Managing Instance Groups

kOps has the concept of "instance groups", which are a group of similar machines. On AWS, they map to an Auto Scaling group.

By default, a cluster has:

• One instance group for each node zone, called nodes-<zone> (e.g. nodes-us-east-1c). These instances are your workers.
• One instance group for each master zone, called master-<zone> (e.g. master-us-east-1c). These normally have minimum size and maximum size = 1, so they will run a single instance. We do this so that the cloud will always relaunch masters, even if everything is terminated at once. We have an instance group per zone because we need to force the cloud to run an instance in every zone, so we can mount the master volumes - we cannot do that across zones.

This page explains some common instance group operations. For more detailed documentation of the various configuration keys, see the InstanceGroup Resource.

Instance Groups Disclaimer

• When there is only one availability zone in a region (eu-central-1) and you would like to run multiple masters, you have to define multiple instance groups for each of those masters. (e.g. master-eu-central-1-a and master-eu-central-1-b and so on...)
• If instance groups are not defined correctly (particularly when there are an even number of master or multiple groups of masters into one availability zone in a single region), etcd servers will not start and master nodes will not check in. This is because etcd servers are configured per availability zone. DNS and Route53 would be the first places to check when these problems are happening.

Listing instance groups

kops get instancegroups

NAME                    ROLE    MACHINETYPE     MIN     MAX     ZONES
master-us-east-1c       Master                  1       1       us-east-1c
nodes-us-east-1c        Node    t2.medium       2       2       us-east-1c

You can also use the kops get ig alias.

Change the instance type in an instance group

First you edit the instance group spec, using kops edit ig nodes-us-east-1c. Change the machine type to t2.large, for example. Now if you kops get ig, you will see the large instance size. Note though that these changes have not yet been applied.

To preview the change:

kops update cluster <clustername>

...
Will modify resources:
  *awstasks.LaunchTemplate LaunchTemplate/mycluster.mydomain.com
    InstanceType t2.medium -> t2.large

Presuming you're happy with the change, go ahead and apply it: kops update cluster <clustername> --yes

This change will apply to new instances only; if you'd like to roll it out immediately to all the instances you have to perform a rolling update.

See a preview with: kops rolling-update cluster

Then restart the machines with: kops rolling-update cluster --yes

This will drain nodes, restart them with the new instance type, and validate them after startup.

Changing the number of nodes

Note: This uses GCE as example. It will look different when AWS is the cloud provider, but the concept and the configuration is the same.

If you kops get ig you should see that you have InstanceGroups for your nodes and for your master:

> kops get ig
NAME            ROLE    MACHINETYPE MIN MAX SUBNETS
master-us-central1-a    Master  n1-standard-1   1   1   us-central1
nodes-us-central1-a Node    n1-standard-2   2   2   us-central1

Let's change the number of nodes to 3. We'll edit the InstanceGroup configuration using kops edit (which should be very familiar to you if you've used kubectl edit). kops edit ig nodes-us-central1-a will open the InstanceGroup in your editor, looking a bit like this:

apiVersion: kops.k8s.io/v1alpha2
kind: InstanceGroup
metadata:
  labels:
    kops.k8s.io/cluster: simple.k8s.local
  name: nodes-us-central1-a
spec:
  image: cos-cloud/cos-stable-57-9202-64-0
  machineType: n1-standard-2
  maxSize: 2
  minSize: 2
  role: Node
  subnets:
  - us-central1
  zones:
  - us-central1-a

Edit minSize and maxSize, changing both from 2 to 3, save and exit your editor. If you wanted to change the image or the machineType, you could do that here as well. There are actually a lot more fields, but most of them have their default values, so won't show up unless they are set. The general approach is the same though.

On saving you'll note that nothing happens. Although you've changed the model, you need to tell kOps to apply your changes to the cloud.

We use the same kops update cluster command that we used when initially creating the cluster; when run without --yes it should show you a preview of the changes, and now there should be only one change:

> kops update cluster
Will modify resources:
  InstanceGroupManager/us-central1-a-nodes-us-central1-a-simple-k8s-local
    TargetSize               2 -> 3

This is saying that we will alter the TargetSize property of the InstanceGroupManager object named us-central1-a-nodes-us-central1-a-simple-k8s-local, changing it from 2 to 3.

That's what we want, so we kops update cluster --yes.

kOps will resize the GCE managed instance group from 2 to 3, which will create a new GCE instance, which will then boot and join the cluster. Within a minute or so you should see the new node join:

> kubectl get nodes
NAME                        STATUS    AGE       VERSION
master-us-central1-a-thjq   Ready     10h       v1.7.2
nodes-us-central1-a-g2v2    Ready     10h       v1.7.2
nodes-us-central1-a-tmk8    Ready     10h       v1.7.2
nodes-us-central1-a-z2cz    Ready     1s       v1.7.2

nodes-us-central1-a-z2cz just joined our cluster!

Changing the image

That was a fairly simple change, because we didn't have to reboot the nodes. Most changes though do require rolling your instances - this is actually a deliberate design decision, in that we are aiming for immutable nodes. An example is changing your image. We're using cos-stable, which is Google's Container OS. Let's try Debian Stretch instead.

If you run gcloud compute images list to list the images available to you in GCE, you should see a debian-9 image:

> gcloud compute images list
...
debian-9-stretch-v20170918                        debian-cloud             debian-9                              READY
...

So now we'll do the same kops edit ig nodes, except this time change the image to debian-cloud/debian-9-stretch-v20170918:

Now kops update cluster will show that you're going to create a new GCE Instance Template, and that the Managed Instance Group is going to use it:

Will create resources:
  InstanceTemplate/nodes-us-central1-a-simple-k8s-local
    Network                 name:default id:default
    Tags                    [simple-k8s-local-k8s-io-role-node]
    Preemptible             false
    BootDiskImage           debian-cloud/debian-9-stretch-v20170918
    BootDiskSizeGB          128
    BootDiskType            pd-standard
    CanIPForward            true
    Scopes                  [compute-rw, monitoring, logging-write, storage-ro]
    Metadata                {cluster-name: <resource>, startup-script: <resource>}
    MachineType             n1-standard-2

Will modify resources:
  InstanceGroupManager/us-central1-a-nodes-us-central1-a-simple-k8s-local
    InstanceTemplate         id:nodes-us-central1-a-simple-k8s-local-1507043948 -> name:nodes-us-central1-a-simple-k8s-local

Note that the BootDiskImage is indeed set to the debian 9 image you requested.

kops update cluster --yes will now apply the change, but if you were to run kubectl get nodes you would see that the instances had not yet been reconfigured. There's a hint at the bottom:

Changes may require instances to restart: kops rolling-update cluster`

These changes require your instances to restart (we'll remove the COS images and replace them with Debian images). kOps can perform a rolling update to minimize disruption, but even so you might not want to perform the update right away; you might want to make more changes or you might want to wait for off-peak hours. You might just want to wait for the instances to terminate naturally - new instances will come up with the new configuration - though if you're not using preemptible/spot instances you might be waiting for a long time.

Fetching images via AWS SSM (AWS Only)

Introduced
kOps 1.25.3

If you are using AWS, you can dynamically fetch instance group images from an AWS SSM Parameter. kOps will automatically fetch SSM Parameter and lookup the AMI ID on every kops update cluster run. This is useful if you often update your images and don't want to update your instance group configuration every time. Your SSM Parameter must start with ssm: and contain the full path of the SSM Parameter.

An example spec looks like this:

metadata:
  name: nodes-us-west-2a
spec:
  image: ssm:/aws/service/canonical/ubuntu/server/18.04/stable/current/amd64/hvm/ebs-gp2/ami-id
  machineType: t3.medium
  maxSize: 1
  minSize: 1
  role: Node

Changing the root volume size or type

The default volume size for Masters is 64 GB, while the default volume size for a node is 128 GB.

The procedure to resize the root volume works the same way:

• Edit the instance group, set rootVolumeSize and/or rootVolumeType to the desired values: kops edit ig nodes-us-east-1c
• rootVolumeType must be one of supported volume types, e.g. gp2 (default), io1 (high performance) or standard (for testing).
• If rootVolumeType is set to io1 then you can define the number of Iops by specifying rootVolumeIops (defaults to 100 if not defined)
• Preview changes: kops update cluster <clustername>
• Apply changes: kops update cluster <clustername> --yes
• Rolling update to update existing instances: kops rolling-update cluster --yes

For example, to set up a 200GB gp2 root volume, your InstanceGroup spec might look like:

metadata:
  name: nodes-us-east-1c
spec:
  machineType: t3.medium
  maxSize: 2
  minSize: 2
  role: Node
  rootVolumeSize: 200
  rootVolumeType: gp2

Another example would be to set up a 200GB io1 root volume with 200 provisioned Iops, which would make your InstanceGroup spec look like:

metadata:
  name: nodes-us-east-1c
spec:
  machineType: t3.medium
  maxSize: 2
  minSize: 2
  role: Node
  rootVolumeSize: 200
  rootVolumeType: io1
  rootVolumeIops: 200

As of kOps 1.19 you can use gp3 volumes for better performance, which would make your InstanceGroup spec look like:

metadata:
  name: nodes-us-east-1c
spec:
  machineType: t3.medium
  maxSize: 2
  minSize: 2
  role: Node
  rootVolumeSize: 200
  rootVolumeType: gp3
  rootVolumeIops: 4000
  rootVolumeThroughput: 200

Encrypting the root volume

Introduced
kOps 1.19

You can encrypt the root volume (note, presently confined to AWS) via the instancegroup specification.

metadata:
  name: nodes-us-east-1a
spec:
  ...
  role: Node
  rootVolumeSize: 200
  rootVolumeEncryption: true
  rootVolumeEncryptionKey: arn:aws:kms:us-east-1:012345678910:key/1234abcd-12ab-34cd-56ef-1234567890ab

In the above example the encryption key is optional. The default key for EBS encryption is used when not specified. The encryption key can specified as the key ID, alias or ARN, as described in the AWS docs.

Adding additional storage to the instance groups

Introduced
kOps 1.12

You can add additional storage (note, presently confined to AWS) via the instancegroup specification.

---
apiVersion: kops.k8s.io/v1alpha2
kind: InstanceGroup
metadata:
  labels:
    kops.k8s.io/cluster: my-beloved-cluster
  name: compute
spec:
  cloudLabels:
    role: compute
  image: coreos.com/CoreOS-stable-1855.4.0-hvm
  machineType: m4.large
  ...
  volumes:
  - device: /dev/xvdd
    size: 20
    type: gp2
    encrypted: true
    key: arn:aws:kms:us-east-1:012345678910:key/1234abcd-12ab-34cd-56ef-1234567890ab

In AWS the above example shows how to add an additional encrypted 20gb EBS volume, which applies to each node within the instancegroup.

Automatically formatting and mounting the additional storage

You can add additional storage via the above volumes collection though this only provisions the storage itself. Assuming you don't wish to handle the mechanics of formatting and mounting the device yourself (perhaps via a hook) you can utilize the volumeMounts section of the instancegroup to handle this for you.

---
apiVersion: kops.k8s.io/v1alpha2
kind: InstanceGroup
metadata:
  labels:
    kops.k8s.io/cluster: my-beloved-cluster
  name: compute
spec:
  cloudLabels:
    role: compute
  image: coreos.com/CoreOS-stable-1855.4.0-hvm
  machineType: m4.large
  ...
  volumeMounts:
  - device: /dev/xvdd
    filesystem: ext4
    path: /var/lib/docker
  volumes:
  - device: /dev/xvdd
    encrypted: true
    size: 20
    type: gp2

The above will provision the additional storage, format and mount the device into the node. Note this feature is purposely distinct from volumes so that it may be reused in areas such as ephemeral storage. Using a c5d.large instance as an example, which comes with a 50gb SSD drive; we can use the volumeMounts to mount this into /var/lib/docker for us.

---
apiVersion: kops.k8s.io/v1alpha2
kind: InstanceGroup
metadata:
  labels:
    kops.k8s.io/cluster: my-beloved-cluster
  name: compute
spec:
  cloudLabels:
    role: compute
  image: coreos.com/CoreOS-stable-1855.4.0-hvm
  machineType: c5d.large
  ...
  volumeMounts:
  - device: /dev/nvme1n1
    filesystem: ext4
    path: /data
  # -- mount the instance storage --
  - device: /dev/nvme2n1
    filesystem: ext4
    path: /var/lib/docker
  volumes:
  - device: /dev/nvme1n1
    encrypted: true
    size: 20
    type: gp2

For AWS you can find more information on device naming conventions here

$ df -h | grep nvme[12]
/dev/nvme1n1      20G   45M   20G   1% /data
/dev/nvme2n1      46G  633M   45G   2% /var/lib/docker

Note: at present its up to the user ensure the correct device names.

Creating a new instance group

Suppose you want to add a new group of nodes, perhaps with a different instance type. You do this using kops create ig <InstanceGroupName> --subnet <zone(s)>. Currently the --subnet flag is required, and it receives the zone(s) of the subnet(s) in which the instance group will be. The command opens an editor with a skeleton configuration, allowing you to edit it before creation.

So the procedure is:

• kops create ig morenodes --subnet us-east-1a

or, in case you need it to be in more than one subnet, use a comma-separated list:

• kops create ig morenodes --subnet us-east-1a,us-east-1b,us-east-1c
• Preview: kops update cluster <clustername>
• Apply: kops update cluster <clustername> --yes
• (no instances need to be relaunched, so no rolling-update is needed)

Creating an instance group of mixed instances types (AWS Only)

Introduced
kOps 1.12

AWS permits the creation of mixed instance EC2 Autoscaling Groups using a mixed instance policy, allowing the users to build a target capacity and make up of on-demand and spot instances while offloading the allocation strategy to AWS.

---
apiVersion: kops.k8s.io/v1alpha2
kind: InstanceGroup
metadata:
  labels:
    kops.k8s.io/cluster: your.cluster.name
  name: compute
spec:
  cloudLabels:
    role: compute
  image: coreos.com/CoreOS-stable-1911.4.0-hvm
  machineType: m4.large
  maxSize: 50
  minSize: 10
  # You can manually set the maxPrice you're willing to pay - it will default to the onDemand price.
  maxPrice: "1.0"
  # add the mixed instance policy here
  mixedInstancesPolicy:
    instances:
    - m4.xlarge
    - m5.large
    - m5.xlarge
    - t2.medium
    onDemandAboveBase: 5
    spotInstancePools: 3

The mixed instance policy permits setting the following configurable below, but for more details please check against the AWS documentation.

Note: as of writing this the kube cluster autoscaler does not support mixed instance groups, in the sense it will still scale groups up and down based on capacity but some of the simulations it does might be wrong as it's not aware of the instance type coming into the group.

Note: when upgrading from a launchconfiguration to launchtemplate with mixed instance policy the launchconfiguration is left undeleted as has to be manually removed.

Moving from one instance group spanning multiple AZs to one instance group per AZ

It may be beneficial to have one IG per AZ rather than one IG spanning multiple AZs. One common example is, when you have a persistent volume claim bound to an AWS EBS Volume this volume is bound to the AZ it has been created in so any resource (e.g. a StatefulSet) depending on that volume is bound to that same AZ. In this case you have to ensure that there is at least one node running in that same AZ, which is not guaranteed by one IG. This however can be guaranteed by one IG per AZ.

So the procedure is:

• kops edit ig nodes
• Remove two of the subnets, e.g. eu-central-1b and eu-central-1c
• Alternatively you can also delete the existing IG and create a new one with a more suitable name
• kops create ig nodes-eu-central-1b --subnet eu-central-1b
• kops create ig nodes-eu-central-1c --subnet eu-central-1c
• Preview: kops update cluster <clustername>
• Apply: kops update cluster <clustername> --yes
• Rolling update to update existing instances: kops rolling-update cluster --yes

Converting an instance group to use spot instances

Follow the normal procedure for reconfiguring an InstanceGroup, but set the maxPrice property to your bid. For example, "0.10" represents a spot-price bid of $0.10 (10 cents) per hour.

An example spec looks like this:

metadata:
  name: nodes-us-east-1a
spec:
  machineType: t2.medium
  maxPrice: "0.01"
  maxSize: 3
  minSize: 3
  role: Node

So the procedure is:

• Edit: kops edit ig nodes-us-east-1a
• Preview: kops update cluster <clustername>
• Apply: kops update cluster <clustername> --yes
• Rolling-update, only if you want to apply changes immediately: kops rolling-update cluster

Adding Taints or Labels to an Instance Group

If you're running Kubernetes 1.6.0 or later, you can also control taints in the InstanceGroup. The taints property takes a list of strings. The following example would add two taints to an IG, using the same edit -> update -> rolling-update process as above.

Additionally, nodeLabels can be added to an IG in order to take advantage of Pod Affinity. Every node in the IG will be assigned the desired labels. For more information see the labels documentation.

metadata:
  name: nodes-us-east-1a
spec:
  machineType: m3.medium
  maxSize: 3
  minSize: 3
  role: Node
  taints:
  - dedicated=gpu:NoSchedule
  - team=search:PreferNoSchedule
  nodeLabels:
    spot: "false"

Resizing the master

(This procedure should be pretty familiar by now!)

Your master instance group will probably be called master-us-west-1c or something similar.

kops edit ig master-us-west-1c

Add or set the machineType:

spec:
  machineType: m3.large

• Preview changes: kops update cluster <clustername>

• Apply changes: kops update cluster <clustername> --yes

• Rolling-update, only if you want to apply changes immediately: kops rolling-update cluster

If you want to minimize downtime, scale the master ASG up to size 2, then wait for that new master to be Ready in kubectl get nodes, then delete the old master instance, and scale the ASG back down to size 1. (A future version of rolling-update will probably do this automatically)

Deleting an instance group

If you decide you don't need an InstanceGroup any more, you delete it using: kops delete ig <name>

Example: kops delete ig morenodes

No kops update cluster nor kops rolling-update is needed, so be careful when deleting an instance group, your nodes will be deleted automatically (and note this is not currently graceful, so there may be interruptions to workloads where the pods are running on those nodes).

EBS Volume Optimization

EBS-Optimized instances can be created by setting the following field:

spec:
  rootVolumeOptimization: true

Booting from a volume in OpenStack

If you want to boot from a volume when you are running in openstack you can set annotations on the instance groups.

# Example for nodes
apiVersion: kops.k8s.io/v1alpha2
kind: InstanceGroup
metadata:
  labels:
    kops.k8s.io/cluster: k8s.dev.local
  name: nodes
  annotations:
    openstack.kops.io/osVolumeBoot: enabled
    openstack.kops.io/osVolumeSize: "15" # In gigabytes
spec:
  detailedInstanceMonitoring: true
  machineType: t2.medium
  maxSize: 2
  minSize: 2
  role: Node

If openstack.kops.io/osVolumeSize is not set it will default to the minimum disk specified by the image.

Working with InstanceGroups

The kOps InstanceGroup is a declarative model of a group of nodes. By modifying the object, you can change the instance type you're using, the number of nodes you have, the OS image you're running - essentially all the per-node configuration is in the InstanceGroup.

We'll assume you have a working cluster - if not, you probably want to read how to get started on GCE.