Now that we shed a bit of light around Kubernetes’ killing activities, we should note that (almost) nothing happens randomly. When there aren’t enough resources to serve the needs of all the Pods, Kubernetes will destroy one or more containers. The decision which one it will be is anything but random. Who will be the unlucky one depends on the assigned Quality of Service (QoS). Those with the lowest priority are killed first.

Since this might be the first time you heard about QoS, we’ll spend some time explaining what it is and how it works.

Pods are the smallest units in Kubernetes. Since almost everything ends up as a Pod (one way or another), it is no wonder that Kubernetes promises specific guarantees to all the Pods running inside the cluster. Whenever we send a request to the API to create or update a Pod, it gets assigned one of the Quality of Service (QoS) classes. They are used to make decisions such as where to schedule a Pod or whether to evict it.

We do not specify QoS directly. Instead, they are assigned based on the decisions we make with resource requests and limits.

Now that we shed a bit of light around Kubernetes’ killing activities, we should note that (almost) nothing happens randomly. When there aren’t enough resources to serve the needs of all the Pods, Kubernetes will destroy one or more containers. The decision which one it will be is anything but random. Who will be the unlucky one depends on the assigned Quality of Service (QoS). Those with the lowest priority are killed first.

Since this might be the first time you heard about QoS, we’ll spend some time explaining what it is and how it works.

Pods are the smallest units in Kubernetes. Since almost everything ends up as a Pod (one way or another), it is no wonder that Kubernetes promises specific guarantees to all the Pods running inside the cluster. Whenever we send a request to the API to create or update a Pod, it gets assigned one of the Quality of Service (QoS) classes. They are used to make decisions such as where to schedule a Pod or whether to evict it.

We do not specify QoS directly. Instead, they are assigned based on the decisions we make with resource requests and limits.