open-nomad/e2e/terraform

Terraform infrastructure

This folder contains Terraform resources for provisioning a Nomad cluster on EC2 instances on AWS to use as the target of end-to-end tests.

Terraform provisions the AWS infrastructure assuming that EC2 AMIs have already been built via Packer. It deploys a specific build of Nomad to the cluster along with configuration files for Nomad, Consul, and Vault.

Setup

You'll need Terraform 0.14.7+, as well as AWS credentials to create the Nomad cluster. This Terraform stack assumes that an appropriate instance role has been configured elsewhere and that you have the ability to AssumeRole into the AWS account.

Optionally, edit the terraform.tfvars file to change the number of Linux clients or Windows clients. The Terraform variables file terraform.full.tfvars is for the nightly E2E test run and deploys a larger, more diverse set of test targets.

region                           = "us-east-1"
instance_type                    = "t2.medium"
server_count                     = "3"
client_count_ubuntu_bionic_amd64 = "4"
client_count_windows_2016_amd64  = "1"
profile                          = "dev-cluster"

Run Terraform apply to deploy the infrastructure:

cd e2e/terraform/
terraform apply

Note: You will likely see "Connection refused" or "Permission denied" errors in the logs as the provisioning script run by Terraform hits an instance where the ssh service isn't yet ready. That's ok and expected; they'll get retried. In particular, Windows instances can take a few minutes before ssh is ready.

Nomad Version

You'll need to pass one of the following variables in either your terraform.tfvars file or as a command line argument (ex. terraform apply -var 'nomad_version=0.10.2+ent')

• nomad_local_binary: provision this specific local binary of Nomad. This is a path to a Nomad binary on your own host. Ex. nomad_local_binary = "/home/me/nomad". This setting overrides nomad_version.
• nomad_url: provision this version from a remote archived binary, e.g. build-binaries CircleCI artifacts zip file urls.
• nomad_version: provision this version from releases.hashicorp.com. Ex. nomad_version = "0.10.2+ent"

If you want to deploy the Enterprise build, include -var 'nomad_enterprise=true'.

If you want to bootstrap Nomad ACLs, include -var 'nomad_acls=true'.

Note: If you bootstrap ACLs you will see "No cluster leader" in the output several times while the ACL bootstrap script polls the cluster to start and and elect a leader.

Profiles

The profile field selects from a set of configuration files for Nomad, Consul, and Vault by uploading the files found in ./config/<profile>. The standard profiles are as follows:

• full-cluster: This profile is used for nightly E2E testing. It assumes at least 3 servers and includes a unique config for each Nomad client.
• dev-cluster: This profile is used for developer testing of a more limited set of clients. It assumes at least 3 servers but uses the one config for all the Linux Nomad clients and one config for all the Windows Nomad clients.

You may create additional profiles for testing more complex interactions between features. You can build your own custom profile by writing config files to the ./config/<custom name> directory.

For each profile, application (Nomad, Consul, Vault), and agent type (server, client_linux, or client_windows), the agent gets the following configuration files, ignoring any that are missing.

• ./config/<profile>/<application>/*: base configurations shared between all servers and clients.
• ./config/<profile>/<application>/<type>/*: base configurations shared between all agents of this type.
• ./config/<profile>/<application>/<type>/indexed/*<index>.<ext>: a configuration for that particular agent, where the index value is the index of that agent within the total count.

For example, with the full-cluster profile, 2nd Nomad server would get the following configuration files:

• ./config/full-cluster/nomad/base.hcl
• ./config/full-cluster/nomad/server/indexed/server-1.hcl

The directory ./config/full-cluster/nomad/server has no configuration files, so that's safely skipped.

Outputs

After deploying the infrastructure, you can get connection information about the cluster:

• $(terraform output --raw environment) will set your current shell's NOMAD_ADDR and CONSUL_HTTP_ADDR to point to one of the cluster's server nodes, and set the NOMAD_E2E variable.
• terraform output servers will output the list of server node IPs.
• terraform output linux_clients will output the list of Linux client node IPs.
• terraform output windows_clients will output the list of Windows client node IPs.

SSH

You can use Terraform outputs above to access nodes via ssh:

ssh -i keys/nomad-e2e-*.pem ubuntu@${EC2_IP_ADDR}

The Windows client runs OpenSSH for convenience, but has a different user and will drop you into a Powershell shell instead of bash:

ssh -i keys/nomad-e2e-*.pem Administrator@${EC2_IP_ADDR}

Teardown

The terraform state file stores all the info.

cd e2e/terraform/
terraform destroy

FAQ

E2E Provisioning Goals

1. The provisioning process should be able to run a nightly build against a variety of OS targets.
2. The provisioning process should be able to support update-in-place tests. (See #7063)
3. A developer should be able to quickly stand up a small E2E cluster and provision it with a version of Nomad they've built on their laptop. The developer should be able to send updated builds to that cluster with a short iteration time, rather than having to rebuild the cluster.

Why not just drop all the provisioning into the AMI?

While that's the "correct" production approach for cloud infrastructure, it creates a few pain points for testing:

• Creating a Linux AMI takes >10min, and creating a Windows AMI can take 15-20min. This interferes with goal (3) above.
• We won't be able to do in-place upgrade testing without having an in-place provisioning process anyways. This interferes with goals (2) above.

Why not just drop all the provisioning into the user data?

• Userdata is executed on boot, which prevents using them for in-place upgrade testing.
• Userdata scripts are not very observable and it's painful to determine whether they've failed or simply haven't finished yet before trying to run tests.