The purpose of auto-scaling is to apply SLA scaling policies to a Cloudbreak-managed Hadoop cluster.

How It Works

The auto-scaling capabilities is based on Ambari Metrics - and Ambari Alerts. Based on the Blueprint used and the running services, Cloudbreak can access all the available metrics from the subsystem and define alerts based on this information.

Beside the default Ambari Metrics, Cloudbreak includes two custom metrics: Pending YARN containers and Pending applications. These two custom metrics works with the YARN subsystem in order to bring application level QoS to the cluster.

In order to use the autoscaling feature with Cloudbreak you will have to enable from the UI or shell.


Auto-scaling supports two Alert types: metric and time based.

Metric-based Alerts

Metric based alerts are using the default (or custom) Ambari metrics. These metrics have a default Threshold value configured in Ambari - nevertheless these thresholds can be configured, changed or altered in Ambari. In order to change the default threshold for a metric please go to Ambari UI and select the Alerts tab and the metric. The values can be changed in the Threshold section.

Metric alerts have a few configurable fields.

Time-based Alerts

Time based alerts are based on cron expressions and allow alerts to be triggered based on time.

Time alerts have a few configurable fields.

Scaling Policies

Scaling is the ability to increase or decrease the capacity of the Hadoop cluster or application based on an alert. When scaling policies are used, the capacity is automatically increased or decreased according to the conditions defined. Cloudbreak will do the heavy lifting and based on the alerts and the scaling policy linked to them it executes the associated policy. We scaling granularity is at the hostgroup level - thus you have the option to scale services or components only, not the whole cluster.

Scaling policies have a few configurable fields.

Cluster Scaling Configuration

An SLA scaling policy can contain multiple alerts. When an alert is triggered a scaling adjustment is applied, however to keep the cluster size within boundaries a cluster size min. and cluster size max. is attached to the cluster - thus a scaling policy can never over or undersize a cluster. Also in order to avoid stressing the cluster we have introduced a cooldown time period (minutes) - though an alert is raised and there is an associated scaling policy, the system will not apply the policy within the configured timeframe. In an SLA scaling policy the triggered rules are applied in order.

Downscale Scaling Considerations

Cloudbreak auto-scaling will try to keep a healthy cluster, thus does several background checks during downscale.

Edit on GitHub