk.com/data/helpdesk/attachments/production/6008246776/original/blob1446229004404.png?1446229001" class="inline-image" data-id="6008246776"> Choose the Cluster (and set the threshold): <img src="https://s3.amazonaws.com/cdn.freshdesk.com/data/helpdesk/attachments/production/6008246850/original/Choose%20cluster.png?1446229082" class="inline-image" data-id="6008246850"> That's it - You get a grid of all the cluster members with color coded health levels. When you view your Dashboard, you can hover over a cluster member to see the actual values, then click on it to jump to the check and begin your analysis. The cluster health is calculated by taking 2 hours of time series data from each of the metrics and running a linear correlation against them. We then walk the matrix to determine the covariance and use a voting algorithm to allow those that are highly correlated to vote that others metrics are “bad.” Basically, the health measurement is determined by how much the current signal looks like the majority of signals. We are continuing to improve, and once we refine this procedure a bit, the new procedure will be documented through the UI.