Star Fork Issue

Design Proposal #002: Existing VPC

Design Proposal #002: Existing VPC

STATUS: This proposal is a working draft, it will get refined and augment as needed. If any non-trivial changes are need to functionality defined here, in particular the user experience, those changes should be suggested via a PR to this proposal document. Any other changes to the text of the proposal or technical corrections are also very welcome.

Initial implementation of eksctl intentionally avoided having too many options for user to customise the VPC configuration. That was in oder to simplify how the tools works. Since then, many users asked for various features in relation to custom VPC.

There a few specific use-cases:

  1. co-location with kops #50
  2. set custom CIDR #279
  3. private/public subnets #120
  4. use any existing VPC #42
  5. use same subnet/AZ for a nodegroup #232

Out of the above, we already have support for 1 & 2 (as of 0.1.8) as well as 3 (as of 0.1.9).

The main challenge with these customisations is user experience and how different flags interact. Ultimately, once Cluster API support is implemented (planned for 0.3.0), some of these flags will not be needed and there will be more fine-grain parameters available via Cluster API. This proposal sets out flags that should serve well for most common use-cases until 0.3.0.

Flags defined in 0.1.9

These are fairly simple, and rather more user-friendly. These are assumed satisfactory for the purpose of this proposal.

Complete set of flags required for using existing VPC/subnets

In order to use eksctl with any custom VPC, one has to provide subnet ID. VPC ID can be derived by making an API call.

So the following flags will be needed:

Also, the following combinations must be valid:

User will have to make sure IDs are correct as well as their private/public designations.

For some use-cases, specifying VPC ID maybe preferred, and in that case the VPC maybe the default one (in which case it is ought to be possible to use default subnets or specific subnets). Hence the following combinations ought to be valid:

Needless to say that when none of --vpc-* flags given, eksctl will create a dedicated VPC and subnets (public as well as private). As more flags are being added, grouping them in --help output would be very helpful; especially because most of these flags are for advanced usage.

## Security Groups

It must be noted that security groups are managed by eksctl only, as certain configuration is required to ensure cluster is is fully functional. Unlike with subnets and VPC, security groups can be update after they were created and any ingress and/or egress rules can be added or removed. If more advanced usage is commonly required, a separate design proposal will be required. It is also expected that advanced functionality will be available via Cluster API (which is planned for 0.3.0).

Additional resources/commands

It would be plausible to also provide utility for managing VPC, i.e.:

This would allow users to create VPCs with recommended configurations without having to worry about the requirements, yet they will be able to do VPC resource management at a different level of access. It should be noted that similar functionality is likely to be required for IAM resources (subject to separate proposal).

Notes on AWS CNI driver

It can be safely assumed (based on simple tests) that the driver is meant to coexist very happily with anything you have in your VPC, from plain EC2 instances, through load balancers and gateways to Lambda functions. It is thereby ought to be possible to share a VPC between two or more cluster running separate instances of the driver. It is expected that driver will try and use what IP addresses are available, which is only a subject to IP address space usage, of course. Allocations and other properties of the driver can be monitored via CloudWatch and/or Prometheus.

However, when very small subnets are used in a VPC, or there are too many resources using up the IP address space, one will ultimately run into issue sooner than later. If that’s the case, VPC peering or some other alternative should be consider. Peering should already be possible, and custom CIDR flag is available as of 0.1.8.

It must be noted, that there exists a plan to provide Weave Net as an overlay option. It maybe possible to use IPv6 also, but more experiments and research will need to be done (it not clear if EKS supports IPv6 properly yet).

Requirements of custom VPC

When user decides to use any of --vpc-* flags, it is up to them to ensure that:

There maybe other requirements imposed by EKS or Kubernetes, and it is entirely up to the user to stay up-to-date on any requirements and/or recommendations, and find ways in which those work well with the practices at their own organisations and any requirements imposed by it. It is not our goal to keep this section of the proposal up to date. However, eksctl create cluster aims to always provide an up-to-date configuration with a dedicated VPC.


eksctl - a CLI for Amazon EKS

Circle CI Coverage StatusGo Report Card

eksctl is a simple CLI tool for creating clusters on EKS - Amazon’s new managed Kubernetes service for EC2. It is written in Go, and uses CloudFormation.

You can create a cluster in minutes with just one command – eksctl create cluster!

Gophers: E, K, S, C, T, & L

Usage

To download the latest release, run:

curl --silent --location "https://github.com/weaveworks/eksctl/releases/download/latest_release/eksctl_$(uname -s)_amd64.tar.gz" | tar xz -C /tmp
sudo mv /tmp/eksctl /usr/local/bin

Alternatively, macOS users can use Homebrew:

brew install weaveworks/tap/eksctl

You will need to have AWS API credentials configured. What works for AWS CLI or any other tools (kops, Terraform etc), should be sufficient. You can use ~/.aws/credentials file or environment variables. For more information read AWS documentation.

To create a basic cluster, run:

eksctl create cluster

A cluster will be created with default parameters

Once you have created a cluster, you will find that cluster credentials were added in ~/.kube/config. If you have kubectl v1.10.x as well as aws-iam-authenticator commands in your PATH, you should be able to use kubectl. You will need to make sure to use the same AWS API credentials for this also. Check EKS docs for instructions. If you installed eksctl via Homebrew, you should have all of these dependencies installed already.

Example output:

$ eksctl create cluster
2018-10-26T16:22:17+01:00 [ℹ]  using region us-west-2
2018-10-26T16:22:19+01:00 [ℹ]  setting availability zones to [us-west-2a us-west-2b us-west-2c]
2018-10-26T16:22:19+01:00 [ℹ]  subnets for us-west-2a - public:192.168.0.0/19 private:192.168.96.0/19
2018-10-26T16:22:19+01:00 [ℹ]  subnets for us-west-2b - public:192.168.32.0/19 private:192.168.128.0/19
2018-10-26T16:22:19+01:00 [ℹ]  subnets for us-west-2c - public:192.168.64.0/19 private:192.168.160.0/19
2018-10-26T16:22:19+01:00 [ℹ]  using "ami-0a54c984b9f908c81" for nodes
2018-10-26T16:22:19+01:00 [ℹ]  creating EKS cluster "floral-unicorn-1540567338" in "us-west-2" region
2018-10-26T16:22:19+01:00 [ℹ]  will create 2 separate CloudFormation stacks for cluster itself and the initial nodegroup
2018-10-26T16:22:19+01:00 [ℹ]  if you encounter any issues, check CloudFormation console or try 'eksctl utils describe-stacks --region=us-west-2 --name=floral-unicorn-1540567338'
2018-10-26T16:22:19+01:00 [ℹ]  creating cluster stack "eksctl-floral-unicorn-1540567338-cluster"
2018-10-26T16:33:03+01:00 [ℹ]  creating nodegroup stack "eksctl-floral-unicorn-1540567338-nodegroup-0"
2018-10-26T16:36:44+01:00 [✔]  all EKS cluster resource for "floral-unicorn-1540567338" had been created
2018-10-26T16:36:44+01:00 [✔]  saved kubeconfig as "/Users/ilya/.kube/config"
2018-10-26T16:36:46+01:00 [ℹ]  the cluster has 0 nodes
2018-10-26T16:36:46+01:00 [ℹ]  waiting for at least 2 nodes to become ready
2018-10-26T16:37:22+01:00 [ℹ]  the cluster has 2 nodes
2018-10-26T16:37:22+01:00 [ℹ]  node "ip-192-168-25-215.us-west-2.compute.internal" is ready
2018-10-26T16:37:22+01:00 [ℹ]  node "ip-192-168-83-60.us-west-2.compute.internal" is ready
2018-10-26T16:37:23+01:00 [ℹ]  kubectl command should work with "~/.kube/config", try 'kubectl get nodes'
2018-10-26T16:37:23+01:00 [✔]  EKS cluster "floral-unicorn-1540567338" in "us-west-2" region is ready
$

To list the details about a cluster or all of the clusters, use:

eksctl get cluster [--name=<name>] [--region=<region>]

To create the same kind of basic cluster, but with a different name, run:

eksctl create cluster --name=cluster-1 --nodes=4

A default StorageClass (gp2 volume type provisioned by EBS) will be added automatically when creating a cluster. If you want to prevent this, use the --storage-class flag. For example:

eksctl create cluster --storage-class=false

To write cluster credentials to a file other than default, run:

eksctl create cluster --name=cluster-2 --nodes=4 --kubeconfig=./kubeconfig.cluster-2.yaml

To prevent storing cluster credentials locally, run:

eksctl create cluster --name=cluster-3 --nodes=4 --write-kubeconfig=false

To let eksctl manage cluster credentials under ~/.kube/eksctl/clusters directory, run:

eksctl create cluster --name=cluster-3 --nodes=4 --auto-kubeconfig

To obtain cluster credentials at any point in time, run:

eksctl utils write-kubeconfig --name=<name> [--kubeconfig=<path>] [--set-kubeconfig-context=<bool>]

To use a 3-5 node Auto Scaling Group, run:

eksctl create cluster --name=cluster-5 --nodes-min=3 --nodes-max=5

To use 30 c4.xlarge nodes and prevent updating current context in ~/.kube/config, run:

eksctl create cluster --name=cluster-6 --nodes=30 --node-type=c4.xlarge --set-kubeconfig-context=false

In order to allow SSH access to nodes, eksctl imports ~/.ssh/id_rsa.pub by default, to use a different SSH public key, e.g. my_eks_node_id.pub, run:

eksctl create cluster --ssh-access --ssh-public-key=my_eks_node_id.pub

To use a pre-existing EC2 key pair in us-east-1 region, you can specify key pair name (which must not resolve to a local file path), e.g. to use my_kubernetes_key run:

eksctl create cluster --ssh-access  --ssh-public-key=my_kubernetes_key --region=us-east-1

To add custom tags for all resources, use --tags.

NOTE: Until [https://github.com/weaveworks/eksctl/issues/25] is resolved, tags cannot be applied to EKS cluster itself, but most of other resources (e.g. EC2 nodes).

eksctl create cluster --tags environment=staging --region=us-east-1

To configure node volume size, use the --node-volume-size flag.

eksctl create cluster --node-volume-size=50

NOTE: In us-east-1 you are likely to get UnsupportedAvailabilityZoneException. If you do, copy the suggested zones and pass --zones flag, e.g. eksctl create cluster --region=us-east-1 --zones=us-east-1a,us-east-1b,us-east-1d. This may occur in other regions, but less likely. You shouldn’t need to use --zone flag otherwise.

To delete a cluster, run:

eksctl delete cluster --name=<name> [--region=<region>]

Scaling nodegroup

The initial nodegroup can be scaled by using the eksctl scale nodegroup command. For example, to scale to 5 nodes:

eksctl scale nodegroup --name=<name> --nodes=5

If the desired number of nodes is greater than the current maximum set on the ASG then the maximum value will be increased to match the number of requested nodes. And likewise for the minimum.

Scaling a nodegroup works by modifying the nodegroup CloudFormation stack via a ChangeSet.

NOTE: Scaling a nodegroup down/in (i.e. reducing the number of nodes) may result in errors as we rely purely on changes to the ASG. This means that the node(s) being removed/terminated aren’t explicitly drained. This may be an area for improvement in the future.

VPC Networking

By default, eksctl create cluster instatiates a dedicated VPC, in order to avoid interference with any existing resources for a variety of reasons, including security, but also because it’s challenging to detect all the settings in an existing VPC. Default VPC CIDR used by eksctl is 192.168.0.0/16, it is divided into 8 (/19) subnets (3 private, 3 public & 2 reserved). Initial nodegroup is create in public subnets, with SSH access disabled unless --allow-ssh is specified. However, this implies that each of the EC2 instances in the initial nodegroup gets a public IP and can be accessed on ports 1025 - 65535, which is not insecure in principle, but some compromised workload could risk an access violation.

If that functionality doesn’t suit you, the following options are currently available.

change VPC CIDR

If you need to setup peering with another VPC, or simply need larger or smaller range of IPs, you can use --vpc-cidr flag to change it. You cannot use just any sort of CIDR, there only certain ranges that can be used in AWS VPC.

use private subnets for initial nodegroup

If you prefer to isolate initial nodegroup from the public internet, you can use --node-private-networking flag. When used in conjunction with --ssh-access flag, SSH port can only be accessed inside the VPC.

use existing VPC: shared with kops

You can use a VPC of an existing Kubernetes cluster managed by kops. This feature is provided to facilitate migration and/or cluster peering.

If you have previously created a cluster with kops, e.g. using commands similar to this:

export KOPS_STATE_STORE=s3://kops
kops create cluster cluster-1.k8s.local --zones=us-west-2c,us-west-2b,us-west-2a --networking=weave --yes

You can create an EKS cluster in the same AZs using the same VPC subnets (NOTE: at least 2 AZs/subnets are required):

eksctl create cluster --name=cluster-2 --region=us-west-2 --vpc-from-kops-cluster=cluster-1.k8s.local

use existing VPC: any custom configuration

Use this feature if you must configure a VPC in a way that’s different to how dedicated VPC is configured by eksctl, or have to use a VPC that already exists so your EKS cluster gets shared access to some resources inside that existing VPC, or you have any other use-case that requires you to manage VPCs separately.

You can use an existing VPC by supplying private and/or public subnets using --vpc-private-subnets and --vpc-public-subnets flags. It is up to you to ensure which subnets you use, as there is no simple way to determine automatically whether a subnets is private or public, because configurations vary. Given these flags, eksctl create cluster will determine the VPC ID automatically, but it will not create any routing tables or other resources, such as internet/NAT gateways. It will, however, create dedicated security groups for the initial nodegroup and the control plane.

You must ensure to provide at least 2 subnets in different AZs. There are other requirements that you will need to follow, but it’s entirely up to you to address those. For example, tagging is not strictly necessary, tests have shown that its possible to create a functional cluster without any tags set on the subnets, however there is no guarantee that this will always hold and tagging is recommended.

There maybe other requirements imposed by EKS or Kubernetes, and it is entirely up to you to stay up-to-date on any requirements and/or recommendations, and implement those as needed/possible.

Default security group settings applied by eksctl may or may not be sufficient for sharing access with resources in other security groups. If you wish to modify the ingress/egress rules of the either of security groups, you might need to use another tool to automate changes, or do it via EC2 console.

If you are in doubt, don’t use a custom VPC. Using eksctl create cluster without any --vpc-* flags will always configure the cluster with a fully-functional dedicated VPC.

To create a cluster using 2x private and 2x public subnets, run:

eksctl create cluster \
  --vpc-private-subnets=subnet-0ff156e0c4a6d300c,subnet-0426fb4a607393184 \
  --vpc-public-subnets=subnet-0153e560b3129a696,subnet-009fa0199ec203c37

To create a cluster using 3x private subnets and make initial nodegroup use those subnets, run:

eksctl create cluster \
  --vpc-private-subnets=subnet-0ff156e0c4a6d300c,subnet-0549cdab573695c03,subnet-0426fb4a607393184 \
  --node-private-networking

To create a cluster using 4x public subnets, run:

eksctl create cluster \
  --vpc-public-subnets=subnet-0153e560b3129a696,subnet-0cc9c5aebe75083fd,subnet-009fa0199ec203c37,subnet-018fa0176ba320e45

GPU Support

If you’d like to use GPU instance types (i.e. p2 or p3 ) then the first thing you need to do is subscribe to the EKS-optimized AMI with GPU Support. If you don’t do this then node creation will fail.

After subscribing to the AMI you can create a cluster specifying the GPU instance type you’d like to use for the nodes. For example:

eksctl create cluster --node-type=p2.xlarge

The AMI resolvers (both static and auto) will see that you want to use a GPU instance type (p2 or p3 only) and they will select the correct AMI.

Once the cluster is created you will need to install the NVIDIA Kubernetes device plugin. Check the repo for the most up to date instructions but you should be able to run this:

kubectl create -f https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/v1.11/nvidia-device-plugin.yml

NOTE: Once addon support has been added as part of 0.2.0 it is envisioned that there will be a addon to install the NVIDIA Kubernetes Device Plugin. This addon could potentially be installed automatically as we know an GPU instance type is being used.

Latest & Custom AMI Support

With the 0.1.2 release we have introduced the --node-ami flag for use when creating a cluster. This enables a number of advanced use cases such as using a custom AMI or querying AWS in realtime to determine which AMI to use (non-GPU and GPU instances).

The --node-ami can take the AMI image id for an image to explicitly use. It also can take the following ‘special’ keywords:

Keyword Description
static Indicates that the AMI images ids embedded into eksctl should be used. This relates to the static resolvers.
auto Indicates that the AMI to use for the nodes should be found by querying AWS. This relates to the auto resolver.

If, for example, AWS release a new version of the EKS node AMIs and a new version of eksctl hasn’t been released you can use the latest AMI by doing the following:

eksctl create cluster --node-ami=auto

With the 0.1.9 release we have introduced the --node-ami-family flag for use when creating the cluster. This makes it possible to choose between different offically supported EKS AMI families.

The --node-ami-family can take following keywords:

Keyword Description
AmazonLinux2 Indicates that the EKS AMI image based on Amazon Linux 2 should be used. (default)
Ubuntu1804 Indicates that the EKS AMI image based on Ubuntu 18.04 should be used.

Project Roadmap

Developer use-case (0.2.0)

It should suffice to install a cluster for development with just a single command. Here are some examples:

To create a cluster with default configuration (2 m5.large nodes), run:

eksctl create cluster

The developer may choose to pre-configure popular addons, e.g.:

It should be possible to combine any or all of these addons.

It would also be possible to add any of the addons after cluster was created with eksctl create addons.

Manage EKS the GitOps way (0.3.0)

Just like kubectl, eksctl aims to be compliant with GitOps model, and can be used as part of a GitOps toolkit!

For example, eksctl apply --cluster-config prod-cluster.yaml will manage cluster state declaratively.

And eksctld will be a controller inside of one cluster that can manage multiple other clusters based on Kubernetes Cluster API definitions (CRDs).

Contributions

Code contributions are very welcome. If you are interested in helping make eksctl great then see our contributing guide.

Get in touch

Create an issue, or login to Weave Community Slack (#eksctl) (signup).

Logo Credits

Original Gophers drawn by Ashley McNamara, unique E, K, S, C, T & L Gopher identities had been produced with Gopherize.me.