Getting Kubernetes Ready For Production

13 Apr 2021

The Story

Slowly and surely, the Kubernetes cluster I’ve set up in my company last year moves towards production usage.

The Kubernetes crash course I wrote last year provides a great way to start a fully working cluster, but there’s quite a bit missing when you have service level targets to meet. Kubernetes by itself is quite stable due to its self-healing nature, but there’s much more we can do to make it better.

With a production workload comes a new set of expectations:

  • Uptime and stability
  • Performance and scalability
  • Security

So, I’ve been starting to look into ways that I can make the Kubernetes cluster more stable, and ways to prevent any sort of downtime in the production Deployments.

Uptime and Stability

During deployment, check the status and rollback if necessary

If you have automatic deployments set up (you should), then it’s important to have a mechanism to rollback the Deployment if it fails for whatever reason.

Fortunately, kubectl comes with a command to check the status of your Deployment. If the pods in your Deployment are anything other than “Running”, then the command will fail.

Keep in mind that you will need your liveness and readiness probes set up properly for this. Kubernetes cannot tell if your application is truly working without these probes.

Here’s a script that checks the status of the Deployment for up to 3 minutes, and if it’s still not ready, the command fails. So it’ll run the command to undo the most recent Deployment, effectively rolling back to what it was before deployment.

if ! kubectl rollout status deployment --timeout 3m ${APP_NAME} -n ${NAMESPACE}; then
  kubectl rollout undo deployment ${APP_NAME} -n ${NAMESPACE}
  exit 1
fi

Node Termination Handler (AWS specific)

The node termination handler is a pod created by AWS to send a signal to your Kubernetes cluster if the node is going to be terminated (either due to spot capacity interruptions, or auto scaling group scale-in events). The node will then be drained, the pods evicted, and a replacement node should spin up to handle these evicted pods if necessary. Set up the EKS Cluster Autoscaler from my previous guide if you haven’t done so.

The setup for this is rather long, and so I will just refer you to the guide written by AWS themselves, here: https://github.com/aws/aws-node-termination-handler

I recommend using the “Queue Processor” mode, as the other mode doesn’t support the Kubernetes Deployment resource.

PriorityClass

A priority class allows you to make certain pods and Deployments more important than others. By default, most of the pods in the kube-system have a higher priority already set. Check out kubectl get priorityclass, you’ll see that your cluster has two priority classes already present.

Now, if your cluster is shared by production and non-production workloads, you will want to set the prodution workloads to have a higher priority (but not higher than system-cluster-critical and system-node-critical). You might wonder why this is important since Kubernetes is self-healing. True, but why play with fire? Outages are often unexpected for a reason, and there’s no harm being cautious.

Here’s one scenario where your production pods might go down. Imagine you’re using spot instances on AWS, and some of your nodes are terminated due to a lack of available spot capacity. Some of your pods will be down because of a lack of node capacity, and they’ll remain down until new nodes come up.

But if you had put your production pods on a higher priority class, Kubernetes would have kicked out the less important pods to make room for your production pods. It’s faster to spin up new pods on existing nodes compared to spinning up new nodes entirely, so if there’s downtime (if there even is one), it’s as short as possible.

Here’s the manifest to create a PriorityClass:

apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: production
value: 100
# If globalDefault is true, then all future Deployments will use this priority class,
# even if not defined in the Deployment manifest.
globalDefault: false
description: "Used for production pods."

You’ll need to assign your Deployment to the priority class. It’s a one-liner under the template spec section:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: your-deployment-name
spec:
  selector:
    matchLabels:
      app: your-deployment-name
  template:
    metadata:
      labels:
        app: your-deployment-name
    spec:
      containers:
      - stuff
      priorityClassName: production

PodDisruptionBudget

A pod disruption budget will allow you to set some sort of minimum or maximum disruption that your Deployment can handle.

There are two settings that you can use, and these can be set in either integers or percentages:

  • minAvailable
  • maxUnavailable

Here’s the manifest to set up a pod disruption budget:

apiVersion: policy/v1beta1
kind: PodDisruptionBudget
metadata:
  name: your-pdb
spec:
  # Example format for percentages: "75%", "5%", etc.
  minAvailable: 1
  selector:
    matchLabels:
      app: your-deployment

This means that there will always be a minimum of 1 pod running at any time.

I believe maxUnavailable is quite self-explanatory. It’s the maximum number of pods that can be unavailable at any time.

Pod Anti-Affinity

This is a way to spread your pods out on all the available nodes, rather than having them all on a single node or two. This way, if a single node goes down for whatever reason, the Deployment will still have some pods running.

This section should be under the template spec of your Deployment manifest, like so:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: your-deployment
spec:
  selector:
    matchLabels:
      app: your-deployment
  template:
    metadata:
      labels:
        app: your-deployment
    spec:
      affinity:
        podAntiAffinity:
          # The other choice is requiredDuringSchedulingIgnoredDuringExecution.
          # Using it will prevent your pod from running on a node
          # that already has an instance of the pod on it.
          preferredDuringSchedulingIgnoredDuringExecution:
          - weight: 100
            podAffinityTerm:
              labelSelector:
                # This key-value is the label you've given to your Deployment resource
                matchExpressions:
                - key: app
                  operator: In
                  values:
                  - your-deployment
              # The topology is the set of resources that the anti-affinity acts on.
              # In this case, it compares among the hostnames (nodes).
              topologyKey: kubernetes.io/hostname
      containers:
      - name: your-deployment
        image: some-url

Integrate monitors and set up alert systems

This doesn’t come with Kubernetes natively, but there are services such as Datadog that offer this. This isn’t actually Kubernetes specific, since you should be doing this no matter what kind of server it is.

The usual monitors should be set up:

  • CPU Utilization
  • Memory Utilization
  • Disk Usage

For Kubernetes specific monitors, there are a few pod statuses that I would like to be notified about:

  • CrashloopBackOff
  • ImagePullBackOff
  • The number of available pods are less than the desired

Performance and Scalability

I’ve already covered a little about auto scaling both your nodes and pods in my previous Kubernetes guide. But here’s a few more things you can do with it.

External Metrics

When it comes to auto scaling, the main two metrics we usually use are CPU utilization and memory utilization. This is easy enough to set up, and is covered in my previous guide.

However, another metric that is often used is queue size, or queue latency. Maybe you have an MQ instance, or Kafka, or even Sidekiq. There are many queue based technologies out there.

There are two things that you’ll need to add: a custom metrics adapter, and an ExternalMetric resource.

Since I’m on the AWS stack, I’ll be using the k8s-cloudwatch-adapter. If your metrics are anywhere other than CloudWatch, you will have to look for an adapter that supports it.

The guide from AWS is here: https://aws.amazon.com/blogs/compute/scaling-kubernetes-deployments-with-amazon-cloudwatch-metrics/, but I’ll add my own guide as well.

To deploy the adapter, run this command. It’s a cluster-wide resource.

kubectl apply -f https://raw.githubusercontent.com/awslabs/k8s-cloudwatch-adapter/master/deploy/adapter.yaml

Your nodes will require the cloudwatch:GetMetricData IAM permission. I won’t cover how to do this.

Next, you’ll need to deploy an ExternalMetric resource. This has to be done in every namespace that requires it.

apiVersion: metrics.aws/v1alpha1
kind: ExternalMetric
metadata:
  name: mq-queue-size
spec:
  name: mq-queue-size
  resource:
    resource: "deployment"
  queries:
    - id: queue_size
      metricStat:
        metric:
          # The easiest way to figure out how to fill in these
          # fields is to open the metric itself on CloudWatch.
          namespace: "AWS/AmazonMQ"
          metricName: "QueueSize"
          dimensions:
            - name: Broker
              value: "broker-name"
            - name: Queue
              value: "queue-name"
        period: 300
        stat: Average
        unit: Count
      returnData: true

Then in your horizontal pod autoscaler, reference the ExternalMetric like this:

kind: HorizontalPodAutoscaler
apiVersion: autoscaling/v2beta1
metadata:
  name: your-autoscaler-name
spec:
  scaleTargetRef:
    apiVersion: apps/v1beta1
    kind: Deployment
    name: your-deployment-name
  minReplicas: 1
  maxReplicas: 10
  metrics:
  - type: External
    external:
      # This is the name of the ExternalMetric defined above
      metricName: mq-queue-size
      targetAverageValue: 5

With this, the Deployment should scale up when the average queue size exceeds 5.

Security

Use the simplest base images, and do not run the app as the root user

This isn’t Kubernetes specific, as it concerns your Docker images.

This is much easier to do on compiled languages such as Go, and harder to do on interpreted languages like Ruby on Rails, since they require the whole source code and Rails libraries to be on the image.

Here’s a Go template for building a Docker image. It’s a two stage Dockerfile, where the application is compiled, then in the second stage, the environment is prepared for the application to run in.

# Test and build
FROM golang as builder

ARG USER=unprivileged_user
ARG WORK_DIR=/app

# Allows cross-compilation for the compiled Go app
# This is needed because the two stages use different OSes, since cross-compilation is required.
# You should be able to remove the need for this if you use a golang-alpine variant (I haven't tested this).
ENV CGO_ENABLED 0

# update os and install required packages
RUN apt-get update -qq && apt-get upgrade -y

# Create directories for the user to avoid running as root
RUN groupadd ${USER} && useradd -m -g ${USER} -l ${USER}
RUN mkdir -p ${WORK_DIR} && chown -R ${USER}:${USER} ${WORK_DIR}

# set WORKDIR and USER
WORKDIR ${WORK_DIR}
USER ${USER}

# Copy the local package files to the container's workspace
COPY . ${WORK_DIR}

# verify the modules
RUN go mod verify

# Run tests
RUN go test -cover ./...

# Build app, produces ./app
RUN go build -o app ./main.go

# ----------------

# Produce docker image to run app
FROM alpine

ARG USER=unprivileged_user
ARG WORK_DIR=/app
ARG PORT=8080

ENV PORT=${PORT}

# Create directories for the user to avoid running as root
# https://github.com/mhart/alpine-node/issues/48
RUN addgroup -S ${USER} && adduser -S ${USER} -G ${USER}

# set WORKDIR
WORKDIR ${WORK_DIR}

# Copy the built file from the previous stage to the container's workspace
COPY --chown=0:0 --from=builder ${WORK_DIR}/app ${WORK_DIR}

# Install CA certificates to prevent x509: certificate signed by unknown authority errors
RUN apk update && \
    apk add ca-certificates && \
    update-ca-certificates && \
    rm -rf /var/cache/apk/*

USER ${USER}
EXPOSE ${PORT}

CMD ["./app"]

If you want to know why the simplest base image should be used, take a look at the difference between these two images, it’s the same code base, with the debian:buster image originally used:

Size and vulnerability stats

A smaller image size will also allow your containers to spin up more quickly, because it has to spend less time downloading the image. So, this doesn’t just have a security benefit, it helps performance as well.

 kubernetes  eks  aws  kubectl

Comments

Related Posts