To run etcd cluster at large-scale, it is important to assign etcd pods to nodes with desired resources, such as SSD, high performance network.
Kuberentes nodes can be attached with labels. Users can assign pods to nodes with given labels. Similarly for the etcd-operator, users can specify
Node Selector in the pod policy to select nodes for etcd pods. For example, users can label a set of nodes with SSD with label
"disk"="SSD". To assign etcd pods to these node, users can specify
"disk"="SSD" node selector in the cluster spec.
Even with container isolation, not all resources are isolated. Thus, performance interference can still affect etcd clusters' performance unexpectedly. We recommend dedicating nodes for each etcd pod to achieve predictable performance.
This feature requires Kuberentes taint support, which is not fully implemented yet. Once Kubernetes supports this feature, we will implement it in the operator.
Kuberentes node can be tainted with keys. Together with node selector feature, users can create nodes that are dedicated for only running etcd clusters. This is the suggested way to run high performance large scale etcd clusters.
Use kubectl to taint the node with kubectl taint nodes etcd dedicated. Then only etcd pods, which tolerate this taint will be assigned to the nodes.