[ci.jenkins.io] Define virtual networking for AWS

[dduportal]

We need to define virtual networking for ci.jenkins.io in AWS.

• We need public dual stack network to allow inbound (SSH, HTTP, HTTPS) on the controller with dualstack (both IPv4 and IPv6) network
  • Ref. https://aws.amazon.com/blogs/networking-and-content-delivery/dual-stack-ipv6-architectures-for-aws-and-hybrid-networks/
• Need 3 IPv4 only subnets. Check https://docs.aws.amazon.com/vpc/latest/userguide/configure-subnets.html to be sure it maps to something doable in EC2
  • No need for multiple availability zones (when possible) as we'll use EBS for disks which does NOT span to different zone - https://docs.aws.amazon.com/ebs/latest/userguide/ebs-volumes.html and a simpler setup (we used EKS in jenkins-infra/aws with multiple AZs which was a nightmare to maintain for no added value)
  • As per https://docs.aws.amazon.com/vpc/latest/userguide/configure-subnets.html#subnet-security, we should use private subnets only (to control allowed outbound connections) not possible, we need a public subnet for gw.
• Outbound requests only need to be IPv4, but we will need at least 2 public egress IPv4 (and the ability to extend it) for the whole VPC (See Internet Gateway/Elastic IPs). - we choose to use 1 gw per subnet for now
• Network restrictions for the VPC:
  • Inbound SSH will be restricted through the VPN IPs
  • HTTP/HTTPS/JNLP will be ok for inbound from everywhere
  • Outbound HTTP/HTTPS/HKP to everywhere
  • Outbound SSH only to GitHub IPs
• Private (internal communications) only allowed communications:
  • Controller should be able to reach the VM agents privately through SSH, should be reachable by both VMs and container agents with HTTP/HTTPS but also JNLP + access the internet (see egress VPC rules above)
    • Inbound from internet to the controller: HTTP/HTTPS from everywhere and SSH only from the VPN.
  • VM agents in their own subnets, allowed to reach ci.jenkins.io (inbound JNLP/HTTP/HTTPS) and to be reached through the ci.jenkins.io VM in SSH (outbound- agents) + access the internet (see egress VPC rules above)
  • EKS API which should be private (ref. https://docs.aws.amazon.com/eks/latest/userguide/cluster-endpoint.html) with private Route53 DNS zone
  • EKS containers should be able to access ci.jenkins.io controller in inbound (HTTP/HTTPS/JNLP) and the internet (see egress VPC rules above)
  • The ECR registry used as a pull-through cache should be reachable only privately by any VM or container (including ci.jenkins.io itself)
  • The ACP instance to be hosted in EKS need to be reachable only privately by the VM agents and container agents only (not needed from ci.jio).

[smerle33]

to provide multiple IPs for the gateway:

https://aws.amazon.com/blogs/networking-and-content-delivery/attach-multiple-ips-to-a-nat-gateway-to-scale-your-egress-traffic-pattern/

we can use https://registry.terraform.io/providers/hashicorp/aws/latest/docs/resources/nat_gateway#secondary_allocation_ids but it's not available yet with the module terraform-aws-modules/terraform-aws-vpc#1109

we decided (with @dduportal) to use one gateway per subnet with one ip per gateway for now as it is not yet available with the module and will imply to change it or define by hand all the resources.

[dduportal]

Update regarding network restrictions:

• https://docs.aws.amazon.com/vpc/latest/userguide/vpc-security-best-practices.html tells us we should use Network ACLs (similar to "Network Security Groups" in Azure to restrict traffic going from and to subnets, while using Security groups for the intra-subnet traffic. It means:
  • ci.jenkins.io controller subnet will mostly be protected by Network ACL as we should only have network interface for the controller itself. The associated subnet should only have SG if we add private endpoints or other VMs.
  • ephemeral agents subnet needs both:
    • Network ACLs to forbid inbound access except with SSH from ci.jio or from the VPN outbound (so we can debug VMs accesses) and outbound access only to the internet (HTTP/HTTPS/HKP, etc. ) or to ci.jio (in HTTPS/JNLP).
    • a SG to forbid network access between, agents themselves (to limit transversal attack risks)

[smerle33]

Update regarding No need for multiple availability zones (when possible)

it is mandatory to have at least 2 Availability Zones as per : https://docs.aws.amazon.com/eks/latest/userguide/network-reqs.html#network-requirements-subnets

but we should still be able to lock our nodes to only one AZ to be able to use our EBS volumes (not spanning on multiple AZs) as per : terraform-aws-modules/terraform-aws-eks#1252