Containers Canonical Kubernetes Nvidia
- By Ubuntu Containers Team
| Channel | Version | Revision | Published | Runs on |
|---|---|---|---|---|
| latest/stable | 11 | 11 | 15 Mar 2021 | |
| latest/candidate | 11 | 11 | 15 Mar 2021 | |
| latest/beta | 11 | 11 | 15 Mar 2021 | |
| latest/edge | 11 | 11 | 15 Mar 2021 |
juju deploy containers-canonical-kubernetes-nvidia
Deploy universal operators easily with Juju, the Universal Operator Lifecycle Manager.
Platform:
Relevant links
Discuss this bundle
Share your thoughts on this charm with the community on discourse.
The Canonical Distribution of Kubernetes
WARNING
This bundle is a Proof of Concept for deploying CDK on GPU-enabled workers. This bundle is currently only usable in AWS.
Overview
This is a scaled-out Kubernetes cluster composed of the following components and features:
- Kubernetes (automated deployment, operations, and scaling)
- Kubernetes cluster with one master and three worker nodes.
- TLS used for communication between nodes for security.
- A CNI plugin (Flannel)
- A load balancer for HA kubernetes-master
- Optional Ingress Controller (on worker)
- Optional Dashboard addon (on master) including Heapster for cluster monitoring
- EasyRSA
- Performs the role of a certificate authority serving self signed certificates to the requesting units of the cluster.
- Etcd (distributed key value store)
- Three node cluster for reliability.
For a more minimal cluster suitable for development and testing, deploy the smaller
kubernetes-core bundle via conjure-up kubernetes-core.
Usage
Installation has been automated via conjure-up:
sudo snap install conjure-up --classic
conjure-up canonical-kubernetes
Conjure-up will prompt you for deployment options (AWS, GCE, Azure, etc.) and credentials.
Proxy configuration
If you are operating behind a proxy (i.e., your charms are running in a limited-egress environment and can not reach IP addresses external to their network), you will need to configure your model appropriately before deploying the Kubernetes bundle.
First, configure your model's http-proxy and https-proxy settings with your
proxy (here we use squid.internal:3128 as an example):
$ juju model-config http-proxy=http://squid.internal:3128 https-proxy=https://squid.internal:3128
Because services often need to reach machines on their own network (including
themselves), you will also need to add localhost to the no-proxy model
configuration setting, along with any internal subnets you're using. The
following example includes two subnets:
$ juju model-config no-proxy=localhost,10.5.5.0/24,10.246.64.0/21
After deploying the bundle, you need to configure the kubernetes-worker charm
to use your proxy:
$ juju config kubernetes-worker http_proxy=http://squid.internal:3128 https_proxy=https://squid.internal:3128
Alternate deployment methods
Deploying with Juju directly
juju deploy canonical-kubernetes
Note: If you're deploying on lxd, use conjure-up instead, as described above. It configures your lxd profile to support running Kubernetes on lxd.
This will deploy the Canonical Distribution of Kubernetes with default machine constraints. This is useful for lab environments, but for real-world use you should provide more CPU and memory to kubernetes-worker units.
You can increase the constraints by downloading and editing the bundle.yaml to fit your needs. Customize the bundle constraints by following this guide.
To deploy your customized bundle:
juju deploy ./bundle.yaml
Note: If you're operating behind a proxy, remember to set the
kubernetes-workerproxy configuration options as described in the Proxy configuration section above.
This bundle exposes the kubeapi-load-balancer and kubernetes-worker charms by default, so they are accessible through their public addresses.
If you would like to remove external access, unexpose them:
juju unexpose kubeapi-load-balancer
juju unexpose kubernetes-worker
To get the status of the deployment, run juju status. For constant updates,
combine it with the watch command:
watch -c juju status --color
Using with your own resources
In order to support restricted-network deployments, the charms in this bundle support juju resources.
This allows you to juju attach the resources built for the architecture of
your cloud.
juju attach kubernetes-master kubectl=/path/to/kubectl.snap
juju attach kubernetes-master kube-apiserver=/path/to/kube-apiserver.snap
juju attach kubernetes-master kube-controller-manager=/path/to/kube-controller-manager.snap
juju attach kubernetes-master kube-scheduler=/path/to/kube-scheduler.snap
juju attach kubernetes-master cdk-addons=/path/to/cdk-addons.snap
juju attach kubernetes-worker kubectl=/path/to/kubectl.snap
juju attach kubernetes-worker kubelet=/path/to/kubelet.snap
juju attach kubernetes-worker kube-proxy=/path/to/kube-proxy.snap
juju attach kubernetes-worker cni=/path/to/cni.tgz
Using a specific Kubernetes version
You can select a specific version or series of Kubernetes by configuring CDK to use a specific snap channel. For example, to use the 1.6 series:
juju config kubernetes-master channel=1.6/stable
juju config kubernetes-worker channel=1.6/stable
After changing the channel, you'll need to manually execute the upgrade action on each kubernetes-worker and kubernetes-master unit, e.g.:
juju run-action kubernetes-master/0 upgrade
...
juju run-action kubernetes-worker/0 upgrade
juju run-action kubernetes-worker/1 upgrade
juju run-action kubernetes-worker/2 upgrade
...
By default, the channel is set to stable on the current minor version of Kubernetes, for example, 1.6/stable. This means your cluster will receive automatic upgrades for new patch releases (e.g. 1.6.2 -> 1.6.3), but not for new minor versions (e.g. 1.6.3 -> 1.7). To upgrade to a new minor version, configure the channel manually as described above.
Interacting with the Kubernetes cluster
After the cluster is deployed you may assume control over the Kubernetes cluster from any kubernetes-master or kubernetes-worker node.
To download the credentials and client application to your local workstation:
Create the kubectl config directory.
mkdir -p ~/.kube
Copy the kubeconfig file to the default location.
juju scp kubernetes-master/0:config ~/.kube/config
Install kubectl locally.
snap install kubectl --classic
Query the cluster.
kubectl cluster-info
Accessing the Kubernetes dashboard
The Kubernetes dashboard addon is installed by default, along with Heapster,
Grafana and InfluxDB for cluster monitoring. The dashboard addons can be
enabled or disabled by setting the enable-dashboard-addons config on the
kubernetes-master application:
juju config kubernetes-master enable-dashboard-addons=true
To access the dashboard, you may establish a secure tunnel to your cluster with the following command:
kubectl proxy
By default, this establishes a proxy running on your local machine and the
kubernetes-master unit. To reach the Kubernetes dashboard, visit
http://localhost:8001/ui
Control the cluster
kubectl is the command line utility to interact with a Kubernetes cluster.
Minimal getting started
To check the state of the cluster:
kubectl cluster-info
List all nodes in the cluster:
kubectl get nodes
Now you can run pods inside the Kubernetes cluster:
kubectl create -f example.yaml
List all pods in the cluster:
kubectl get pods
List all services in the cluster:
kubectl get services
For expanded information on kubectl beyond what this README provides, please see the kubectl overview which contains practical examples and an API reference.
Additionally if you need to manage multiple clusters, there is more information about configuring kubectl in the kubectl config guide
Using Ingress
The kubernetes-worker charm supports deploying an NGINX ingress controller. Ingress allows access from the Internet to containers running web services inside the cluster.
First allow the Internet access to the kubernetes-worker charm with with the following Juju command:
juju expose kubernetes-worker
In Kubernetes, workloads are declared using pod, service, and ingress definitions. An ingress controller is provided to you by default and deployed into the default namespace of the cluster. If one is not available, you may deploy it with:
juju config kubernetes-worker ingress=true
Ingress resources are DNS mappings to your containers, routed through endpoints.
As an example for users unfamiliar with Kubernetes, we packaged an action to both deploy an example and clean itself up.
To deploy 5 replicas of the microbot web application inside the Kubernetes cluster run the following command:
juju run-action kubernetes-worker/0 microbot replicas=5
This action performs the following steps:
It creates a deployment titled 'microbots' comprised of 5 replicas defined during the run of the action. It also creates a service named 'microbots' which binds an 'endpoint', using all 5 of the 'microbots' pods.
Finally, it will create an ingress resource, which points at a xip.io domain to simulate a proper DNS service.
Running the packaged example
Run a Juju action to create the example microbot web application:
$ juju run-action kubernetes-worker/0 microbot replicas=3
Action queued with id: db7cc72b-5f35-4a4d-877c-284c4b776eb8
$ juju show-action-output db7cc72b-5f35-4a4d-877c-284c4b776eb8
results:
address: microbot.104.198.77.197.xip.io
status: completed
timing:
completed: 2016-09-26 20:42:42 +0000 UTC
enqueued: 2016-09-26 20:42:39 +0000 UTC
started: 2016-09-26 20:42:41 +0000 UTC
Note: Your FQDN will be different and contain the address of the cloud instance.
At this point, you can inspect the cluster to observe the workload coming online.
List the pods
$ kubectl get pods
NAME READY STATUS RESTARTS AGE
default-http-backend-kh1dt 1/1 Running 0 1h
microbot-1855935831-58shp 1/1 Running 0 1h
microbot-1855935831-9d16f 1/1 Running 0 1h
microbot-1855935831-l5rt8 1/1 Running 0 1h
nginx-ingress-controller-hv5c2 1/1 Running 0 1h
List the services and endpoints
$ kubectl get services,endpoints
NAME CLUSTER-IP EXTERNAL-IP PORT(S) AGE
svc/default-http-backend 10.1.225.82 <none> 80/TCP 1h
svc/kubernetes 10.1.0.1 <none> 443/TCP 1h
svc/microbot 10.1.44.173 <none> 80/TCP 1h
NAME ENDPOINTS AGE
ep/default-http-backend 10.1.68.2:80 1h
ep/kubernetes 172.31.31.139:6443 1h
ep/microbot 10.1.20.3:80,10.1.68.3:80,10.1.7.4:80 1h
List the ingress resources
$ kubectl get ingress
NAME HOSTS ADDRESS PORTS AGE
microbot-ingress microbot.52.38.62.235.xip.io 172.31.26.109 80 1h
When all the pods are listed as Running, you are ready to visit the address listed in the HOSTS column of the ingress listing.
Note: It is normal to see a 502/503 error during initial application deployment.
As you refresh the page, you will be greeted with a microbot web page, serving from one of the microbot replica pods. Refreshing will show you another microbot with a different hostname as the requests are load-balanced across the replicas.
Clean up example
There is also an action to clean up the microbot applications. When you are done using the microbot application you can delete them from the pods with one Juju action:
juju run-action kubernetes-worker/0 microbot delete=true
If you no longer need Internet access to your workers, remember to unexpose the kubernetes-worker charm:
juju unexpose kubernetes-worker
To learn more about Kubernetes Ingress and how to configure the Ingress Controller beyond defaults (such as TLS and websocket support) view the nginx-ingress-controller project on github.
Scale out Usage
Scaling kubernetes-worker
The kubernetes-worker nodes are the load-bearing units of a Kubernetes cluster.
By default, pods are automatically spread across the kubernetes-worker units that you have deployed.
To add more kubernetes-worker units to the cluster:
juju add-unit kubernetes-worker
or specify machine constraints to create larger nodes:
juju set-constraints kubernetes-worker cpu-cores=8 mem=32G
juju add-unit kubernetes-worker
Refer to the machine constraints documentation for other machine constraints that might be useful for the kubernetes-worker units.
Scaling Etcd
Etcd is the key-value store for the Kubernetes cluster. For reliability the bundle defaults to three instances in this cluster.
For more scalability, we recommend between 3 and 9 etcd nodes. If you want to add more nodes:
juju add-unit etcd
The CoreOS etcd documentation has a chart for the optimal cluster size to determine fault tolerance.
Adding optional storage
Using Juju Storage, the Canonical Distribution of Kubernetes allows you to connect with durable storage devices such as Ceph.
Deploy a minimum of three ceph-mon and three ceph-osd charms:
juju deploy cs:ceph-mon -n 3
juju deploy cs:ceph-osd -n 3
Relate the charms:
juju add-relation ceph-mon ceph-osd
List the storage pools available to Juju for your cloud:
$ juju storage-pools
Name Provider Attrs
ebs ebs
ebs-ssd ebs volume-type=ssd
loop loop
rootfs rootfs
tmpfs tmpfs
Note: This listing is for the Amazon Web Services public cloud. Different clouds will have different pool names.
Add a storage pool to the ceph-osd charm by NAME,SIZE,COUNT:
juju add-storage ceph-osd/0 osd-devices=ebs,10G,1
juju add-storage ceph-osd/1 osd-devices=ebs,10G,1
juju add-storage ceph-osd/2 osd-devices=ebs,10G,1
Next relate the storage cluster with the Kubernetes cluster:
juju add-relation kubernetes-master ceph-mon
We are now ready to enlist Persistent Volumes in Kubernetes, which our workloads can use via Persistent Volume Claims (PVC).
juju run-action kubernetes-master/0 create-rbd-pv name=test size=50
This example created a "test" Radios Block Device (rbd) in the size of 50 MB.
You should see the PV become enlisted and be marked as available:
$ watch kubectl get pv
NAME CAPACITY ACCESSMODES STATUS CLAIM REASON AGE
test 50M RWO Available 10s
To consume these Persistent Volumes, your pods will need a Persistent Volume Claim associated with them, a task that is outside the scope of this README. See the Persistent Volumes documentation for more information.
Known Limitations and Issues
The following are known issues and limitations with the bundle and charm code:
Destroying the the easyrsa charm will result in loss of public key infrastructure (PKI).
Deployment locally on LXD will require the use of conjure-up to tune settings on the host's LXD installation to support Docker and other components.
If resources fail to download during initial deployment for any reason, you will need to download and install them manually. For example, if kubernetes-master is missing its resources, download them from the resources section of the sidebar here and install them by running, for example:
juju attach kubernetes-master kube-apiserver=/path/to/snap.You can find resources for the canonical-kubernetes charms here: