Fix terraform resource churn

[mrwacky42]

Possibly related to this issue with the Terraform AWS provider:
hashicorp/terraform-provider-aws#1521

Terraform or the AWS API report 0 IOPS for gp2 volumes after creation.
If I create a gp2 volume with these modules, every subsequent run of
Terraform will attempt to destroy and recreate the LC in order to make
IOPs match. This change sets IOPS to 0 for non-io1 type volumes.

Since the number of IOPS for gp2 volume is based on the size, Terraform
arguably should only accept the iops parameter for io1 volumes.

Interestingly, for the etcd nodes, this problem does not occur because
the Terraform state shows 100 IOPS even for gp2 volumes.

[coreosbot]

Can one of the admins verify this patch?

[coreosbot]

Can one of the admins verify this patch?

[coreosbot]

Can one of the admins verify this patch?

[squat]

@mrwacky42 this generally looks pretty good but seems fragile. Why not update the etcd node launch config as well? And what happens when that node's IOPs in Terraform state is no longer 100?

[mrwacky42]

@squat - I based this change on the current behavior of Terraform and AWS.

The etcd nodes are not using autoscaling, so there is no launch configuration, just aws_instance resources.

For aws_instance resources, Terraform state currently reports 100 iops (for non-io1 EBS volumes). For aws_launch_configuration resources, Terraform state currently reports 0 iops.

There is a PR attached to the original issue I linked above to fix the aws_instance to behave consistent with aws_launch_configuration: hashicorp/terraform-provider-aws#1573

[squat]

thank I see now. From the issue you linked, it looks like even settings iops to 100 for gp2 volumes will not always work. @fraenkel writes that for large volume sizes the iops can be reported to be 150, so we may still have resource churn.

[squat]

As @mrwacky42 correctly, notes, iops defaults to 0 for gp2 volumes in launch configurations, but defaults to size*3 for instance resources. This explains why @simnalamburt was seeing the issue in #1397 (comment) for only the volumes used in launch configurations but not the volumes used etcd nodes. This PR partially undoes #1397. We should merge #953 next to ensure that iops is set correctly for volumes in instance resources.

cc @s-urbaniak @alexsomesan

[squat]

closing and reopening to trigger tests

[squat]

@mrwacky42 tests are failing with:

misformatted files (run 'terraform fmt .' to fix): platforms/aws/variables.tf

can you please fix the formatting?

[mrwacky42]

@squat - Maybe I shouldn't have rebased. Did I confuse Jenkins?

[squat]

@mrwacky42 no, rebasing was good 👍. Linting is still failing. Can you please make docs && make examples and commit the result?

[mrwacky42]

I cheated and ran make docs examples, then rebased again.

😀

[squat]

@mrwacky42 had an unrelated flake in one of the azure tests. Merging!