Spot instances on kubernetes
0 likes103 views
This document discusses strategies for using Spot Instances on Kubernetes clusters managed by Kops. It recommends: 1) Creating multiple instance groups with different instance types and purchasing options to optimize workloads and reduce risks. 2) Setting up Spot interrupt handlers to gracefully handle instances being reclaimed by AWS. 3) Experimenting with adding a small Spot instance group to the cluster, and observing how Kops handles provisioning Spot Instances.
Spot instances on kubernetes
- 1. Spot Instances on K8s (with Kops) Mustafa Kirimli
- 2. Must Haves Having multiples instance groups on Kops will give you freedom to experiment different set of underlying infra for Kubernetes. Multiple instance groups on kops
- 3. Must Haves Having multiple instance/machine specs from first day will give you room to optimize workload based on requirement/consumption. You can eliminate your risk to fall in short on one of the purchasing option. Mixed instance types and multiple purchasing/lifecycle options
- 4. Must Haves Be aware if AWS is asking you to give instance back in 2 minutes (rarely happening and hasn’t happened last 3-4 months to us) Spot interrupt handler
- 5. Steps We Have Taken ● Create a small instance group on kops ● Put taints to new instance group ● Set toleration to least important service ● Trigger scale-in / scale-out to see how it getting provisioned ● Note taken actions on dev and follow same steps on prod cluster Experimenting spot instances on running cluster
- 6. Steps We Have Taken ● When spot instances are getting provisioning there is small difference on initializing steps. ● First ec2 object is getting created and then (when spot request fulfilled) ec2 tags are getting assigned ● Check your kops version whether handling spot provisioning if not put small check on kops instance group Be sure spot instances are getting provisioned properly
- 7. Steps We Have Taken ● “most-pods” or “price” preferably can be set as an expander value on Cluster Autoscaler when working with spot instances.Utilise CA expander option to control which instance group need to be scaled first
- 8. 79% Cost saving on our compute power with mixed instance types Monthly double digit ($$K) saving
- 9. Roadmap ● Having mixed instance type on one IG (ec2 fleet) ● Dynamic bidding on spot ● Trying *5d instance types to reduce cost more ● Tweaking CA to take faster action on spot interruption