Provisioning cloud resources (AWS, GCP, Azure) in Kubernetes

Published in April 2020

Provisioning cloud resources (AWS, GCP, Azure) in Kubernetes

TL;DR: You can create and connect to managed cloud resources from Kubernetes using the Service Catalog, a tool such as Kubeform or cloud-specific operators such as Config Connector and AWS Operator Service.

Table of content

Managing cloud providers' services

Generally, there are two kinds of applications that you can deploy in Kubernetes: stateless and stateful applications.

Stateless applications don't hold any state and are an excellent use case for Kubernetes.

With stateless apps, you can take advantage of horizontal scaling and create copies of your app without worrying about synchronisation.

You can also schedule Pods anywhere in your infrastructure without having particular constraints.

On the contrary, stateful applications present quite a few challenges:

  • Pods should have access to persistent disks which could be located on specific nodes.
  • It is often impractical to move volumes with Terabytes of data to other Nodes. Stateful apps are usually scheduled and respawned on the same Nodes.
  • Stateful apps should not compete for I/O. Workloads should be isolated from noisy neighbours.
  • There should be a backup and retention policy in place.
  • You might need to configure failover and leader election, sharding and rebalancing.

You can be successful at running stateful applications on Kubernetes such as databases or message brokers if you invest the time to learn how to operate them.

But what if you don't have the expertise or time and still need to provide a production-grade database to your teams?

What if you are a team of one and prefer using an external message broker service with a guaranteed Service Level Agreement (SLA)?

A popular option is to leverage managed services from a cloud provider.

Cloud providers such as Amazon Web Services, Google Cloud Platform and Azure offer several managed services such as:

When you need a MySQL database for your app, you don't have to worry about provisioning it from scratch.

You can ask your cloud provider to provision an instance.

Since the cloud vendor offers APIs for its managed services, you can also script creating, updating and deleting databases.

You can use tools such as Terraform or AWS CloudFormation to create the database programmatically when you create an environment for development or production.

The community has contributed several scripts and modules to make it easier to provision databases, messages brokers, etc.

Here's a short selection:

Utilising managed services is convenient because you can focus on delivering your app, and the cloud provider on running a reliable service.

What if you could use the same managed services in your Kubernetes cluster?

If Amazon RDS, Cloud SQL and Azure Database are so convenient, why not use those instead of rolling out and maintaining your database in Kubernetes?

The good news is that you can use managed services in Kubernetes.

However, there is little consensus on how you should provision those resources and make them available to your apps in the cluster.

There are three options you can choose from:

  1. Externally provisioned — you can create the managed resource on your cloud provider and store the details such as connection detail, username and password in Kubernetes as secrets.
  2. Internally provisioned using Custom Resource Definitions (CRDs) — You can map the managed resources to Kubernetes objects and have a component that creates the resources on the cloud provider.
  3. Internally provisioned using infrastructure as code (IaC) — You can script creating the resources. The script is stored and executed in the cluster.

Let's explore the three options in more detail.

Creating Secrets manually

You can provision your managed databases or message brokers manually, or you could use infrastructure-as-code tools such as Terraform or AWS CloudFormation.

And you can pass the details such as username or password to your apps using a Kubernetes Secret.

Let's have a look at an example.

You can provision an Amazon RDS database with:


export DB_USERNAME=master
export DB_PASSWORD=secret99
aws rds create-db-instance \ --allocated-storage 20 \ --db-instance-class db.m1.small \ --db-instance-identifier test-instance \ --engine mysql \ --master-username $DB_USERNAME \ --master-user-password $DB_PASSWORD
aws rds wait db-instance-available \ --db-instance-identifier test-instance
aws rds describe-db-instances \ --db-instance-identifier test-instance

The output of the last command is a JSON payload with the endpoint of the database.


{ "DBInstances": [ { "DBInstanceIdentifier": "test-instance", "DBInstanceClass": "db.m5.large", "Engine": "postgres", "DBInstanceStatus": "available", "MasterUsername": "master", "Endpoint": {
 "Address": "",
 "Port": 5432, "HostedZoneId": "Z29XKXDKYMONMX" } } ]

You can combine username, password, connection string and the rest of the database details in a Kubernetes secret with:


kubectl create secret generic my-secret \ --from-literal=username=master \ --from-literal=password=secret99 \

You can use the values from the Secret in any Pod.

The example below shows how you can inject username, password and connection string as environment variables:


apiVersion: v1
kind: Pod
metadata: name: my-pod
spec: containers: - name: my-container image: app env: - name: DB_USERNAME valueFrom:
 secretKeyRef: name: my-secret key: username
 - name: DB_PASSWORD valueFrom:
 secretKeyRef: name: my-secret key: password
 - name: DB_URL valueFrom:
 secretKeyRef: name: my-secret key: connection_url

Let's have a look at some of the advantages of managing resources externally:

  1. If you've already use managed services, the steps above are familiar. The only change is storing the values in a Kubernetes secret.
  2. It is flexible as you can create the managed services manually, through the command line, with Terraform or Clouformation, etc.
  3. If you want to share an existing managed service, you only need to copy the details in the Kubernetes Secret.

Managing resources externally has some drawbacks:

  1. You need to have a strategy to keep the credentials in sync. You could manually copy them or have a script that does this on your behalf.
  2. You need to create the database or message broker ahead of time.
  3. The managed service is deployed separately from the app. You can't use kubectl apply -f and create an entire stack with a single command.

If you're willing to accept the trade-offs, managing resources externally is a viable option to have apps in Kubernetes consuming cloud provider services such as a database.

But what if you want to manage those resources as if they were part of Kubernetes?

Imagine you could create a new type of Kubernetes object such as a database object:


kind: ManagedDatabase
metadata: name: rds-mysql
spec: version: '11.5' storageType: Durable

Every time you submit that YAML resource to the cluster, a script calls the cloud provider APIs and creates the database.

When the database is ready, a Secret is created with the connection string, username, password, etc.

Imagine, with a single kubectl apply -f you could deploy the app (as a Deployment) and the database (as a managed service).

Does that work in Kubernetes?

Can you create a ManagedDatabase object?

You can get close.

The Kubernetes Service Catalog

Cloud providers have an extensive collection of managed services.

And of the managed services can be scripted: the cloud provider exposes an API to create, update, consume and delete resources.

So, before you start using services from a cloud provider, you could do the following:

  1. Collect all the managed services that you wish to use into a list.
  2. Create a Kubernetes object for each item in that list.
  3. Create a component that can map custom Kubernetes objects to the cloud provider managed services.

In other words:

  1. Your cluster knows that there's a new ManagedDatabase resource.
  2. When you submit a YAML definition for a ManagedDatabase, there's a script that calls the cloud provider APIs and creates the database.

As you can imagine, you could end up with a long but useful list.

You could also open-source it and share it with your friends and colleagues since they might find it useful.

The Kubernetes community did precisely that.

However, they developed a more generic solution called the Service Catalog.

The Service Catalog is a pluggable system that integrates to the cloud provider and lists all managed services available.

It also maps resources — if you use Amazon Web Services (AWS), you can create an RDS database object in Kubernetes.

In turn, the Service Catalog calls the API on the cloud provider, creates the resource and populates a Secret in Kubernetes with all the details.

The Service Catalog has integrations for all major cloud providers (Azure, Amazon Web Service, Google) and you can write your own as well.

Let's have a look at how it works in Amazon Web Services (AWS).

If you need help setting up the Service Catalog on Amazon Web Services (AWS) you might find this repository useful.

Once the Service Catalog is installed, you can list all available resources with:


kubectl get ClusterServiceClasses
079bab7e-5fb3-5e50-a91a-46ce25ec44d5 elasticsearch sb
20ac3b60-8526-5327-9ea3-4f7893bed097 dynamodb sbs
499aa9c6-085f-5e46-9c59-f063acb20538 sns sb
9ff60dab-f1de-53c7-baf4-47c70ed95419 redshift sb
ad60756e-7482-5077-9883-2fe99ec45fd4 route53 sb
baa70642-2fab-5615-ad19-05885f2be690 rdsmysql sb
ca3e9a89-0310-530b-8a5d-21ef88888822 s3 sb
cf1eda6f-8515-5d75-ba12-4a82c283d30d sqs sb
d56a51f6-75f2-5f42-a613-3d9aff5e33f1 rdsmariadb sb
e15beaa5-e813-5b58-aa27-0b1f61d91a80 rdspostgresql sb
f4b249ea-71ea-59dd-9d0b-4f7f1c651b5d rdsmssql sb

Please note that the above output is truncated.

You can inspect the class for the S3 service with:


kubectl get ClusterServiceClass ca3e9a89-0310-530b-8a5d-21ef88888822 -o yaml

The response is a YAML definition:


kind: ClusterServiceClass
metadata: name: ca3e9a89-0310-530b-8a5d-21ef88888822
spec: bindable: true bindingRetrievable: false clusterServiceBrokerName: sb description: AWS Service Broker - Amazon S3 externalID: ca3e9a89-0310-530b-8a5d-21ef88888822 externalMetadata: displayName: Amazon S3 documentationUrl:' imageUrl: longDescription: Amazon Simple Storage Service (Amazon S3) is storage for the Internet. You can use Amazon S3 to store and retrieve any amount of data at any time, from anywhere on the web. You can accomplish these tasks using the simple and intuitive web interface of the AWS Management Console. outputsAsIs: false providerDisplayName: Amazon Web Services externalName: s3 planUpdatable: false tags: - AWS - S3 - Object Storage

The ClusterServiceClass is a description of what the Service does.

Notice how the description doesn't contain any information on how much space you could use or what other parameters you could tweak in the service.

All the configuration and parameters are collected in a Service Plan.

You can retrieve all the Service plans for the AWS S3 Service with:


kubectl get ClusterServicePlans
3050c541-df46-55e2-9ecc-71fed5085852 custom sb ca3e9a89-0310-530b-8a5d-21ef88888822
733fc2af-ade8-533d-85f3-0bb1051aa2f0 production sb ca3e9a89-0310-530b-8a5d-21ef88888822

You can inspect the plan for the AWS S3 Service with:


kubectl get ClusterServicePlan 3050c541-df46-55e2-9ecc-71fed5085852 -o yaml

Please note that 3050c541-df46-55e2-9ecc-71fed5085852 is the name of the custom plan.

The response is a (rather long) YAML definition:


kind: ClusterServicePlan
metadata: name: 3050c541-df46-55e2-9ecc-71fed5085852
spec: clusterServiceClassRef: name: ca3e9a89-0310-530b-8a5d-21ef88888822 description: S3 Bucket pre-configured with custom configuration externalName: custom free: false instanceCreateParameterSchema: type: object
 properties: BucketName: default: Auto description: Must contain only lowercase letters, numbers, periods (.), and hyphens. If set to Auto, a bucket name will be generated (-),Cannot end in numbers title: BucketName type: string 

But how do you use the S3 resource in your Pods?

Let's have a look at the code.

To create an instance of S3, you have to combine ClusterServiceClass and ClusterServicePlan in a ServiceInstance resource:


kind: ServiceInstance
metadata: name: my-s3-instance
 clusterServiceClassExternalName: s3
 clusterServicePlanExternalName: custom
 parameters: BucketName: my-test-20200325

As soon as you submit the resource to the cluster, the Service broker calls the APIs on Amazon Web Services (AWS) and creates the S3 bucket.

You can check the progress with:


kubectl describe ServiceInstance my-s3-instance

When the broker completes, you should see a message confirming that the resource is created.

But do you trust it?

You can verify that the resource is created for real by listing the S3 buckets on AWS with:


aws s3api list-buckets
{ "Buckets":[ { "Name":"my-test-20200325", "CreationDate":"2020-03-25T06:36:45+00:00" }, { "Name": "my-test-20200325-logging", "CreationDate": "2020-03-25T06:36:18+00:00" } ], "Owner":{ "DisplayName":"learnk8s", "ID":"8b10" }

Excellent, the resource exists!

How do you access it in the cluster, though?

You need to bind the resource to use it.

You should create a ServiceBinding resource:


kind: ServiceBinding
metadata: name: my-s3-binding
spec: instanceRef: name: my-s3-instance

At this point, the broker server will retrieve and store the credentials to access S3 into a Kubernetes secret.

You can check the progress with:


kubectl describe servicebinding my-s3-binding

Did it create a Kubernetes Secret?


kubectl get secrets
default-token-zbl2r 3
my-s3-binding Opaque 6

Excellent, let's inspect the my-s3-binding Secret:


kubectl get secret my-s3-binding -o yaml
apiVersion: v1
data: BUCKET_ARN: Y2F0Y2htZWlmeW91Y2Fu BUCKET_NAME: Y2F0Y2htZWlmeW91Y2Fu LOGGING_BUCKET_NAME: c21hcnRjb29raWU= S3_AWS_ACCESS_KEY_ID: c21hcnRjb29raWU= S3_AWS_SECRET_ACCESS_KEY: Y2F0Y2htZWlmeW91Y2Fu S3_REGION: c21hcnRjb29raWU=
kind: Secret
metadata: name: my-s3-binding
type: Opaque

Great, you can use the value in the Secret to connect to the S3 bucket.

In other words, you have to use four Custom Resource Definitions (CRDs) to use the Service Catalog:

  1. Managed services are described in a ClusterServiceClass.
  2. If you want to use a service, you can choose an appropriate plan with ClusterServicePlans.
  3. You can combine the service and its plan into a ServiceInstance. As soon as you submit it, the cloud provider creates the resource.
  4. Creating resources is not enough. When you want to use it, you should claim it with a ServiceBinding.

Service Catalog limitations

In the previous example, you provisioned an S3 bucket in Amazon Web Services (AWS).

You can also create other services such as Databases (RDS) and message brokers (SQS, AMQP).

You can also use the Service Catalog with another cloud vendor such as Google Cloud or Azure.

There's one caveat, though.

There's no easy way to import or bind to existing resources.

Imagine you have three clusters:

  1. One for the development environment.
  2. One for the testing environment.
  3. One for the production environment.

The production environment has its own database, but you want to share a single database between testing a development.

How can you do so with the Service Catalog?

If you create the database from the development cluster using the Service Catalog, how do you import it in the testing environment?

You can't, unless you copy the Secret from one cluster to the other.

The Service Catalog is designed to create new resources only.

There's no way to import existing services.

So you're back at square one: provisioning resources and manually copying secrets.

What options do you have to create and import resources?

Is there anything similar to the Service Catalog?

Terraform to the rescue

The Service Catalog runs a controller that calls the cloud provider APIs to create resources.

What if you could replace that controller with another?

That's the idea behind Kubeform — a controller that crates cloud resources using Terraform.

Terraform is a convenient tool for creating cloud resources using code.

When you install Kubeform in your cluster, you can create resources like this:


kind: S3Bucket
metadata: name: my-bucket
spec: bucket: my-bucket-20200325 providerRef: name: aws

Kubeform runs a controller that receives the YAML resource, invokes Terraform and create the S3 bucket.

It looks very similar to the Service Catalog, so what's the difference?

Kubeform calls the Terraform provider for Amazon Web Services (AWS), Google Cloud, Azure etc.

Terraform is a mature product, and it is battle-tested in production by small and large companies.

As a consequence, it has the most up to date list of configuration parameters.

As an example, when configuring an S3 Bucket, the Service catalog exposes about twelve properties that you can customise such as:

  • BucketAccessControl
  • BucketName
  • EnableGlacierLifeCycle
  • EnableLogging
  • etc.

For the same object, Terraform has over a hundred parameters.

Also, when you use the Service Catalog, all the parameters in the ServiceInstance object are not checked.

You can add as many properties as you want, Kubernetes won't stop you:


kind: ServiceInstance
metadata: name: my-s3-instance
spec: clusterServiceClassExternalName: s3 clusterServicePlanExternalName: custom parameters: BucketName: my-test-20200325
 something: random_string

However, you will see an error as soon as the Service Catalog passes the arguments to the cloud provider.

In contrast, Kubeform exposes Custom Resource Definitions (CRDs).

kind: S3Bucket
metadata: name: my-bucket
spec: bucket: my-bucket-20200325
 something: random_string
 providerRef: name: aws

Kubernetes will validate the resource and make sure that only defined fields are accepted before the resource is stored in etcd.

So, should you always use Kubeform instead of the Service Catalog?

Terraform stores the current state of the infrastructure into a state file.

If you run a command twice, Terraform provision the difference between the current and previous state only.

Kubeform stores that state in etcd — similarly to the Service Catalog.

But there's no easy way to retrieve the state or fix it when it gets corrupted or out of sync.

Or to use the state to manually import resources.

So if you want to share a database between development and testing environment, you're out of luck — again.

Exploring more options on Amazon Web Services (AWS)

Amazon Web Services offers the Service Catalog integration as well as the AWS Service Operator — a component that lives in the cluster and is capable of creating managed resources.

The operator works similarly to the Service Broker, but instead of using ServiceInstances and ServiceBindings, you use Custom Resource Definitions (CRDs).

Here's an example of how you can define a DynamoDB table:


kind: DynamoDB
metadata: name: dynamo-table
spec: hashAttribute: name: name type: S rangeAttribute: name: created_at type: S readCapacityUnits: 5 writeCapacityUnits: 5

The AWS Service Operator combines a dedicated controller with Custom Resource Definitions (CRDs), making it an excellent solution to provision managed services on AWS.

But there's something you should know.

The AWS Service operator was initially released in 2018 and is currently deprecated without migration path.

At the time of writing, the engineering team at Amazon is rewriting the operator from the ground up.

While it's an excellent alternative to the Service Catalog, it's not a viable option at the moment.

Exploring more options on Google Cloud Platform (GCP)

Google Cloud offers the Config Connector — a component that lives in the cluster and behaves like the Service Catalog.

You can use the Config Connector to create managed services such as Cloud Spanner and Big Query using Custom Resource Definitions (CRDs).

Here's the definition for Spanner:


kind: SpannerInstance
metadata: name: spannerinstance-sample
spec: config: regional-us-west1 displayName: Spanner Instance Sample numNodes: 1

The solution is remarkably similar to the AWS Service Operator, but the operator actually exists.

And, on the plus side, Google designed Config Connector to work nicely with existing service.

So you can import pre-existing resources with Config Connector and use them in your cluster.

If you have an existing Cloud SQL or Cloud Pub/Sub you can import those in the cluster too.

  • If you liked the idea behind Kubeform and you want to explore more options, you should check out the Terraform controller and AWS CloudFormation operator.
  • Crossplane offers a set of Kubernetes Custom Resource Definitions (CRDs) and controllers that provision and manage services on cloud providers. In other words, Crossplane is a Service Catalog equivalent but built with CRDs.
  • The Service Binding Operator is an exciting twist on the Service Broker. It focuses on binding service and apps and doesn't deal with cloud provider-specific implementations.

There are several ways to configure cloud managed services such as databases in your cluster.

Some options give you full integration, and you can create resources without noticing that they are not part of the cluster.

With the Service Catalog, for example, you submit a YAML definition in the cluster that triggers the creation of a managed resource in your cloud provider.

Managing resources with Kubernetes feels natural, and you can offer a more consistent and unified experience to the user of your cluster.

But what's the price?

Most of the solutions discussed above are quite complicated:

  • You need to create and understand the lifecycle of a resource in the Service Catalog.
  • Kubeform is an ingenious tool to provision infrastructure with Terraform, but it's hard to debug and fix issues with the state.
  • Proprietary tools such as Google Config Connector and AWS Service Operator are great, but they are not mature.

And even if you can manage the complexity, there is a deal-breaker for most of them: you can't import existing managed services.

You can't import an existing database in your cloud provider with the following tools:

  • Service Catalog
  • Kubeform
  • AWS Service operator

That's not something that you can easily ignore.

So what's the alternative?

You can still provision resources externally and then manually create a Secret in the cluster with the credential.

It's manual and limiting, but it's straightforward, and it works.

Is there anything better than that?

Not without a lot of trade-offs.

The team behind the Service Catalog has the ups-broker a component that could be used to import existing services, but it's not well documented and only meant for testing.

The Service Binding Operator is one of the most exciting experiments in this area.

However, it is designed to run on OpenShift and supports only a few cloud services.

It's also quite complicated and requires an external binding server.

Closing notes and further reading

A few notes on the Service Catalog:

Many thanks to Daniel Weibel and Salman Iqbal for offering some invaluable suggestions.

That's all folks!

If you enjoyed this article, you might find the following articles interesting:

Be the first to be notified when a new article or Kubernetes experiment is published.

*We'll never share your email address, and you can opt-out at any time.

  • In-depth Kubernetes training that is practical and easy to understand

  • ⎈ Instructor-led workshops

    Deep dive into containers and Kubernetes with the help of our instructors and become an expert in deploying applications at scale.

  • ⎈ Online courses

    Learn Kubernetes online with hands-on, self-paced courses. No need to leave the comfort of your home.

  • ⎈ Corporate training

    Train your team in containers and Kubernetes with a customised learning path — remotely or on-site.