Improve the AWS VPC architecture

[0x2b3bfa0]

(Follow-up of iterative/cml#484)

We're using the first VPC and the first subnet to deploy EC2 instances, but this behavior can be problematic in some cases.

We can either create a new VPC as part of the provisioning process or just allow users to specify which one should be used. The first solution could be more transparent for users and allow us to keep the common interface as pure as possible.

[haviland-nh]

First posting here!

Just wanted to add an upvote here to provide VPC and Subnet as a parameter. I think this feature would add flexibility to the provider, without requiring it to do more than it has to.

[DavidGOrtega]

We have to create the network resource. This is not new actually nor anything that we have decided recently. In the very beginning the provider needed to have two additional resources:

• Network
• Volumes

The network resource will be a very simple network abstraction on top of the cloud vendors. The goal is to have like CML a very yet powerful implementation and the major goal to solve here for ML purposes is the need to create clusters of machines that allow us to do distributed training.
Achieving that goal we will also provide that reusable network where all the basic needs are going to be covered, similar to any cloud vendor capabilities.

[0x2b3bfa0]

Closing in favor of #123 and children

[0x2b3bfa0]

Reopening after the organization, related to (but not mixed with) #124

[dacbd]

I think we can close this, a target vpc can be selected with aws_security_group and a target subnet with aws_subnet_id

perhaps a larger "networking" issue exists so that this can functionally exist with gcp/az @0x2b3bfa0 ?

[0x2b3bfa0]

Related to #289 and #299; more specifically to making use of existing resources.

Yes, we can close this for the reasons you mention.