Kubernetes Static Pod Autoscaler (k8s-spa)

• Use cases
• Installation
• Usage

k8s-spa is a command line utility that will automatically calculate memory requests for your pods based on:

1. real memory usage of your pods, based on Prometheus metrics
2. your risk tolerance (the higher your risk, the lower the requests)

It is similar to the Vertical Pod Autoscaler (VPA), in the sense that they both calculate requests for pods. However, there are several differences:

• k8s-spa calculates the requests offline, that is, there is no need to deploy anything to your Kubernetes cluster in order to use this tool.
• k8s-spa takes into account the memory metrics of all your pods in order to calculate the requests. This allows k8s-spa to further optimize the request values, since there isn't any harm in a pod using more memory than requested as long as there is slack from other pods that can absorb the excess memory usage. Additionally, k8s-spa is able to take into account memory correlations between pods

Use cases

Who would benefit from using this tool?

• K8s users that aren't autoscaling their pods: instead of going through several cycles of trial and error to set the requests of your pods, you can just get a reliable request value by running k8s-spa.
• People using the Horizontal Pod Autoscaler (HPA): The VPA doesn't play well with the HPA (Horizontal Pod Autoscaler) unless it is configured with custom metrics. You can use k8s-spa to automatically calculate your requests and still be able to use the HPA.

Installation

Source

Install cargo via rustup.

cargo install k8s-spa

Binary

Currently, the only binary distribution is through a docker container image:

docker run telefonica/k8s-spa:0.3.1

You can build the docker image with:

docker build -f Cargo.toml .

Usage

There are two phases: importing data and analyzing data.

Importing

Right now, we only support importing memory metrics from a Prometheus server:

k8s-spa import prometheus
        --end-date <end_date>
        --output <output>
        --start-date <start_date>
        --url <url>

In the future, other data sources could be implemented, like the Metric Server API from Kubernetes.

Analyzing

Once you have imported your data into a file, you can analyze it to get the request values:

k8s-spa analyze --data <data> --risk-tolerance <risk>

The risk-tolerance parameter indicates how much risk of OOMKilled are you willing to accept for your particular scenario. The default value is 0.05 (i.e. 5%).

Request calculation details

More documentation coming soon. For now, take a look at the code if you want to understand the gory details.