Terraform module for scalable self hosted GitHub action runners

This Terraform module creates the required infrastructure needed to host GitHub Actions self hosted, auto scaling runners on AWS spot instances. It provides the required logic to handle the life cycle for scaling up and down using a set of AWS Lambda functions. Runners are scaled down to zero to avoid costs when no workflows are active.

• Motivation
• Overview
  • ARM64 support via Graviton/Graviton2 instance-types
• Usages
  • Setup GitHub App (part 1)
  • Setup terraform module
  • Setup GitHub App (part 2)
  • Encryption
  • Idle runners
• Examples
• Sub modules
  • ARM64 configuration for submodules
• Debugging
• Requirements
• Providers
• Inputs
• Outputs
• Contribution
• Philips Forest

Motivation

GitHub Actions self hosted runners provide a flexible option to run CI workloads on infrastructure of your choice. Currently there is no option provided to automate the creation and scaling of action runners. This module takes care of creating the AWS infrastructure to host action runners on spot instances. It provides lambda modules to orchestrate the life cycle of the action runners.

Lambda is chosen as runtime for two major reasons. First it allows to create small components with minimal access to AWS and GitHub. Secondly it provides a scalable setup with minimal costs that works on repo level and scales to organization level. The lambdas will create Linux based EC2 instances with Docker to serve CI workloads that can run on Linux and/or Docker. The main goal is to support Docker based workloads.

A logical question would be why not Kubernetes? In the current approach we stay close to the way the GitHub action runners are available today. The approach is to install the runner on a host where the required software is available. With this setup we stay quite close to the current GitHub approach. Another logical choice would be AWS Auto Scaling groups. This choice would typically require much more permissions on instance level to GitHub. And besides that, scaling up and down is not trivial.

Overview

The moment a GitHub action workflow requiring a self-hosted runner is triggered, GitHub will try to find a runner which can execute the workload. This module reacts to GitHub's check_run event for the triggered workflow and creates a new runner if necessary.

For receiving the check_run event, a GitHub App needs to be created with a webhook to which the event will be published. Installing the GitHub App in a specific repository or all repositories ensures the check_run event will be sent to the webhook.

In AWS a API gateway endpoint is created that is able to receive the GitHub webhook events via HTTP post. The gateway triggers the webhook lambda which will verify the signature of the event. This check guarantees the event is sent by the GitHub App. The lambda only handles check_run events with status created. The accepted events are posted on a SQS queue. Messages on this queue will be delayed for a configurable amount of seconds (default 30 seconds) to give the available runners time to pick up this build.

The "scale up runner" lambda is listening to the SQS queue and picks up events. The lambda runs various checks to decide whether a new EC2 spot instance needs to be created. For example, the instance is not created if the build is already started by an existing runner, or the maximum number of runners is reached.

The Lambda first requests a registration token from GitHub which is needed later by the runner to register itself. This avoids that the EC2 instance, which later in the process will install the agent, needs administration permissions to register the runner. Next the EC2 spot instance is created via the launch template. The launch template defines the specifications of the required instance and contains a user_data script. This script will install the required software and configure it. The registration token for the action runner is stored in the parameter store (SSM) from which the user data script will fetch it and delete it once it has been retrieved. Once the user data script is finished the action runner should be online and the workflow will start in seconds.

Scaling down the runners is at the moment brute-forced, every configurable amount of minutes a lambda will check every runner (instance) if it is busy. In case the runner is not busy it will be removed from GitHub and the instance terminated in AWS. At the moment there seems no other option to scale down more smoothly.

Downloading the GitHub Action Runner distribution can be occasionally slow (more than 10 minutes). Therefore a lambda is introduced that synchronizes the action runner binary from GitHub to an S3 bucket. The EC2 instance will fetch the distribution from the S3 bucket instead of the internet.

Secrets and private keys which are passed to the lambdas as environment variables are encrypted by default by a KMS key managed by the module. Alternatively you can pass your own KMS key. Encryption via KMS can be complete disabled by setting encrypt_secrets to false.

Permission are managed on several places. Below the most important ones. For details check the Terraform sources.

• The GitHub App requires access to actions and publish check_run events to AWS.
• The scale up lambda should have access to EC2 for creating and tagging instances.
• The scale down lambda should have access to EC2 to terminate instances.

Besides these permissions, the lambdas also need permission to CloudWatch (for logging and scheduling), SSM and S3. For more details about the required permissions see the documentation of the IAM module which uses permission boundaries.

ARM64 support via Graviton/Graviton2 instance-types

When using the default example or top-level module, specifying an instance_type that matches a Graviton/Graviton 2 (ARM64) architecture (e.g. a1 or any 6th-gen g or gd type), the sub-modules will be automatically configured to provision with ARM64 AMIs and leverage GitHub's ARM64 action runner. See below for more details.

Usages

Examples are provided in the example directory. Please ensure you have installed the following tools.

• Terraform, or tfenv.
• Bash shell or compatible
• Docker (optional, to build lambdas without node).
• AWS cli (optional)
• Node and yarn (for lambda development).

The module supports two main scenarios for creating runners. On repository level a runner will be dedicated to only one repository, no other repository can use the runner. On organization level you can use the runner(s) for all the repositories within the organization. See https://help.github.com/en/actions/hosting-your-own-runners/about-self-hosted-runners for more information. Before starting the deployment you have to choose one option.

GitHub workflows fail immediately if there is no action runner available for your builds. Since this module supports scaling down to zero, builds will fail in case there is no active runner available. We recommend to create an offline runner with matching labels to the configuration. Create this runner manually by following the GitHub instructions for adding a new runner on your local machine. If you stop the process after the step of running the config.sh script the runner will remain offline. This offline runner ensures that builds will not fail immediately and stay queued until there is an EC2 runner to pick it up.

The setup consists of running Terraform to create all AWS resources and manually configuring the GitHub App. The Terraform module requires configuration from the GitHub App and the GitHub app requires output from Terraform. Therefore you first create the GitHub App and configure the basics, then run Terraform, and afterwards finalize the configuration of the GitHub App.

Setup GitHub App (part 1)

Go to GitHub and create a new app. Beware you can create apps your organization or for a user. For now we support only organization level apps.

1. Create app in Github
2. Choose a name
3. Choose a website (mandatory, not required for the module).
4. Disable the webhook for now (we will configure this later).
5. Permissions for all runners:

• Repository:
  • Actions: Read-only (check for queued jobs)
  • Checks: Read-only (receive events for new builds)
  • Metadata: Read-only (default/required)

6. Permissions for repo level runners only:

• Repository:
  • Administration: Read & write (to register runner)

7. Permissions for organization level runners only:

• Organization
  • Administration: Read & write (to register runner)
  • Self-hosted runners: Read & write (to register runner)

8. Save the new app.
9. On the General page, make a note of the "App ID" and "Client ID" parameters.
10. Create a new client secret and also write it down.
11. Generate a new private key and save the app.private-key.pem file.

Setup terraform module

Download lambdas

To apply the terraform module, the compiled lambdas (.zip files) need to be available either locally or in an S3 bucket. They can be either downloaded from the GitHub release page or build locally.

To read the files from S3, set the lambda_s3_bucket variable and the specific object key for each lambda.

The lambdas can be downloaded manually from the release page or using the download-lambda terraform module (requires curl to be installed on your machine). In the download-lambda directory, run terraform init && terraform apply. The lambdas will be saved to the same directory.

For local development you can build all the lambdas at once using .ci/build.sh or individually using yarn dist.

Service-linked role

To create spot instances the AWSServiceRoleForEC2Spot role needs to be added to your account. You can do that manually by following the AWS docs. To use terraform for creating the role, either add the following resource or let the module manage the the service linked role by setting create_service_linked_role to true. Be aware this is an account global role, so maybe you don't want to mange it via a specific deployment.

resource "aws_iam_service_linked_role" "spot" {
  aws_service_name = "spot.amazonaws.com"
}

Terraform module

Next create a second terraform workspace and initiate the module, or adapt one of the examples.

Note that github_app.key_base64 needs to be the base64-encoded .pem file, i.e., the output of base64 app.private-key.pem (not directly the content of app.private-key.pem).

module "github-runner" {
  source  = "philips-labs/github-runner/aws"
  version = "0.9.1"

  aws_region = "eu-west-1"
  vpc_id     = "vpc-123"
  subnet_ids = ["subnet-123", "subnet-456"]

  environment = "gh-ci"

  github_app = {
    key_base64     = "base64string"
    id             = "1"
    client_id      = "c-123"
    client_secret  = "client_secret"
    webhook_secret = "webhook_secret"
  }

  webhook_lambda_zip                = "lambdas-download/webhook.zip"
  runner_binaries_syncer_lambda_zip = "lambdas-download/runner-binaries-syncer.zip"
  runners_lambda_zip                = "lambdas-download/runners.zip"
  enable_organization_runners = true
}

ARM64 support: Specify an a1 or *6g* (6th-gen Graviton2) instance type to stand up an ARM64 runner, otherwise the default is x86_64.

Run terraform by using the following commands

terraform init
terraform apply

The terraform output displays the API gateway url (endpoint) and secret, which you need in the next step.

The lambda for syncing the GitHub distribution to S3 is triggered via CloudWatch (by default once per hour). After deployment the function is triggered via S3 to ensure the distribution is cached.

Setup GitHub App (part 2)

Go back to the GitHub App and update the following settings.

1. Enable the webhook.
2. Provide the webhook url, should be part of the output of terraform.
3. Provide the webhook secret.
4. Enable the check_run event for the webhook.
5. In the "Install App" section, install the App in your organization, either in all or in selected repositories.

You are now ready to run action workloads on self hosted runner. Remember that builds will fail if there is no (offline) runner available with matching labels.

Encryption

The module support 3 scenario's to manage environment secrets and private key of the Lambda functions.

Encrypted via a module managed KMS key (default)

This is the default, no additional configuration is required.

Encrypted via a provided KMS key

You have to create an configure you KMS key. The module will use the context with key: Environment and value var.environment as encryption context.

resource "aws_kms_key" "github" {
  is_enabled = true
}

module "runners" {

  ...
  manage_kms_key = false
  kms_key_id     = aws_kms_key.github.key_id
  ...

No encryption

Not advised but you can disable the encryption as by setting the variable encrypt_secrets to false.

Idle runners

The module will scale down to zero runners be default, by specifying a idle_config config idle runners can be kept active. The scale down lambda checks if any of the cron expressions matches the current time with a marge of 5 seconds. When there is a match the number of runners specified in the idle config will be kept active. In case multiple cron expressions matches only the first one is taken in to account. Below an idle configuration for keeping runners active from 9 to 5 on working days.

idle_config = [{
   cron      = "* * 9-17 * * 1-5"
   timeZone  = "Europe/Amsterdam"
   idleCount = 2
}]

Supported config

Cron expressions are parsed by cron-parser. The supported syntax.

*    *    *    *    *    *
┬    ┬    ┬    ┬    ┬    ┬
│    │    │    │    │    |
│    │    │    │    │    └ day of week (0 - 7) (0 or 7 is Sun)
│    │    │    │    └───── month (1 - 12)
│    │    │    └────────── day of month (1 - 31)
│    │    └─────────────── hour (0 - 23)
│    └──────────────────── minute (0 - 59)
└───────────────────────── second (0 - 59, optional)

For time zones please check TZ database name column for the supported values.

Examples

Examples are located in the examples directory. The following examples are provided:

• Default: The default example of the module
• Permissions boundary: Example usages of permissions boundaries.

Sub modules

The module contains several submodules, you can use the module via the main module or assemble your own setup by initializing the submodules yourself.

The following submodules are the core of the module and are mandatory:

• runner-binaries-syncer - Syncs the action runner distribution.
• runners - Scales the action runners up and down
• webhook - Handles GitHub webhooks

The following sub modules are optional and are provided as example or utility:

• download-lambda - Utility module to download lambda artifacts from GitHub Release
• setup-iam-permissions - Example module to setup permission boundaries

ARM64 configuration for submodules

When not using the top-level module and specifying an a1 or *6g* (6th-gen Graviton2) instance_type, the runner-binaries-syncer and runners submodules need to be configured appropriately for pulling the ARM64 GitHub action runner binary and leveraging the arm64 AMI for the runners.

When configuring runner-binaries-syncer

• runner_architecture - set to arm64, defaults to x64

When configuring runners

• ami_filter - set to ["amzn2-ami-hvm-2*-arm64-gp2"], defaults to ["amzn2-ami-hvm-2.*-x86_64-ebs"]

Debugging

In case the setup does not work as intended follow the trace of events:

• In the GitHub App configuration, the Advanced page displays all webhook events that were sent.
• In AWS CloudWatch, every lambda has a log group. Look at the logs of the webhook and scale-up lambdas.
• In AWS SQS you can see messages available or in flight.
• Once an EC2 instance is running, you can connect to it in the EC2 user interface using Session Manager. Check the user data script using cat /var/log/user-data.log. By default several log files of the instances are streamed to AWS CloudWatch, look for a log group named <environment>/runners. In the log group you should see at least the log streams for the user data installation and runner agent.
• Registered instances should show up in the Settings - Actions page of the repository or organization (depending on the installation mode).

Requirements

No requirements.

Providers

Inputs

[
  "amazon"
]

object({
    key_base64     = string
    id             = string
    client_id      = string
    client_secret  = string
    webhook_secret = string
  })

list(object({
    cron      = string
    timeZone  = string
    idleCount = number
  }))

list(object({
    log_group_name   = string
    prefix_log_group = bool
    file_path        = string
    log_stream_name  = string
  }))

[
  {
    "file_path": "/var/log/messages",
    "log_group_name": "messages",
    "log_stream_name": "{instance_id}",
    "prefix_log_group": true
  },
  {
    "file_path": "/var/log/user-data.log",
    "log_group_name": "user_data",
    "log_stream_name": "{instance_id}",
    "prefix_log_group": true
  },
  {
    "file_path": "/home/ec2-user/actions-runner/diag/Runner**.log",
    "log_group_name": "runner",
    "log_stream_name": "{instance_id}",
    "prefix_log_group": true
  }
]

Outputs

Contribution

We welcome contribution, please checkout the contribution guide. Be-aware we use pre commit hooks to update the docs.

Philips Forest

This module is part of the Philips Forest.

                                                     ___                   _
                                                    / __\__  _ __ ___  ___| |_
                                                   / _\/ _ \| '__/ _ \/ __| __|
                                                  / / | (_) | | |  __/\__ \ |_
                                                  \/   \___/|_|  \___||___/\__|

                                                                 Infrastructure

Talk to the forestkeepers in the forest-channel on Slack.