• Introduction
• Prerequisites
  • AWS security groups
  • AWS credentials
• Prepare the Runner Manager instance
• Registering the GitLab Runner
• Configuring the GitLab Runner
  • The global section
  • The runners section
  • The runners.docker section
  • The runners.cache section
  • The runners.machine section
  • Getting it all together
• Cutting down costs with Amazon EC2 Spot instances
  • Caveats of Spot instances
• Conclusion

Autoscaling GitLab Runner on AWS EC2

One of the biggest advantages of GitLab Runner is its ability to automatically spin up and down VMs to make sure your builds get processed immediately. It’s a great feature, and if used correctly, it can be extremely useful in situations where you don’t use your Runners 24/7 and want to have a cost-effective and scalable solution.

Introduction

In this tutorial, we’ll explore how to properly configure a GitLab Runner in AWS that will serve as the Runner Manager where it will spawn new Docker machines on demand.

In addition, we’ll make use of Amazon’s EC2 Spot instances which will greatly reduce the costs of the Runner instances while still using quite powerful autoscaling machines.

Prerequisites

Your GitLab instance is going to need to talk to the Runners over the network, and that is something you need think about when configuring any AWS security groups or when setting up your DNS configuration.

For example, you can keep the EC2 resources segmented away from public traffic in a different VPC to better strengthen your network security. Your environment is likely different, so consider what works best for your situation.

AWS security groups

Docker Machine will attempt to use a default security group with rules for port 2376 and SSH 22, which is required for communication with the Docker daemon. Instead of relying on Docker, you can create a security group with the rules you need and provide that in the Runner options as we will see below. This way, you can customize it to your liking ahead of time based on your networking environment. You have to make sure that ports 2376 and 22 are accessible by the Runner Manager instance.

AWS credentials

You’ll need an AWS Access Key tied to a user with permission to scale (EC2) and update the cache (via S3). Create a new user with policies for EC2 (AmazonEC2FullAccess) and S3 (AmazonS3FullAccess). To be more secure, you can disable console login for that user. Keep the tab open or copy paste the security credentials in an editor as we’ll use them later during the Runner configuration.

Prepare the Runner Manager instance

The first step is to install GitLab Runner in an EC2 instance that will serve as the Runner Manager that spawns new machines. This doesn’t have to be a powerful machine since it will not run any jobs itself, a t2.micro instance will do. This machine will be a dedicated host since we need it always up and running, thus it will be the only standard cost.

Install the prerequisites:

1. Log in to your server
2. Install GitLab Runner from the official GitLab repository
3. Install Docker
4. Install Docker Machine

Now that the Runner is installed, it’s time to register it.

Registering the GitLab Runner

Before configuring the GitLab Runner, you need to first register it, so that it connects with your GitLab instance:

1. Obtain a Runner token
2. Register the Runner
3. When asked the executor type, enter docker+machine

You can now move on to the most important part, configuring the GitLab Runner.

Configuring the GitLab Runner

Now that the Runner is registered, you need to edit its configuration file and add the required options for the AWS machine driver.

Let’s first break it down to pieces.

The global section

In the global section, you can define the limit of the jobs that can be run concurrently across all Runners (concurrent). This heavily depends on your needs, like how many users your Runners will accommodate, how much time your builds take, etc. You can start with something low like 10, and increase or decrease its value going forward.

The check_interval option defines how often the Runner should check GitLab for new jobs, in seconds.

Example:

concurrent = 10
check_interval = 0

Read more about all the options you can use.

The runners section

From the [[runners]] section, the most important part is the executor which must be set to docker+machine. Most of those settings are taken care of when you register the Runner for the first time.

limit sets the maximum number of machines (running and idle) that this Runner will spawn. For more information, check the relationship between limit, concurrent and IdleCount.

Example:

[[runners]]
  name = "gitlab-aws-autoscaler"
  url = "<URL of your GitLab instance>"
  token = "<Runner's token>"
  executor = "docker+machine"
  limit = 20

Read more about all the options you can use under [[runners]].

The runners.docker section

In the [runners.docker] section you can define the default Docker image to be used by the child Runners if it’s not defined in .gitlab-ci.yml. By using privileged = true, all Runners will be able to run Docker in Docker which is useful if you plan to build your own Docker images via GitLab CI/CD.

Next, we use disable_cache = true to disable the Docker executor’s inner cache mechanism since we will use the distributed cache mode as described in the following section.

Example:

  [runners.docker]
    image = "alpine"
    privileged = true
    disable_cache = true

Read more about all the options you can use under [runners.docker].

The runners.cache section

To speed up your jobs, GitLab Runner provides a cache mechanism where selected directories and/or files are saved and shared between subsequent jobs. While not required for this setup, it is recommended to use the distributed cache mechanism that GitLab Runner provides. Since new instances will be created on demand, it is essential to have a common place where the cache is stored.

In the following example, we use Amazon S3:

  [runners.cache]
    Type = "s3"
    Shared = true
    [runners.cache.s3]
      ServerAddress = "s3.amazonaws.com"
      AccessKey = "<your AWS Access Key ID>"
      SecretKey = "<your AWS Secret Access Key>"
      BucketName = "<the bucket where your cache should be kept>"
      BucketLocation = "us-east-1"

Here’s some more information to further explore the cache mechanism:

• Reference for runners.cache
• Reference for runners.cache.s3
• Deploying and using a cache server for GitLab Runner
• How cache works

The runners.machine section

This is the most important part of the configuration and it’s the one that tells GitLab Runner how and when to spawn new or remove old Docker Machine instances.

We will focus on the AWS machine options, for the rest of the settings read about the:

• Autoscaling algorithm and the parameters it’s based on - depends on the needs of your organization
• Autoscaling periods - useful when there are regular time periods in your organization when no work is done, for example weekends

Here’s an example of the runners.machine section:

  [runners.machine]
    IdleCount = 1
    IdleTime = 1800
    MaxBuilds = 10
    MachineDriver = "amazonec2"
    MachineName = "gitlab-docker-machine-%s"
    MachineOptions = [
      "amazonec2-access-key=XXXX",
      "amazonec2-secret-key=XXXX",
      "amazonec2-region=us-central-1",
      "amazonec2-vpc-id=vpc-xxxxx",
      "amazonec2-subnet-id=subnet-xxxxx",
      "amazonec2-zone=x",
      "amazonec2-use-private-address=true",
      "amazonec2-tags=runner-manager-name,gitlab-aws-autoscaler,gitlab,true,gitlab-runner-autoscale,true",
      "amazonec2-security-group=xxxxx",
      "amazonec2-instance-type=m4.2xlarge",
    ]
    [[runners.machine.autoscaling]]
      Periods = ["* * 9-17 * * mon-fri *"]
      IdleCount = 50
      IdleTime = 3600
      Timezone = "UTC"
    [[runners.machine.autoscaling]]
      Periods = ["* * * * * sat,sun *"]
      IdleCount = 5
      IdleTime = 60
      Timezone = "UTC"

The Docker Machine driver is set to amazonec2 and the machine name has a standard prefix followed by %s (required) that is replaced by the ID of the child Runner: gitlab-docker-machine-%s.

Now, depending on your AWS infrastructure, there are many options you can set up under MachineOptions. Below you can see the most common ones.

Read more about all the options you can use under [runners.machine].

Getting it all together

Here’s the full example of /etc/gitlab-runner/config.toml:

concurrent = 10
check_interval = 0

[[runners]]
  name = "gitlab-aws-autoscaler"
  url = "<URL of your GitLab instance>"
  token = "<Runner's token>"
  executor = "docker+machine"
  limit = 20
  [runners.docker]
    image = "alpine"
    privileged = true
    disable_cache = true
  [runners.cache]
    Type = "s3"
    Shared = true
    [runners.cache.s3]
      ServerAddress = "s3.amazonaws.com"
      AccessKey = "<your AWS Access Key ID>"
      SecretKey = "<your AWS Secret Access Key>"
      BucketName = "<the bucket where your cache should be kept>"
      BucketLocation = "us-east-1"
  [runners.machine]
    IdleCount = 1
    IdleTime = 1800
    MaxBuilds = 100
    [[runners.machine.autoscaling]]
      Periods = ["* * 9-17 * * mon-fri *"]
      IdleCount = 50
      IdleTime = 3600
      Timezone = "UTC"
    [[runners.machine.autoscaling]]
      Periods = ["* * * * * sat,sun *"]
      IdleCount = 5
      IdleTime = 60
      Timezone = "UTC"
    MachineDriver = "amazonec2"
    MachineName = "gitlab-docker-machine-%s"
    MachineOptions = [
      "amazonec2-access-key=XXXX",
      "amazonec2-secret-key=XXXX",
      "amazonec2-region=us-central-1",
      "amazonec2-vpc-id=vpc-xxxxx",
      "amazonec2-subnet-id=subnet-xxxxx",
      "amazonec2-use-private-address=true",
      "amazonec2-tags=runner-manager-name,gitlab-aws-autoscaler,gitlab,true,gitlab-runner-autoscale,true",
      "amazonec2-security-group=docker-machine-scaler",
      "amazonec2-instance-type=m4.2xlarge",
    ]

Cutting down costs with Amazon EC2 Spot instances

As described by Amazon:

Amazon EC2 Spot instances allow you to bid on spare Amazon EC2 computing capacity. Since Spot instances are often available at a discount compared to On-Demand pricing, you can significantly reduce the cost of running your applications, grow your application’s compute capacity and throughput for the same budget, and enable new types of cloud computing applications.

In addition to the runners.machine options you picked above, in /etc/gitlab-runner/config.toml under the MachineOptions section, add the following:

    MachineOptions = [
      "amazonec2-request-spot-instance=true",
      "amazonec2-spot-price=",
    ]

In this configuration with an empty amazonec2-spot-price, AWS sets your bidding price for a Spot instance to the default On-Demand price of that instance class. If you omit the amazonec2-spot-price completely, Docker Machine will set the maximum price to a default value of $0.50 per hour.

You may further customize your Spot instance request:

    MachineOptions = [
      "amazonec2-request-spot-instance=true",
      "amazonec2-spot-price=0.03",
      "amazonec2-block-duration-minutes=60"
    ]

With this configuration, Docker Machines are created using Spot instances with a maximum Spot request price of $0.03 per hour and the duration of the Spot instance is capped at 60 minutes. The 0.03 number mentioned above is just an example, so be sure to check on the current pricing based on the region you picked.

To learn more about Amazon EC2 Spot instances, visit the following links:

• https://aws.amazon.com/ec2/spot/
• https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/spot-requests.html
• https://aws.amazon.com/ec2/spot/getting-started/

Caveats of Spot instances

While Spot instances is a great way to use unused resources and minimize the costs of your infrastructure, you must be aware of the implications.

Running CI jobs on Spot instances may increase the failure rates because of the Spot instances pricing model. If the maximum Spot price you specify exceeds the current Spot price you will not get the capacity requested. Spot pricing is revised on an hourly basis. Any existing Spot instances that have a maximum price below the revised Spot instance price will be terminated within two minutes and all jobs on Spot hosts will fail.

As a consequence, the auto-scale Runner would fail to create new machines while it will continue to request new instances. This eventually will make 60 requests and then AWS won’t accept any more. Then once the Spot price is acceptable, you are locked out for a bit because the call amount limit is exceeded.

If you encounter that case, you can use the following command in the Runner Manager machine to see the Docker Machines state:

docker-machine ls -q --filter state=Error --format "{{.NAME}}"

Conclusion

In this guide we learned how to install and configure a GitLab Runner in autoscale mode on AWS.

Using the autoscale feature of GitLab Runner can save you both time and money. Using the Spot instances that AWS provides can save you even more, but you must be aware of the implications. As long as your bid is high enough, there shouldn’t be an issue.

You can read the following use cases from which this tutorial was (heavily) influenced:

• HumanGeo - Scaling GitLab CI
• Substrakt Health - Autoscale GitLab CI Runners and save 90% on EC2 costs