An open source system for automating deployment, scaling, and operations of applications.

Friday, July 14, 2017

How Watson Health Cloud Deploys Applications with Kubernetes

Today’s post is by Sandhya Kapoor, Senior Technologist, Watson Platform for Health, IBM
For more than a year, Watson Platform for Health at IBM deployed healthcare applications in virtual machines on our cloud platform. Because virtual machines had been a costly, heavyweight solution for us, we were interested to evaluate Kubernetes for our deployments.

Our design was to set up the application and data containers in the same namespace, along with the required agents using sidecars, to meet security and compliance requirements in the healthcare industry.
I was able to run more processes on a single physical server than I could using a virtual machine. Also, running our applications in containers ensured optimal usage of system resources.
To orchestrate container deployment, we are using Armada infrastructure, a Kubernetes implementation by IBM for automating deployment, scaling, and operations of application containers across clusters of hosts, providing container-centric infrastructure.
With Kubernetes, our developers can rapidly develop highly available applications by leveraging the power and flexibility of containers, and with integrated and secure volume service, we can store persistent data, share data between Kubernetes pods, and restore data when needed.
Here is a snapshot of Watson Care Manager, running inside a Kubernetes cluster:

Before deploying an app, a user must create a worker node cluster. I can create a cluster using the kubectl cli commands or create it from a Bluemix dashboard.
Our clusters consist of one or more physical or virtual machines, also known as worker nodes, that are loosely coupled, extensible, and centrally monitored and managed by the Kubernetes master. When we deploy a containerized app, the Kubernetes master decides where to deploy the app, taking into consideration the deployment requirements and available capacity in the cluster.
A user makes a request to Kubernetes to deploy the containers, specifying the number of replicas required for high availability. The Kubernetes scheduler decides where the pods (groups of one or more containers) will be scheduled and which worker nodes they will be deployed on, storing this information internally in Kubernetes and etcd. The deployment of pods in worker nodes is updated based on load at runtime, optimizing the placement of pods in the cluster.
Kubelet running in each worker node regularly polls the kube API server. If there is new work to do, kubelet pulls the configuration information and takes action, for example, spinning off a new pod.

Process Flow:
UCD – IBM UrbanCode Deploy is a tool for automating application deployments through your environments. WH Cluster – Kubernetes worker node.

Usage of GitLab in the Process Flow:
We stored all our artifacts in GitLab, which includes the Docker files that are required for creating the image, YAML files needed to create a pod, and the configuration files to make the Healthcare application run.
GitLab and Jenkins interaction in the Process Flow:
We use Jenkins for continuous integration and build automation to create/pull/retag the Docker image and push the image to a Docker registry in the cloud.
Basically, we have a Jenkins job configured to interact with GitLab project to get the latest artifacts and, based on requirements, it will either create a new Docker image from scratch by pulling the needed intermediate images from Docker/Bluemix repository or update the Docker image.
After the image is created/updated the Jenkins job pushes the image to a Bluemix repository to save the latest image to be pulled by UrbanCode Deploy (UCD) component.
Jenkins and UCD interaction in the Process Flow:
The Jenkins job is configured to use the UCD component and its respective application, application process, and the UCD environment to deploy the application. The Docker image version files that will be used by the UCD component are also passed via Jenkins job to the UCD component.
Usage of UCD in the Process Flow:
UCD is used for deployment and the end-to end deployment process is automated here. UCD component process involves the following steps:
  • Download the required artifacts for deployment from the Gitlab.
  • Login to Bluemix and set the KUBECONFIG based on the Kubernetes cluster used for creating the pods.
  • Create the application pod in the cluster using kubectl create command.
  • If needed, run a rolling update to update the existing pod.

Deploying the application in Armada:
Provision a cluster in Armada with <x> worker nodes. Create Kubernetes controllers for deploying the containers in worker nodes, the Armada infrastructure pulls the Docker images from IBM Bluemix Docker registry to create containers. We tried deploying an application container and running a logmet agent (see Reading and displaying logs using logmet container, below) inside the containers that forwards the application logs to an IBM cloud logging service. As part of the process, YAML files are used to create a controller resource for the UrbanCode Deploy (UCD). UCD agent is deployed as a DaemonSet controller, which is used to connect to the UCD server. The whole process of deployment of application happens in UCD. To support the application for public access, we created a service resource to interact between pods and access container services. For storage support, we created persistent volume claims and mounted the volume for the containers.

UCD: IBM UrbanCode Deploy is a tool for automating application deployments through your environments. Armada: Kubernetes implementation of IBM. WH Docker Registry: Docker Private image registry. Common agent containers: We expect to configure our services to use the WHC mandatory agents. We deployed all ion containers.

Reading and displaying logs using logmet container:
Logmet is a cloud logging service that helps to collect, store, and analyze an application’s log data. It also aggregates application and environment logs for consolidated application or environment insights and forwards them. Metrics are transmitted with collectd. We chose a model that runs a logmet agent process inside the container. The agent takes care of forwarding the logs to the cloud logging service configured in containers.

The application pod mounts the application logging directory to the storage space, which is created by persistent volume claim, and stores the logs, which are not lost even when the pod dies. Kibana is an open source data visualization plugin for Elasticsearch. It provides visualization capabilities on top of the content indexed on an Elasticsearch cluster.

Exposing services with Ingress:
Ingress controllers are reverse proxies that expose services outside cluster through URLs. They act as an external HTTP load balancer that uses a unique public entry point to route requests to the application.
To expose our services to outside the cluster, we used Ingress. In Armada, if we create a paid cluster, an Ingress controller is automatically installed for us to use. We were able to access services through Ingress by creating a YAML resource file that specifies the service path.

Sandhya Kapoor, Senior Technologist, Watson Platform for Health, IBM

  • Post questions (or answer questions) on Stack Overflow
  • Join the community portal for advocates on K8sPort
  • Follow us on Twitter @Kubernetesio for latest updates
  • Connect with the community on Slack
  • Get involved with the Kubernetes project on GitHub

Thursday, June 29, 2017

Kubernetes 1.7: Security Hardening, Stateful Application Updates and Extensibility

Today we’re announcing Kubernetes 1.7, a milestone release that adds security, storage and extensibility features motivated by widespread production use of Kubernetes in the most demanding enterprise environments. 

At-a-glance, security enhancements in this release include encrypted secrets, network policy for pod-to-pod communication, node authorizer to limit kubelet access and client / server TLS certificate rotation. 

For those of you running scale-out databases on Kubernetes, this release has a major feature that adds automated updates to StatefulSets and enhances updates for DaemonSets. We are also announcing alpha support for local storage and a burst mode for scaling StatefulSets faster. 

Also, for power users, API aggregation in this release allows user-provided apiservers to be served along with the rest of the Kubernetes API at runtime. Additional highlights include support for extensible admission controllers, pluggable cloud providers, and container runtime interface (CRI) enhancements.

What’s New
  • The Network Policy API is promoted to stable. Network policy, implemented through a network plug-in, allows users to set and enforce rules governing which pods can communicate with each other. 
  • Node authorizer and admission control plugin are new additions that restrict kubelet’s access to secrets, pods and other objects based on its node.
  • Encryption for Secrets, and other resources in etcd, is now available as alpha. 
  • Kubelet TLS bootstrapping now supports client and server certificate rotation.
  • Audit logs stored by the API server are now more customizable and extensible with support for event filtering and webhooks. They also provide richer data for system audit.

Stateful workloads:
  • StatefulSet Updates is a new beta feature in 1.7, allowing automated updates of stateful applications such as Kafka, Zookeeper and etcd, using a range of update strategies including rolling updates.
  • StatefulSets also now support faster scaling and startup for applications that do not require ordering through Pod Management Policy. This can be a major performance improvement. 
  • Local Storage (alpha) was one of most frequently requested features for stateful applications. Users can now access local storage volumes through the standard PVC/PV interface and via StorageClasses in StatefulSets.
  • DaemonSets, which create one pod per node already have an update feature, and in 1.7 have added smart rollback and history capability.
  • A new StorageOS Volume plugin provides highly-available cluster-wide persistent volumes from local or attached node storage.


Additional Features:
  • Alpha support for external admission controllers is introduced, providing two options for adding custom business logic to the API server for modifying objects as they are created and validating policy. 
  • Policy-based Federated Resource Placement is introduced as Alpha providing placement policies for the federated clusters, based on custom requirements such as regulation, pricing or performance.


  • Third Party Resource (TPR) has been replaced with Custom Resource Definitions (CRD) which provides a cleaner API, and resolves issues and corner cases that were raised during the beta period of TPR. If you use the TPR beta feature, you are encouraged to migrate, as it is slated for removal by the community in Kubernetes 1.8.

The above are a subset of the feature highlights in Kubernetes 1.7. For a complete list please visit the release notes.

This release is possible thanks to our vast and open community. Together, we’ve already pushed more than 50,000 commits in just three years, and that’s only in the main Kubernetes repo. Additional extensions to Kubernetes are contributed in associated repos bringing overall stability to the project. This velocity makes Kubernetes one of the fastest growing open source projects -- ever. 

Kubernetes adoption has been coming from every sector across the world. Recent user stories from the community include: 

  • GolfNow, a member of the NBC Sports Group, migrated their application to Kubernetes giving them better resource utilization and slashing their infrastructure costs in half.
  • Bitmovin, provider of video infrastructure solutions, showed us how they’re using Kubernetes to do multi-stage canary deployments in the cloud and on-prem.
  • Ocado, world’s largest online supermarket, uses Kubernetes to create a distributed data center for their smart warehouses. Read about their full setup here.
  • Is Kubernetes helping your team? Share your story with the community. See our growing resource of user case studies and learn from great companies like Box that have adopted Kubernetes in their organization. 

Huge kudos and thanks go out to the Kubernetes 1.7 release team, led by Dawn Chen of Google. 

Kubernetes 1.7 is available for download on GitHub. To get started with Kubernetes, try one of the these interactive tutorials

Get Involved
Join the community at CloudNativeCon + KubeCon in Austin Dec. 6-8 for the largest Kubernetes gathering ever. Speaking submissions are open till August 21 and discounted registration ends October 6.

The simplest way to get involved is joining one of the many Special Interest Groups (SIGs) that align with your interests. Have something you’d like to broadcast to the Kubernetes community? Share your voice at our weekly community meeting, and these channels:

  • Post questions (or answer questions) on Stack Overflow
  • Join the community portal for advocates on K8sPort
  • Follow us on Twitter @Kubernetesio for latest updates
  • Connect with the community on Slack
  • Share your Kubernetes story

Many thanks to our vast community of contributors and supporters in making this and all releases possible.

-- Aparna Sinha, Group Product Manager, Kubernetes Google and Ihor Dvoretskyi, Program Manager, Kubernetes Mirantis

Wednesday, May 31, 2017

Draft: Kubernetes container development made easy

Today's post is by Brendan Burns, Director of Engineering at Microsoft Azure and Kubernetes co-founder.

About a month ago Microsoft announced the acquisition of Deis to expand our expertise in containers and Kubernetes. Today, I’m excited to announce a new open source project derived from this newly expanded Azure team: Draft. 

While by now the strengths of Kubernetes for deploying and managing applications at scale are well understood. The process of developing a new application for Kubernetes is still too hard. It’s harder still if you are new to containers, Kubernetes, or developing cloud applications.

Draft fills this role. As it’s name implies it is a tool that helps you begin that first draft of a containerized application running in Kubernetes. When you first run the draft tool, it automatically discovers the code that you are working on and builds out the scaffolding to support containerizing your application. Using heuristics and a variety of pre-defined project templates draft will create an initial Dockerfile to containerize your application, as well as a Helm Chart to enable your application to be deployed and maintained in a Kubernetes cluster. Teams can even bring their own draft project templates to customize the scaffolding that is built by the tool.

But the value of draft extends beyond simply scaffolding in some files to help you create your application. Draft also deploys a server into your existing Kubernetes cluster that is automatically kept in sync with the code on your laptop. Whenever you make changes to your application, the draft daemon on your laptop synchronizes that code with the draft server in Kubernetes and a new container is built and deployed automatically without any user action required. Draft enables the “inner loop” development experience for the cloud.

Of course, as is the expectation with all infrastructure software today, Draft is available as an open source project, and it itself is in “draft” form :) We eagerly invite the community to come and play around with draft today, we think it’s pretty awesome, even in this early form. But we’re especially excited to see how we can develop a community around draft to make it even more powerful for all developers of containerized applications on Kubernetes.

To give you a sense for what Draft can do, here is an example drawn from the Getting Started page in the GitHub repository.

There are multiple example applications included within the examples directory. For this walkthrough, we'll be using the python example application which uses Flask to provide a very simple Hello World webserver.

$ cd examples/python

Draft Create

We need some "scaffolding" to deploy our app into a Kubernetes cluster. Draft can create a Helm chart, a Dockerfile and a draft.toml with draft create:

$ draft create
--> Python app detected
--> Ready to sail
$ ls
Dockerfile  chart/  draft.toml  requirements.txt

The chart/ and Dockerfile assets created by Draft default to a basic Python configuration. This Dockerfile harnesses the python:onbuild image, which will install the dependencies in requirements.txt and copy the current directory into /usr/src/app. And to align with the service values in chart/values.yaml, this Dockerfile exposes port 80 from the container.

The draft.toml file contains basic configuration about the application like the name, which namespace it will be deployed to, and whether to deploy the app automatically when local files change.

$ cat draft.toml
   name = "tufted-lamb"
   namespace = "default"
   watch = true
   watch_delay = 2

Draft Up

Now we're ready to deploy to a Kubernetes cluster.
Draft handles these tasks with one draft up command:
  • reads configuration from draft.toml
  • compresses the chart/ directory and the application directory as two separate tarballs
  • uploads the tarballs to draftd, the server-side component
  • draftd then builds the docker image and pushes the image to a registry
  • draftd instructs helm to install the Helm chart, referencing the Docker registry image just built
With the watch option set to true, we can let this run in the background while we make changes later on…

$ draft up
--> Building Dockerfile
Step 1 : FROM python:onbuild
onbuild: Pulling from library/python
Successfully built 38f35b50162c
--> Pushing
The push refers to a repository []
5a3c633ae76c9bdb81b55f5d4a783398bf00658e: digest: sha256:9d9e9fdb8ee3139dd77a110fa2d2b87573c3ff5ec9c045db6009009d1c9ebf5b size: 16384
--> Deploying to Kubernetes
   Release "tufted-lamb" does not exist. Installing it now.
--> Status: DEPLOYED
--> Notes:
    1. Get the application URL by running these commands:
    NOTE: It may take a few minutes for the LoadBalancer IP to be available.
          You can watch the status of by running 'kubectl get svc -w tufted-lamb-tufted-lamb'
 export SERVICE_IP=$(kubectl get svc --namespace default tufted-lamb-tufted-lamb -o jsonpath='{.status.loadBalancer.ingress[0].ip}')
 echo http://$SERVICE_IP:80

Watching local files for changes...

Interact with the Deployed App

Using the handy output that follows successful deployment, we can now contact our app. Note that it may take a few minutes before the load balancer is provisioned by Kubernetes. Be patient!

$ export SERVICE_IP=$(kubectl get svc --namespace default tufted-lamb-tufted-lamb -o jsonpath='{.status.loadBalancer.ingress[0].ip}')
$ curl http://$SERVICE_IP

When we curl our app, we see our app in action! A beautiful "Hello World!" greets us.

Update the App

Now, let's change the "Hello, World!" output in to output "Hello, Draft!" instead:

$ cat <<EOF >
from flask import Flask

app = Flask(__name__)

def hello():
   return "Hello, Draft!\n"

if __name__ == "__main__":'', port=8080)

Draft Up(grade)

Now if we watch the terminal that we initially called draft up with, Draft will notice that there were changes made locally and call draft up again. Draft then determines that the Helm release already exists and will perform a helm upgrade rather than attempting another helm install:

--> Building Dockerfile
Step 1 : FROM python:onbuild
Successfully built 9c90b0445146
--> Pushing
The push refers to a repository []
--> Deploying to Kubernetes
--> Status: DEPLOYED
--> Notes:
    1. Get the application URL by running these commands:
    NOTE: It may take a few minutes for the LoadBalancer IP to be available.
          You can watch the status of by running 'kubectl get svc -w tufted-lamb-tufted-lamb'
 export SERVICE_IP=$(kubectl get svc --namespace default tufted-lamb-tufted-lamb -o jsonpath='{.status.loadBalancer.ingress[0].ip}')
 echo http://$SERVICE_IP:80

Now when we run curl http://$SERVICE_IP, our first app has been deployed and updated to our Kubernetes cluster via Draft!
We hope this gives you a sense for everything that Draft can do to streamline development for Kubernetes. Happy drafting!

--Brendan Burns, Director of Engineering, Microsoft Azure

  • Post questions (or answer questions) on Stack Overflow
  • Join the community portal for advocates on K8sPort
  • Follow us on Twitter @Kubernetesio for latest updates
  • Connect with the community on Slack
  • Get involved with the Kubernetes project on GitHub