Our next task before deployment is to add affinities to the configuration.
In the previous section we created nodegroups with the label intent and values control-apps and apps. Managed node groups automatically create a label eks.amazonaws.com/capacityType and sets its value to SPOT and ON_DEMAND respectively to identify which nodes are Spot Instances and which are On-Demand Instances. For us to adhere to the criteria above we will need to add two affinity properties:
Read about how affinities can be used to assign pods to nodes and modify the monte-carlo-pi-service.yml file accordingly.
As per the Assign Pods to Nodes documentation the objective is to add the following section to the monte-carlo-pi-service.yml.
The following affinities must be added at the spec.template.spec level
affinity:
nodeAffinity:
preferredDuringSchedulingIgnoredDuringExecution:
- weight: 1
preference:
matchExpressions:
- key: eks.amazonaws.com/capacityType
operator: In
values:
- SPOT
requiredDuringSchedulingIgnoredDuringExecution:
nodeSelectorTerms:
- matchExpressions:
- key: intent
operator: In
values:
- apps
If you are still struggling with the implementation, the solution file is available here : monte-carlo-pi-service-final.yml
When creating self managed node groups, we created two nodegroups that were tainted with
spotInstance: "true:PreferNoSchedule". PreferNoSchedule is used to indicate we prefer pods not to be scheduled on Spot Instances. NoSchedule can also be used to enforce a hard discrimination as a taint. To overcome the spotInstance: "true:PreferNoSchedule" taint, we need to create a toleration in the deployment. Read about how tolerations are applied and modify the monte-carlo-pi-service.yml file accordingly.
As per the tolerations documentation
the objective is to add the following section to the monte-carlo-pi-service.yml.
The following toleration must be added at the spec.template.spec level
tolerations:
- key: "spotInstance"
operator: "Equal"
value: "true"
effect: "PreferNoSchedule"