terraform-aws-jenkins-ha-agents

A module for deploying Jenkins in a highly available and highly scalable manner.

Related blog post can be found on the Neiman Marcus Medium page.

Features

Highly available architecture with agents and master in an autoscaling group
EFS volume used for master node persistence
Jenkins versions incremented through variable
Complete Infrastructure as code deployment, no plugin configuration required
Spot instance pricing for agents
Custom user data available
Auto update plugins

Terraform & Module Version

Terraform 0.13 - Pin module version to ~> v3.0. Submit pull-requests to master branch.

Terraform 0.12 - Pin module version to ~> v2.0. Submit pull-requests to terraform12 branch. Only bug fixes will be accepted. All new developement will be on Terraform 0.13.

Terraform 0.11 - Deprecated in this module.

Usage

To be used with a local map of tags.

Minimum Configuration

module "jenkins_ha_agents" {
  source  = "neiman-marcus/jenkins-ha-agents/aws"
  version = "x.x.x"

  admin_password  = "foo"
  bastion_sg_name = "bastion-sg"
  domain_name     = "foo.io."

  private_subnet_name = "private-subnet-*"
  public_subnet_name  = "public-subnet-*"

  r53_record = "jenkins.foo.io"
  region     = "us-west-2"

  ssl_certificate = "*.foo.io"
  ssm_parameter   = "/jenkins/foo"

  tags     = local.tags
  vpc_name = "prod-vpc"
}

Full Configuration with Custom Userdata and Plugins

main.tf

module "jenkins_ha_agents" {
  source  = "neiman-marcus/jenkins-ha-agents/aws"
  version = "x.x.x"

  admin_password    = "foo"
  agent_max         = 6
  agent_min         = 2
  agent_volume_size = 16

  ami_name          = "amzn2-ami-hvm-2.0.*-x86_64-gp2"
  ami_owner         = "amazon"
  api_ssm_parameter = "/api_key"

  auto_update_plugins_cron = "0 0 31 2 *"

  efs_mode                   = "provisioned"
  efs_provisioned_throughput = 3

  application     = "jenkins"
  bastion_sg_name = "bastion-sg"
  domain_name     = "foo.io."

  agent_lt_version  = "$Latest"
  master_lt_version = "$Latest"

  key_name          = "foo"
  scale_down_number = -1
  scale_up_number   = 1

  custom_plugins              = data.template_file.custom_plugins.rendered
  extra_agent_userdata        = data.template_file.extra_agent_userdata.rendered
  extra_agent_userdata_merge  = "list(append)+dict(recurse_array)+str()"
  extra_master_userdata       = data.template_file.extra_master_userdata.rendered
  extra_master_userdata_merge = "list(append)+dict(recurse_array)+str()"

  retention_in_days = 90

  executors              = 4
  instance_type          = ["t3a.xlarge", "t3.xlarge", "t2.xlarge"]
  jenkins_version        = "2.249.1"
  password_ssm_parameter = "/admin_password"

  cidr_ingress        = ["0.0.0.0/0"]
  private_subnet_name = "private-subnet-*"
  public_subnet_name  = "public-subnet-*"

  r53_record      = "jenkins.foo.io"
  region          = "us-west-2"
  ssl_certificate = "*.foo.io"

  ssm_parameter = "/jenkins/foo"
  swarm_version = "3.23"
  tags          = local.tags
  vpc_name      = "prod-vpc"
}

data "template_file" "custom_plugins" {
  template = file("init/custom_plugins.cfg")
}

data "template_file" "extra_agent_userdata" {
  template = file("init/extra_agent_userdata.cfg")

  vars {
    foo = "bar"
  }
}

data "template_file" "extra_master_userdata" {
  template = file("init/extra_master_userdata.cfg")

  vars {
    foo = "bar"
  }
}

init/custom_plugins.cfg

---
#cloud-config

write_files:
  - path: /root/custom_plugins.txt
    content: |
      cloudbees-folder
    permissions: "000400"
    owner: root
    group: root

init/extra_agent_userdata.cfg

---
runcmd:
  - echo 'foo = ${foo}'

init/extra_master_userdata.cfg

---
runcmd:
  - echo 'foo = ${foo}'

Examples

Inputs

Name	Description	Type	Default	Required
admin_password	The master admin password. Used to bootstrap and login to the master. Also pushed to ssm parameter store for posterity.	string	`N/A`	yes
agent_lt_version	The version of the agent launch template to use. Only use if you need to programatically select an older version of the launch template. Not recommended to change.	string	`$Latest`	no
agent_max	The maximum number of agents to run in the agent ASG.	int	`6`	no
agent_min	The minimum number of agents to run in the agent ASG.	int	`2`	no
agent_volume_size	The size of the agent volume.	int	`16`	no
ami_name	The name of the amzn2 ami. Used for searching for AMI id.	string	`amzn2-ami-hvm-2.0.*-x86_64-gp2`	no
ami_owner	The owner of the amzn2 ami.	string	`amazon`	no
api_ssm_parameter	The path value of the API key, stored in ssm parameter store.	string	`/api_key`	no
application	The application name, to be interpolated into many resources and tags. Unique to this project.	string	`jenkins`	no
auto_update_plugins_cron	Cron to set to auto update plugins. The default is set to February 31st, disabling this functionality. Overwrite this variable to have plugins auto update.	string	`0 0 31 2 *`	no
bastion_sg_name	The bastion security group name to allow to ssh to the master/agents.	string	`N/A`	yes
cidr_ingress	IP address cidr ranges allowed access to the LB.	string	`["0.0.0.0/0"]`	no
custom_plugins	Custom plugins to install when bootstrapping. Created from a template outside of the module.	string	`empty`	no
domain_name	The root domain name used to lookup the route53 zone information.	string	`N/A`	yes
efs_mode	The EFS throughput mode. Options are bursting and provisioned. To set the provisioned throughput in mibps, configure `efs_provisioned_throughput` variable.	string	`bursting`	no
efs_provisioned_throughput	The EFS provisioned throughput in mibps. Ignored if EFS throughput mode is set to bursting.	int	`3`	no
executors	The number of executors to assign to each agent. Must be an even number, divisible by two.	int	`4`	no
extra_agent_userdata	Extra agent user-data to add to the default built-in. Created from a template outside of the module.	string	`empty`	no
extra_agent_userdata_merge	Control how cloud-init merges custom agent user-data sections.	string	`list(append) + dict(recurse_array) + str()`	no
extra_master_userdata	Extra master user-data to add to the default built-in. Created from a template outside of the module.	string	`empty`	no
extra_master_userdata_merge	Control how cloud-init merges custom master user-data sections.	string	`list(append) + dict(recurse_array) + str()`	no
instance_type	The type of instances to use for both ASG's. The first value in the list will be set as the master instance.	list	`t3a.xlarge, t3.xlarge, t2.xlarge`	no
jenkins_version	The version number of Jenkins to use on the master. Change this value when a new version comes out, and it will update the launch configuration and the autoscaling group.	string	`2.249.1`	no
key_name	SSH Key to launch instances.	string	`null`	no
master_lt_version	The version of the master launch template to use. Only use if you need to programatically select an older version of the launch template. Not recommended to change.	string	`$Latest`	no
password_ssm_parameter	The path value of the master admin passowrd, stored in ssm parameter store.	string	`/admin_password`	no
private_subnet_name	The name prefix of the private subnets to pull in as a data source.	string	`N/A`	yes
public_subnet_name	The name prefix of the public subnets to pull in as a data source.	string	`N/A`	yes
r53_record	The FQDN for the route 53 record.	string	`N/A`	yes
region	The AWS region to deploy the infrastructure too.	string	`N/A`	yes
retention_in_days	How many days to retain cloudwatch logs.	int	`90`	no
scale_down_number	Number of agents to destroy when scaling down.	int	`-1`	no
scale_up_number	Number of agents to create when scaling up.	int	`1`	no
ssl_certificate	The name of the SSL certificate to use on the load balancer.	string	`N/A`	yes
ssm_parameter	The full ssm parameter path that will house the api key and master admin password. Also used to grant IAM access to this resource.	string	`N/A`	yes
swarm_version	The version of swarm plugin to install on the agents. Update by updating this value.	int	`3.23`	no
tags	tags to define locally, and interpolate into the tags in this module.	string	`N/A`	yes
vpc_name	The name of the VPC the infrastructure will be deployed to.	string	`N/A`	yes

Outputs

Name	Description
agent_asg_name	The name of the agent asg. Use for adding to addition outside resources.
agent_iam_role	The agent IAM role attributes. Use for attaching additional iam policies.
master_asg_name	The name of the master asg. Use for adding to addition outside resources.
master_iam_role	The master IAM role name. Use for attaching additional iam policies.
r53_record_fqdn	The fqdn of the route 53 record.

Known Issues/Limitations

N/A

Notes

It can take a decent amount of time for initial master bootstrap (approx 11 minutes on t2.micro).
- This is normal, and sped up by higher instance types.
- After the master is built and EFS is populated with Jenkins installation configuration, master boot times come down considerably.
During initial bootstrapping the master reboots several times.
You should not need to change the admin password in the Jenkins wizard.
- This is done through bootstrapping.
- If you run through the wizard and it asks you to change the admin password, wait a short time and reload the page.
- Do not click 'Continue as admin' on the password reset page, just wait and reload.

Breaking Changes

v2.5.0

Giving custom names to ASG's has been removed. This should only impact external resources created outside of the module.
ASG's no longer rehydrate with launch template/configuration revisions. You will need to manaully rehydrate your ASG's with new instances.
Spot pricing variable has been removed as the agent ASG was moved to launch template, and does not require this parameter (defaults to on-demand max price).
Instance type variable has been changed to a list to accomodate multiple launch template overrides. If you use a non-default value, you will have to change your variable to a list.

v2.1.0

This version of the module pulls all public and private subnets using a wildcard.
- This allows for more than two hardcoded subnets.
- You may have to destroy several resources and create them again, including mount targets.
- As long as you do not delete your EFS volume, there should be no data loss.
CIDR blocks have been consolidated to reduce redundant configuration.

How it works

The architecture, on the surface, is simple, but has a lot of things going on under the hood. Similar to a basic web-application architecture, a load balancer sits in front of the master auto scaling group, which connects directly to the agent autoscaling group.

Master Node Details

The Master node sits in an autoscaling group, using the Amazon Linux 2 AMI. The autoscaling group is set to a minimum and maximum of one instance. The autoscaling group does not scale out or in. It can be in one of two availability zones. It is fronted by an ELB which can control the autoscaling group based on a health check. If port 8080 is not functioning properly, the ELB will terminate the instance.

The name of the master autoscaling group is identical to the master launch configuration. This is intentional. If the launch configuration is updated, the master autoscaling group will be recreated with the new launch configuration.

Data are persisted through an EFS volume, with a mount target in each availability zone.

During initial launch, the master will generate an API key and publish it to SSM Parameter store.

Agent Nodes Details

Agent nodes are also set in an autoscaling group, using the Amazon Linux 2 AMI, set in the same availability zones.

Agents connect to the master node through the Jenkins SWARM plugin. The agents are smart enough to get the master's IP address using the AWS CLI and API key from the parameter store. Agents launch, configure themselves, and connect to the master. If agents cannot connect or get disconnected, the agent will self-terminate, causing the autoscaling group to create a new instance. This helps in the case that the agents launch, and the master has not yet published the API key to the parameter store. After it is published, the agents and master will sync up. If the master is terminated, the agents will automatically terminate.

Agents are spot instances, keeping cost down.

Agent Scaling Details

Agents scale based on CPU, and on the Jenkins build queue. The master node will poll itself to see how many executors are busy and send a CloudWatch metric alarm. If the number of executors available is less than half, then the autoscaling group will scale up. If executors are idle, then the agents will scale down. This is configured in the cloud-init user data.

Updating Jenkins/SWARM Version

To update Jenkins or the SWARM plugin, update the variable in the terraform.tfvars files and redeploy the stack. The master will rebuild with the new version of Jenkins, maintaining configuration on the EFS volume. The agents will redeploy with the new version of SWARM.

Auto Updating Plugins

The master has the ability to check for plugin updates, and automatically install them. By default, this feature is disabled. To enable it, set the auto_update_plugins_cron argument. Finally, it saves the list of plugins, located in /var/lib/jenkins/plugin-updates/archive for further review. You are encouraged to use something like AWS Backup to take daily backups of your EFS volume, and set the cron to a time during a maintenance window.

Diagram

FAQ

Why not use ECS or Fargate?

ECS still requires managing instances with an autoscaling group, in addition to the ECS containers and configuration. Just using autoscaling groups is less management overhead.

Fargate cannot be used with the master node as it cannot currently mount EFS volumes. It is also more costly than spot pricing for the agents.

Why not use a plugin to create agents?

The goal is to completely define the deployment with code. If a plugin is used and configured for agent deployment, defining the solution as code would be more challenging. With the SWARM plugin, and the current configuration, the infrastructure deploys instances, and the instance user data connects. The master is only used for scaling in and out based on executor load.

Possible Improvements

Below are a list of possible improvements identified. Please feel free to develop and test. These may or may not be implemented.

Fargate agents instead of instances
Fargate master with EFS mount
EFS mount helper
Add instance protection to agents actively executing jobs
Add signaling to the master and agent bootstraping process
IAM policy document resources instead of plain json
~~Add the ability to include custom iam policy details from variable inputs~~ / Added in v2.5.0
~~Move towards launch templates instead of launch configuration~~ / Added in v2.5.0

Authors

Raul Dominguez - Project maintenance.

Conduct / Contributing / License

Refer to our contribution guidelines to contribute to this project. See CONTRIBUTING.md.
All contributions must follow our code of conduct. See CONDUCT.md.
This project is licensed under the Apache 2.0 license. See LICENSE.

Acknowledgments

Cloudonaut.io Template - Original cloudformation template, this project is based on.
Special thanks to Clay Danford for the creation and development of this module.

y0zg/terraform-aws-jenkins-ha-agents