Kubernetes on Business Information Technology Services

Arden University

Sat, 01 Jan 2022 00:00:00 +0000

For Arden, there were two projects to deliver: an Amazon Elastic Kubernetes Service (EKS) based platform and development of their serverless integrations.

In order to help with the delivery of these projects, some open source contributions were made.

Elastic Kubernetes Service (EKS) Platform

A new platform was required to migrate existing on-premises and Amazon EC2 workloads to. The new platform had to allow developers to easily manage their own deployments, allow for centralised logging and monitoring, had to be secure and allow for management of things such as DNS records and TLS certificates to be fully automated.

The platform had to be delivered ‘as code’ so it could be easily managed, upgraded and replicated. This was done using mainly Terraform with some base Kubernetes services being deployed using the Terraform Helm provider.

Any base Kubernetes services that ArgoCD did not depend on explicitly were managed and deployed by ArgoCD along with the business applications hosted on the platform.

Managing Deployments with ArgoCD

The main interface into the platform for developers was through ArgoCD. This allowed developers to deploy applications, see the status of their workloads and get immediate logs and events to diagnose failures. Being able to deploy and diagnose simple issues themselves allowed developers to spend less time being blocked waiting for infrastructure engineers to help out.

ArgoCD was integrated with Microsoft Entra to allow single-sign on which improves security and convenience. Developers could be assigned to individual projects meaning they only have access to the applications that they are working with.

ArgoCD Image Updater was also used in development environments so that successful builds of an application could be deployed automatically.

Monitoring & Logging

For the monitoring and logging, open source tools were used as they provide excellent flexibility and of course value.

For monitoring, Prometheus was used with Grafana. Prometheus was also configured to make a future expansion to Thanos as easy as possible, should the need ever arise.

Loki was used for logging, it integrated perfectly with the monitoring metrics since it also uses Grafana as the main user interface.

Alerts were configured with Grafana unified alerts, allowing alerts to be raised from either metric or logging data.

Automated Management

Tools such as Certificate Manager and External DNS were used to allow automation of day to day administrative tasks.

Certificate Manager automatically provisions and renews TLS certificates for applications hosted within the platform whilst External DNS creates and cleans up DNS records.

Future Ready

Whilst not immediately implemented due to possible conflicts with the business the platform was developed and tested with IPv6 support and support for ARM based compute allowing for easy adoption later on.

Serverless Integrations Development

Arden was implementing a new student record system called SITS. This was a huge project that affected all areas of the business. A significant problem was integrating existing systems that do not understand SITS, such as: the virtual learning environment, Active Directory, time tabling products and many more.

To enable this communication between systems, a group of serverless ‘middlewares’ were used to keep things scalable and event driven.

Since the integrated systems spanned across AWS and Azure as well as including software as a service (SaaS) solutions, many technologies were used to integrate them. These technologies included: Typescript based Lambda functions deployed with the Amazon CDK, Azure Logic Apps, Azure Functions, Amazon SNS, Amazon SQS and Azure Service Bus.

These integrations handled large amounts of traffic especially as SITS went live at Arden. They also helped the business implement complex processes with systems that cannot be directly integrated.

Open Source Contributions

The following open source contributions were made whilst delivering these projects:

Drivvn

Thu, 02 Jul 2020 00:00:00 +0000

A project to pick up support for and improve an existing Kubernetes based infrastructure platform hosted on Microsoft Azure.

Skip to Result

Stabilising Platform

Initially, the primary focus had been to automate continual time consuming tasks and to reduce the number of support tickets. This has been achieved by implementing technologies such as External DNS and Certificate Manager. Reliability has been improved by consolidating and improving complex build pipelines by utilising YAML libraries for Azure DevOps and introducing infrastructure as code with Terraform.

Evolving Platform

After stabilising the platform focus shifted to evolving the platform to be more performant, scalable and developer friendly. In addition to everything outlined below, lots of work went into upgrades of Kubernetes clusters and services, improvements as and when they were identified and migrations to new Azure subscriptions/tenants.

Developer Friendly

To help improve continuous delivery, ArgoCD was implemented. This allowed developers to publish Docker images in their build pipelines which would then automatically be picked up and deployed into a Kubernetes environment. ArgoCD also allowed the configuration for the application to be completely separated from other infrastructure as code meaning developers could have full control over the configuration of their deployed applications without requiring permissions to alter other parts of infrastructure.

Infrastructure as Code

There was a large focus on infrastructure as code. Terraform was used to automate the creation of all base infrastructure: such as resource groups, virtual networks, security groups and Kubernetes clusters - including all supporting infrastructure and Active Directory objects. This allowed much quicker and reliable provision of core infrastructure.

Any new services deployed were also fully automated so that required databases, Kubernetes permissions and the application themselves could also be deployed in a reliable and repeatable way.

Monitoring & Logging

As the platform grew it became important to be able to monitor workloads. Initially Prometheus was used with Grafana as well as some other custom exporters to get certain metrics that were otherwise unavailable.

However as things grew Prometheus was not scaling so eventually the monitoring stack was upgraded with Thanos. This allowed gathering of huge amounts of metrics from multiple Prometheus instances and storing the data in cost efficient blob storage. Thanos also allowed a ‘single pane of glass’ view of the entire estate of multiple clusters through Grafana.

Centralised logging was also set up utilising ElasticSearch at first and later New Relic.

Ingress

A lot of work was carried out around ingress to the platform. Initially cluster ingresses were standardised with Ingress Nginx. Security and CDN capability was then added using Cloudflare and Cloudflare Access. This allowed many apps to use single sign-on that originally did not support it. It also enabled zero trust access to internal services without having to set up and maintain a VPN.

Result

Due to multiple reasons - the good work that Drivvn do, the pandemic and just the way things are going - significantly more people are now buying cars online. Making the infrastructure more scalable and resilient has helped make it possible to meet these suddenly increasing demands.

Having the infrastructure as code ready allowed us to work with one of the Drivvn development teams to deliver a new product from inception to launch within a few weeks. The initial infrastructure was provisioned and available within hours. We then worked with the team to refine the build pipelines and infrastructure as they developed the product. Since launch, the infrastructure supporting this product has been extremely reliable. Having ArgoCD set up allowed developers to manage their own configuration and ‘self-serve’ deployments all the way to production - with easy roll backs if required.

Using confd to Inject Secrets into Kubernetes Pods

Mon, 06 Aug 2018 19:22:56 +0100

Introduction

Whilst using Kubernetes over the past few months, one challenge I repeatedly faced was to get secrets - such as passwords, SSH keys or certificate keys - securely into applications running on Kubernetes.

Whilst this is quite easy if the container image is under your full control, to achieve this with an ‘off the shelf’ image is a little more tricky.

One tool I came across recently was confd - which has helped a lot with this challenge and below I will outline how.

confd Basics

confd is a tool for rendering configuration files from predefined templates using values (secrets) that are stored in a backend. A backend could be etcd, Amazon SSM Parameter Store, Hashicorp Vault or many others.

The examples below will be using the Amazon SSM Parameter Store backend. For Kubernetes clusters running on AWS this works really well as Amazon IAM roles can be used, mitigating the use for storing the backend password anywhere.

I won’t go into too much detail on the basics of confd. I would recommend you look at the below links:

confd Image

First a Docker container that has confd available will be required. At the time of writing I could not find an official image available - so baked my own. This should be a simple and small image, based on something like Alpine Linux with only confd installed and not much else.

Since creating this guide, many better ways of handling secrets in Kubernetes are available such as CSI drivers and the use of Terraform with Kubernetes secrets. Due to this and the fact confd has not had a release since 2018 the pre-baked images are no longer available.

Example 1: Injecting Secrets into Environment

Many ‘off the shelf’ images allow for loading secrets from environment variables. One example of this is Grafana.

Starting Example

Lets start with injecting secrets as simply as possible - plain text in the deployment spec:

- name: grafana
  imagePullPolicy: IfNotPresent
  image: "grafana/grafana:latest"
  env:
  - name: GF_SECURITY_ADMIN_USER
    value: admin
  - name: GF_SECURITY_ADMIN_PASSWORD
    value: supersecurepassword123

We want the Grafana image to get the secrets above, by itself, without having to manage the Grafana image ourselves.

Add the Secrets to Amazon SSM Parameter Store

Add the two secrets to the Amazon SSM Parameter store using the AWS console.

/grafana-username: the Grafana administrators username
/grafana-password: the Grafana administrators password

Create an Amazon IAM Role

To allow the containers to access the SSM Parameters, they need to be granted access by IAM.

In addition, access to decrypt using the Amazon KMS key used to encypt the parameters will also need to be granted.

Example (do not copy and paste!):

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Action": [
                "ssm:GetParameter"
            ],
            "Resource": [
              "arn:aws:ssm:eu-west-1:123456123:parameter/grafana-username",
              "arn:aws:ssm:eu-west-1:123456123:parameter/grafana-password"
            ]
        },
        {
            "Effect": "Allow",
            "Action": [
                "kms:Decrypt"
            ],
            "Resource": "arn:aws:kms:eu-west-1:111122223333:key/1234abcd-12ab-34cd-56ef-1234567890ab"
        }
    ]
}

confd Configurations

confd uses TOML configuration files to define what you want it to process. The below TOML will process template grafana.env.tmpl (defined later) and put the output in /shared-config/grafana.env with mode 0400. As Grafana by default runs as UID:GID 472:472 we make sure the environment file is owned by the same user & group.

apiVersion: v1
kind: ConfigMap
metadata:
  name: grafana-confd-configs
  namespace: monitoring
data:
  grafana.env.toml: |
    [template]
    src    = "grafana.env.tmpl"
    dest   = "/shared-config/grafana.env"
    uid    = 472
    gid    = 472
    mode   = "0400"
    keys   = [
      "/grafana-username",
      "/grafana-password"
    ]

confd Templates

The templates are the configuration files to render. As we want to set environment variables, the following template works well:

apiVersion: v1
kind: ConfigMap
metadata:
  name: grafana-confd-templates
  namespace: monitoring
data:
  grafana.env.tmpl: |
    export GF_SECURITY_ADMIN_USER="{{     getv "/grafana-username" }}"
    export GF_SECURITY_ADMIN_PASSWORD="{{ getv "/grafana-password" }}"

Override Launcher

A new launcher script should be created for the main container (Grafana in this example). The new launcher script should import rendered environment variables and then start the original entrypoint script.

Please note the following:

Always use . instead of source. A lot of containers do not have a full bash shell.
Always exec to start the original entry point - so that it remains as PID 1.
/run.sh is the original entry point of the Grafana image.
Make sure "${@}" is passed to the original entry point, so arguments still work.

To find the original entry point of an image, download the image with docker pull and then use docker inspect to find the entry point.

apiVersion: v1
kind: ConfigMap
metadata:
  name: grafana-launcher
  namespace: monitoring
data:
  launcher.sh: |
    #!/bin/bash -e
    ###############################################################################

    echo ":: Loading extra environment variables..."
    . "/shared-config/grafana.env"

    ###############################################################################

    echo ":: Launching Grafana..."
    exec "/run.sh" "${@}"

    ###############################################################################

Modifying the Deployment

The final step is to make the Grafana deployment run confd based on the supplied configuration before Grafana is started. To do that we use an initContainer.

Notice:

We set the region to eu-west-1, but you need to set this to the region your SSM parameters are stored.
We have three volume mounts:
- grafana-shared-config is a shared emptyDir volume for the main Grafana container and the confd initContainer.
- grafana-confd-configs will refer to the confd configurations configuration map defined above.
- grafana-confd-templates will refer to the confd templates configuration map defined above.

initContainers:
  - name: grafana-confd
    image: "rlees85/secrets-loader:latest"
    command: [ 'confd', '-onetime', '-backend', 'ssm' ]
    env:
      - name: AWS_DEFAULT_REGION
        value: eu-west-1
    volumeMounts:
      - name: grafana-shared-config
        mountPath: /shared-config
      - name: grafana-confd-configs
        mountPath: /etc/confd/conf.d
      - name: grafana-confd-templates
        mountPath: /etc/confd/templates

The grafana-shared-config and grafana-launcher mount should be added to the main Grafana container.

- name: grafana-shared-config
  mountPath: /shared-config
- name: grafana-launcher
  mountPath: /launcher

All volumes should be correctly defined in the deployment. Please note that in the particular deployment used in this example grafana-config was already present.

volumes:
- name: grafana-config
  configMap:
    name: grafana
- name: grafana-shared-config
  emptyDir: {}
- name: grafana-confd-configs
  configMap:
    defaultMode: 0400
    name: grafana-confd-configs
- name: grafana-confd-templates
  configMap:
    defaultMode: 0400
    name: grafana-confd-templates
- name: grafana-launcher
  configMap:
    defaultMode: 0500
    name: grafana-launcher

The initContainer needs access to Amazon SSM Parameter store. Make sure kube2iam is configured on the Kubernetes Cluster and add the appropriate annotation to the Grafana deployment.

annotations:
  iam.amazonaws.com/role: grafana

Finally, we can override the Grafana containers start-up command to use the new launcher script:

- name: grafana
  imagePullPolicy: IfNotPresent
  image: "grafana/grafana:latest"
  command: [ '/launcher/launcher.sh' ]

Conclusion

The confd initContainer now runs before Grafana starts and outputs the templated secrets to shared storage. The main Grafana container then sources these secrets from shared storage before running the original image entry point.

$ kubectl -n monitoring logs grafana-7646488856-4f4gx -c grafana-confd
2018-08-06T19:34:55Z grafana-7646488856-4f4gx confd[1]: INFO Backend set to ssm
2018-08-06T19:34:55Z grafana-7646488856-4f4gx confd[1]: INFO Starting confd
2018-08-06T19:34:55Z grafana-7646488856-4f4gx confd[1]: INFO Backend source(s) set to
2018-08-06T19:34:56Z grafana-7646488856-4f4gx confd[1]: INFO Target config /shared-config/grafana.env out of sync
2018-08-06T19:34:56Z grafana-7646488856-4f4gx confd[1]: INFO Target config /shared-config/grafana.env has been updated

$ kubectl -n monitoring logs grafana-7646488856-4f4gx
t=2018-08-06T19:35:07+0000 lvl=info msg="Starting Grafana" logger=server version=5.2.1 commit=2040f61 compiled=2018-06-29T09:17:46+0000

...

t=2018-08-06T19:35:07+0000 lvl=info msg="Config overridden from Environment variable" logger=settings var="GF_SECURITY_ADMIN_USER=admin"
t=2018-08-06T19:35:07+0000 lvl=info msg="Config overridden from Environment variable" logger=settings var="GF_SECURITY_ADMIN_PASSWORD=*********"

...

Example 2: Rendering Configuration Files and/or Keys

Please read through example 1 first. A lot of things will not be covered again and are assumed to be already set up (SSM parameters, KMS keys and IAM permissions).

In this example we have a much more complicated application, that requires secrets to be loaded into its configuration files. Additionally, the application integrates with other services - and therefore needs an SSH private key to be injected at run-time.

Starting Example

In this example, the deployment spec has no secrets in. The secrets are baked directly in to the image. This may be undesirable for example if the image has to pass through a pipeline - developers perhaps should not have access to production secrets.

Let’s say the following file is baked directly into the image:

$ cat /etc/application.d/50-config.properties
mysql.db.username=application
mysql.db.password=application123
integration.ssh-key=/etc/application/ssh.pem

NOTE: When loading multi-line parameters (such as SSH keys) into Amazon SSM Parameter store use the CLI tool and not the console! If the console is used new lines are lost.

confd Configurations

In a similar fashion to the first example, we template a configuration file and SSH key based on templates to shared storage.

apiVersion: v1
kind: ConfigMap
metadata:
  name: application-confd-configs
data:
  99-secrets.properties.toml: |
    [template]
    src    = "99-secrets.properties.tmpl"
    dest   = "/shared-config/99-secrets.properties"
    mode   = "0400"
    keys   = [
      "/application-db-username",
      "/application-db-password"
    ]    
  integration-key.pem.toml: |
    [template]
    src    = "integration-key.pem.tmpl"
    dest   = "/shared-config/integration-key.pem"
    mode   = "0400"
    keys   = [
      "/application-integration-key"
    ]

confd Templates

As before, the templates referred to by the TOML configurations are defined below:

apiVersion: v1
kind: ConfigMap
metadata:
  name: application-confd-templates
data:
  99-secrets.properties.tmpl: |
    mysql.db.username={{ getv "/application-db-username" }}
    mysql.db.password={{ getv "/application-db-password" }}
    integration.ssh-key=/shared-config/integration-key.pem    
  integration-key.pem.tmpl: |
    {{ getv "/application-integration-key" }}

Override Launcher

The same as the first example a new launcher script should be created for the main container. The script should import rendered configuration files into a folder that the application can pick them up.

The extra configuration file already points to the rendered SSH key so no further action is required for the key.

apiVersion: v1
kind: ConfigMap
metadata:
  name: application-launcher
  namespace: monitoring
data:
  launcher.sh: |
    #!/bin/bash -e
    ###############################################################################

    echo ":: Loading extra configuration files..."
    find "/shared-config" -maxdepth 1 -type f -name "*.properties" -exec cp -sfv {} "/etc/application.d/" \;

    ###############################################################################

    echo ":: Launching Application..."
    exec "/opt/startup/startup.sh" "${@}"

    ###############################################################################

If the image you are working with does not have find there are many other ways to achieve the same thing.

Modifying the Deployment

The deployment needs to be modified the same way as in example 1 above.

Conclusion

This shows even complicated configurations can be setup with confd whilst still using off-the-shelf images.

Eclipse Group Solutions

Mon, 06 Nov 2017 00:00:00 +0000

A very exciting ‘greenfield’ project creating a new infrastructure platform using Amazon Web Services, Docker and Kubernetes for SAP Commerce based e-commerce websites.

It had been a great pleasure working with the team at Eclipse whilst delivering this project.

Skip to Result

Project Breifing

The project aim was to create an Amazon Web Services based, fully automated infrastructure platform to host SAP Commerce e-commerce websites. The platform must be constructed in a way that it can be used on other cloud service providers with little effort later on.

Developers and testers needed the ability to get code and features through testing pipelines much quicker. In addition, production websites need to handle peak demand seamlessly.

As well as delivering a platform, existing teams were to be assisted with picking up new tools, technologies and concepts to enable on-going support of the platform.

Solution

The final solution comprised of many components. These components are outlined below.

Infrastructure as Code

Writing infrastructure as code was key to this solution. This is what enabled infrastructure to be provisioned in a reliable and repeatable way at the click of a button. By taking advantage of Terraform module sources it was possible to define a collection of infrastructure objects (such as subnets, route tables, gateways) in a single place but allow variables (such as name, CIDR ranges) to be passed in depending on the environment being built. As a result all infrastructure met defined standards, human error is vastly reduced and development/production parity was achieved.

By taking advantage of Amazon Web Services availability zones and infrastructure as code, all production environments were highly available and could withstand the loss of an Amazon data centre without any downtime.

Immutable Software

To allow software (SAP Commerce in particular) to run in a dynamically scaled environment there were a few challenges to overcome. There needed to be a way to start SAP Commerce very quickly in ‘scale-up’ situations additionally the state of any running SAP Commerce instances had to be externalised in case of scale-down.

Docker was chosen to containerise software. Containerisation enabled the application and all dependencies, configurations, etc to be packed into an image that can be started very quickly. By using SAP Commerce ‘aspects’ a single Docker image can be capable of running in multiple environments in multiple modes. A single image can be promoted all the way through the testing pipeline just by using tags.

To make running containers stateless, services such as Amazon S3 and EFS were used.

Service Orchestration

To enable service-level auto scaling, auto healing, multi-tenant clustering and service health checks, Kubernetes was implemented. This allowed the platform to meet the scaling requirements. The healing features of Kubernetes allowed the platform to be more resilient to virtual machine failure or network outages resulting in a higher service availability.

Technologies & Tools

Multiple technologies and tools were used to deliver this project.

Amazon Web Services

The initial cloud provider chosen was Amazon. Amazon Web Services (AWS) is a mature cloud service with endless offerings. It is effortless to build highly-available and highly-performant infrastructure stacks. Most tools (such as Terraform and Kops below) offer excellent support for AWS. By using the many AWS services available such as IAM and availability zones it was possible to create secure and resilient infrastructure.

Hashicorp Terraform

Terraform was chosen to build foundation and networking infrastructure. Terraform has excellent Amazon Web Services support and code can easily be ported to work with other cloud providers too - including OpenStack for managing resources on-premises. Terraform made it very easy to meet the project requirements to stay cloud agnostic and to fully automate infrastructure.

Docker

Containerisation was chosen to help simplify development. Rather than pushing a codebase that may have a complete different set of steps to deploy depending on target environment, containerisation makes that a single image that may take the target environment as a parameter. Additionally, all dependencies and libraries required to run the application are packed into the container, meaning that the same container can be ran locally, on tin or in the cloud with minimal effort. Docker was chosen specifically due to it being well proven and mature.

Kubernetes

Kubernetes is fast becoming an industry standard for container orchestration. It can also run on any cloud provider or even on-premises on tin. Native support for service-level auto-scaling and ‘cluster autoscaler’ add-on for the scaling of underlying virtual machines allowed scaling requirements to be met. The many different types of services (deployments, statefulsets, daemonsets), specifications (disruption budgets, affinities) and probes (liveness, readiness) made it possible to build a platform that is resilient against hypervisor, network or even data centre failures.

Other Tools

Other tools used to deliver this project include: Kops, Hasicorp Packer and Ansible.

Result

The result of this project allows Eclipse to offer their customers an improved hosting service. Development environments can now be provisioned rapidly and easily decommissioned when not in use to match project demands. Automatic scaling allows customers websites to seamlessly handle high load during sales and events whilst running economically during quiet periods. The platform allows for zero-downtime code deployments and platform updates. High availability is achieved by always running across multiple availability zones (data centres) in addition to having automatic health checks and repairs.