New to KubeDB? Please start here.

Distributed MariaDB Compute Resource Autoscaling

This guide will give an overview on how KubeDB Autoscaler operator autoscales the compute resources i.e. cpu and memory of a distributed MariaDB cluster using mariadbautoscaler crd.

Before You Begin

• You should be familiar with the following KubeDB concepts:
  • MariaDB
  • MariaDBAutoscaler
  • MariaDBOpsRequest
  • Distributed MariaDB Overview

How Compute Autoscaling Works

The following diagram shows how KubeDB Autoscaler operator autoscales the resources of MariaDB database components. Open the image in a new tab to see the enlarged version.

The Auto Scaling process consists of the following steps:

1. At first, the user creates a PlacementPolicy Custom Resource (CR) with monitoring.prometheus.url configured for each spoke cluster. This allows the autoscaler to scrape metrics from the Prometheus instance running in each spoke cluster.

2. The user creates a MariaDB Custom Resource Object (CRO) with spec.distributed: true and a reference to the PlacementPolicy.

3. KubeDB Community operator watches the MariaDB CRO.

4. When the operator finds a MariaDB CRO, it creates required number of PetSets and distributes them across spoke clusters as defined by the PlacementPolicy.

5. Then, in order to set up autoscaling of the CPU & Memory resources of the MariaDB database the user creates a MariaDBAutoscaler CRO with desired configuration.

6. KubeDB Autoscaler operator watches the MariaDBAutoscaler CRO.

7. KubeDB Autoscaler operator utilizes the modified version of Kubernetes official VPA-Recommender for different components of the database, as specified in the mariadbautoscaler CRO. It generates recommendations based on resource usages by querying Prometheus endpoints configured in the PlacementPolicy, & stores them in the status section of the autoscaler CRO.

8. If the generated recommendation doesn’t match the current resources of the database, then KubeDB Autoscaler operator creates a MariaDBOpsRequest CRO to scale the database to match the recommendation provided by the VPA object.

9. KubeDB Ops-Manager operator watches the MariaDBOpsRequest CRO.

10. Lastly, the KubeDB Ops-Manager operator will scale the database component vertically as specified on the MariaDBOpsRequest CRO.

Key Difference from Non-Distributed Autoscaling: For distributed MariaDB, the PlacementPolicy must include a monitoring.prometheus.url for each spoke cluster’s distributionRules entry. The autoscaler uses these Prometheus endpoints to collect resource metrics from pods running across multiple Kubernetes clusters.

In the next docs, we are going to show a step by step guide on Autoscaling of Distributed MariaDB database using MariaDBAutoscaler CRD.