Add option to launch all nodes into the same AZ

[armandmcqueen]

Why do you want this feature?
We are running ML training workloads on EKS that are very sensitive to network latency. For performance reasons, we need our nodes to exist in a single AZ. Using --zones with a single AZ raises an error.

What feature/behavior/change do you want?
When launching a new cluster, I would like to be able to pass in a single AZ with the --zones option or to have a --single-az flag that forces nodes to launch into the same AZ.

Any workflow that is less unwieldy than manually altering the AutoScaling group's subnet list would be helpful.

[gruebel]

Hi,
This is not possible due to API restrictions of AWS. You need to specify at least two Subnets in two different AZs.

[armandmcqueen]

I'm fine with having multiple subnets for EKS. I just need the worker autoscaling group to live in a single AZ.

[errordeveloper]

@armandmcqueen I'm curious to know what happens if you pass e.g. --zones us-west-2a,us-west-2a,us-west-2a?

@gruebel is that an EKS API requirement, or something else?

I think in the future, we should be able to allow this to happen for individual node-groups, perhaps --single-az would makes sense in that context. It'd be nice to allow users to have single-AZ nodegroups at the same time as nodegroups that span multiple AZs.

[gruebel]

@errordeveloper yeah it is a requirement by the EKS API, if you try it, you will directly get an error message.
I also tried to put in three times the same AZ, then you will get the error and can see in CloudFormation Stack that you need two Subnets in two different AZs.

[errordeveloper]

We will be able to do this very easily soon (namely after #257).

[errordeveloper]

@richardcase as you are working on #281, there's now a NodeGroup.AvailibilityZones field that can be leveraged for this, and I think it would be sufficient to just expose it as a flag as there is plumbing already in place.

https://github.com/weaveworks/eksctl/blob/99b5d58d340660ba7299ee470888b022754a2d8e/pkg/cfn/builder/nodegroup.go#L113-L122

[errordeveloper]

This is definitely done now, and we have ironed out a bug in this week - #432.

[dnascimento]

Does it assume only one subnet per AZ? Can we deploy on multiple subnets in the same AZ?

[errordeveloper]

No, it is single subnet per AZ. If you have a particular use-case, please open an issue with more details of that use-case.

…

On Thu, 19 Sep 2019, 3:04 pm Dário Nascimento, ***@***.***> wrote: Does it assume only one subnet per AZ? Can we deploy on multiple subnets in the same AZ? — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#232>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAB5MS6W5YVZCDKCE25ZDTDQKOBHNANCNFSM4FYA6CFA> .