Mastering Kubernetes

It all starts and ends with hardware. In order to run your workloads, you need some real hardware provisioned. That includes actual physical machines, with certain compute capabilities (CPUs or cores), memory, and some local persistent storage (spinning disks or SSDs). In addition, you will need some shared persistent storage and to hook up all these machines using networking so they can find and talk to each other. At this point, you run multiple virtual machines on the physical machines or stay at the bare-metal level (no virtual machines). Kubernetes can be deployed on a bare-metal cluster (real hardware) or on a cluster of virtual machines. Kubernetes in turn can orchestrate the containers it manages directly on bare-metal or on virtual machines. In theory, a Kubernetes cluster can be composed of a mix of bare-metal and virtual machines, but this is not very common.

Containers are ideal to package microservices because, while providing isolation to the microservice, they are very lightweight and you don't incur a lot of overhead when deploying many microservices as you do with virtual machines. That makes containers ideal for cloud deployment, where allocating a whole virtual machine for each microservice would be cost prohibitive.

All major cloud providers, such as AWS, GCE, and Azure, provide container hosting services these days. Some of them, such as Google's GKE, are based on Kubernetes. Others, such as Microsoft Azure's container service, are based on other solutions (Apache Mesos). By the way, AWS has the ECS (the containers service over EC2), which uses their own orchestration solution. The great thing about Kubernetes is that it can be deployed on all those clouds. Kubernetes has a cloud provider interface that allows any cloud provider to implement it and integrate Kubernetes seamlessly.

In the olden days, when systems were small, each server had a name. Developers and users knew exactly what software was running on each machine. I remember that, in many of the companies I worked for, we had multi-day discussions to decide on a naming theme for our servers. For example, composers and Greek mythology characters were popular choices. Everything was very cozy. You treated your servers like beloved pets. When a server died it was a major crisis. Everybody scrambled to try to figure out where to get another server, what was even running on the dead server, and how to get it working on the new server. If the server stored some important data, then hopefully you had an up-to-date backup and maybe you'd even be able to recover it.

Obviously, that approach doesn't scale. When you have a few tens or hundreds of servers, you must start treating them like cattle. You think about the collective and not individuals. You may still have some pets (that is, your build machines), but your web servers are just cattle.

Kubernetes takes the cattle approach to the extreme and takes full responsibility for allocating containers to specific machines. You don't need to interact with individual machines (nodes) most of the time. This works best for stateless workloads. For stateful applications, the situation is a little different, but Kubernetes provides a solution called StatefulSet, which we'll discuss soon.

In this section, we covered the idea of container orchestration and discussed the relationships between hosts (physical or virtual) and containers, as well as the benefits of running containers in the cloud, and finished with a discussion about cattle versus pets. In the following section, we will get to know the world of Kubernetes and learn its concepts and terminology.

A cluster is a collection of hosts storage and networking resources that Kubernetes uses to run the various workloads that comprise your system. Note that your entire system may consist of multiple clusters. We will discuss this advanced use case of federation in detail later.

A node is a single host. It may be a physical or virtual machine. Its job is to run pods. Each Kubernetes node runs several Kubernetes components, such as a kubelet and a kube proxy. Nodes are managed by a Kubernetes master. The nodes are worker bees of Kubernetes and shoulder all the heavy lifting. In the past they were called minions. If you read some old documentation or articles, don't get confused. Minions are nodes.

The master is the control plane of Kubernetes. It consists of several components, such as an API server, a scheduler, and a controller manager. The master is responsible for the global, cluster-level scheduling of pods and handling of events. Usually, all the master components are set up on a single host. When considering high-availability scenarios or very large clusters, you will want to have master redundancy. I will discuss highly available clusters in detail in Chapter 4, High Availability and Scaling.

A pod is the unit of work in Kubernetes. Each pod contains one or more containers. Pods are always scheduled together (always run on the same machine). All the containers in a pod have the same IP address and port space; they can communicate using localhost or standard inter-process communication. In addition, all the containers in a pod can have access to shared local storage on the node hosting the pod. The shared storage will be mounted on each container. Pods are important feature of Kubernetes. It is possible to run multiple applications inside a single Docker container by having something like supervisor as the main Docker application that runs multiple processes, but this practice is often frowned upon, for the following reasons:

Pods provide a great solution for managing groups of closely related containers that depend on each other and need to co-operate on the same host to accomplish their purpose. It's important to remember that pods are considered ephemeral, throwaway entities that can be discarded and replaced at will. Any pod storage is destroyed with its pod. Each pod gets a unique ID (UID), so you can still distinguish between them if necessary.

Labels are key-value pairs that are used to group together sets of objects, very often pods. This is important for several other concepts, such as replication controller, replica sets, and services that operate on dynamic groups of objects and need to identify the members of the group. There is a NxN relationship between objects and labels. Each object may have multiple labels, and each label may be applied to different objects. There are certain restrictions by design on labels. Each label on an object must have a unique key. The label key must adhere to a strict syntax. It has two parts: prefix and name. The prefix is optional. If it exists then it is separated from the name by a forward slash (/) and it must be a valid DNS sub-domain. The prefix must be 253 characters long at most. The name is mandatory and must be 63 characters long at most. Names must start and end with an alphanumeric character (a-z, A-Z, 0-9) and contain only alphanumeric characters, dots, dashes, and underscores. Values follow the same restrictions as names. Note that labels are dedicated for identifying objects and not for attaching arbitrary metadata to objects. This is what annotations are for (see the following section).

Annotations let you associate arbitrary metadata with Kubernetes objects. Kubernetes just stores the annotations and makes their metadata available. Unlike labels, they don't have strict restrictions about allowed characters and size limits. In my experience, you always need such metadata for complicated systems, and it is nice that Kubernetes recognizes this need and provides it out of the box so you don't have to come up with your own separate metadata store and mapping object to their metadata.

We've covered most, if not all, of Kubernetes' concepts; there are a few more I mentioned briefly. In the next section, we will continue our journey into Kubernetes architecture by looking into its design motivations, the internals and implementation, and even pick at the source code.

Label selectors are used to select objects based on their labels. Equality-based selectors specify a key name and a value. There are two operators, = (or ==) and !=, for equality or inequality based on the value. For example:

role = webserver

This will select all objects that have that label key and value.

Label selectors can have multiple requirements separated by a comma. For example:

role = webserver, application != foo

Set-based selectors extend the capabilities and allow selection based on multiple values:

role in (webserver, backend)

Replication controllers and replica sets both manage a group of pods identified by a label selector and ensure that a certain number is always up and running. The main difference between them is that replication controllers test for membership by name equality and replica sets can use set-based selection. Replica sets are newer and designated as the next-generation replication controllers. They are still in beta and are not fully supported by all the tools at the time of writing. Hopefully, by the time you read this, they will be full-fledged members.

Kubernetes guarantees that you will always have the same number of pods running as you specified in a replication controller or a replica set. Whenever the number drops due to a problem with the hosting node or the pod itself, Kubernetes will fire up new instances. Note that, if you manually start pods and exceed the specified number, the replication controller will kill some extra pods.

Replication controllers used to be central to many workflows, such as rolling updates and running one-off jobs. As Kubernetes evolved, it introduced direct support for many of these workflows, with dedicated objects such as Deployment, Job, and DaemonSet. We will meet them all later.

Services are used to expose some functionality to users or other services. They usually encompass a group of pods, usually identified by – you guessed it – a label. You can have services that provide access to external resources, or to pods you control directly at the virtual IP level. Native Kubernetes services are exposed through convenient endpoints. Note that services operate at layer 3 (TCP/UDP). Kubernetes 1.2 added the Ingress object, which provides access to HTTP objects. More on that later. Services are published or discovered via one of two mechanisms: DNS, or environment variables. Services can be load-balanced by Kubernetes. But, developers can choose to manage load balancing themselves in case of services that use external resources or require special treatment.

There are many gory details associated with IP addresses, virtual IP addresses, and port spaces. We will discuss them in depth in a future chapter.

Local storage on the pod is ephemeral and goes away with the pod. Sometimes that's all you need, if the goal is just to exchange data between containers of the node, but sometimes it's important for the data to outlive the pod, or it's necessary to share data between pods. The volume concept supports that need. Note that, while Docker has a volume concept too, it is quite limited (although getting more powerful). Kubernetes uses its own separate volumes. Kubernetes also supports additional container types such as rkt, so it couldn't rely on Docker volumes even in principle.

There are many volume types. Kubernetes currently directly supports each volume type. In the future, another layer of indirection may be added and an abstract volume plugin may be developed. The emptyDir volume type mounts a volume on each container that is backed by default by whatever is available on the hosting machine. You can request a memory medium if you want. This storage is deleted when the pod is terminated for any reason. There are many volume types for specific cloud environments, various networked filesystems, and even Git repositories. An interesting volume type is the persistentDiskClaim, which abstracts the details a little bit and uses the default persistent storage in your environment (typically in a cloud provider).

Pods come and go, and if you care about their data then you can use persistent storage. That's all good. But sometimes you want Kubernetes to manage a distributed data store such as Kubernetes or MySQL Galera. These clustered stores keep the data distributed across uniquely identified nodes. You can't model that with regular pods and services. Enter StatefulSet. If you remember earlier, I discussed pets versus cattle and how cattle is the way to go. Well, StatefulSet sits somewhere in the middle. StatefulSet ensures (similar to a replication controller) that a given number of pets with unique identities are running at any given time. Pets have the following properties:

StatefulSet can help with peer discovery as well as adding or removing pets.

Secrets are small objects that contain sensitive info such as credentials and tokens. They are stored as plaintext in etcd, accessible by the Kubernetes API server, and can be mounted as files into pods (using dedicated secret volumes that piggyback on regular data volumes) that need access to them. The same secret can be mounted into multiple pods. Kubernetes itself creates secrets for its components, and you can create your own secrets. Another approach is to use secrets as environment variables. Note that secrets in a pod are always stored in memory (tmpfs in the case of mounted secrets) for better security.

Each object in Kubernetes is identified by a UID and a name. The name is used to refer to the object in API calls. Names should be up to 253 characters long and use lowercase alphanumeric characters, dash (-) and dot (.). If you delete an object, you can create another object with the same name as the deleted object, but the UIDs must be unique across the lifetime of the cluster. The UIDs are generated by Kubernetes, so you don't have to worry about it.

A namespace is a virtual cluster. You can have a single physical cluster that contains multiple virtual clusters segregated by namespaces. Each virtual cluster is totally isolated from other virtual clusters, and they can only communicate through public interfaces. Note that Node objects and persistent volumes don't live in a namespace. Kubernetes may schedule pods from different namespaces to run on the same node. Likewise, pods from different namespaces can use the same persistent storage.

When using namespaces, you have to consider network policies and resource quotas to ensure proper access and distribution of the physical cluster resources.

All happy (working) distributed systems are alike, to paraphrase Tolstoy in Anna Karenina. That means that, to function properly, all well-designed distributed systems must follow some best practices and principles. Kubernetes doesn't want to be just a management system. It wants to support and enable these best practices and provide high-level services to developers and administrators. Let's look at some of those described as design patterns.

This is the main API of Kubernetes. It is huge. All the concepts we discussed before, and many auxiliary concepts, have corresponding API objects and operations. If you have the right permissions you can list, get, create, and update objects. Here is a detailed documentation of one of the most common operations, get a list of all the pods:

GET /api/v1/pods

It accepts various query parameters (all optional):

The autoscaling API is very focused and lets you control the horizontal pod autoscaler, which manages a group of pods based on CPU utilization and even application-specific metrics. You can list, query, create, update, and destroy autoscaler objects using the /apis/autoscaling/v1 endpoint.

The master components typically run on one node, but in a highly available or very large cluster, they may be spread across multiple nodes.

Nodes in the cluster need a couple of components to interact with the cluster master components, receive workloads to execute, and update the cluster on their status.

The runtime interface for containers is specified in the Kubernetes project on GitHub. Kubernetes is open source, so we can look at it at the following URL:

https://github.com/kubernetes/kubernetes/blob/master/pkg/kubelet/container/runtime.go.

I'll present here snippets from this file without the elaborate comments. Even if you're not a full-fledged programmer and know nothing about the Go language, you should be able to grasp the scope and responsibilities of a runtime engine from the viewpoint of Kubernetes:

type Runtime interface {
  Type() string

  Version() (Version, error)

  APIVersion() (Version, error)

  Status() error

  GetPods(all bool) ([]*Pod, error)
}

The fact that it is an interface means that Kubernetes doesn't provide an implementation. The first group of methods provides general information about the runtime: Type, Version, APIVersion, and Status. You can also get all the pods:

SyncPod(pod *api.Pod, apiPodStatus api.PodStatus, podStatus *PodStatus, pullSecrets []api.Secret, backOff *flowcontrol.Backoff) PodSyncResult

KillPod(pod *api.Pod, runningPod Pod, gracePeriodOverride *int64) error

GetPodStatus(uid types.UID, name, namespace string) (*PodStatus, error)

GetNetNS(containerID ContainerID) (string, error)

GetPodContainerID(*Pod) (ContainerID, error)

GetContainerLogs(pod *api.Pod, containerID ContainerID, logOptions *api.PodLogOptions, stdout, stderr io.Writer) (err error)

DeleteContainer(containerID ContainerID) error

The next group of methods deal mostly with pods appropriately as this is the main abstraction in the Kubernetes conceptual model. Then there is the GetPodContainerID(), which gets you from a container to a pod, and a few more container-related methods:

The last three items, ContainerCommandRunner, ContainerAttacher, and ImageService, are interfaces that the runtime interface inherits. This means that whoever implements the runtime interface also needs to implement the methods of these interfaces. The interfaces are defined in the same file. Just the interface names provide a lot of information about what they do. Kubernetes obviously needs to run commands in containers, and it needs to attach containers to its pods and pull container images. I encourage you to pursue this file and get familiar with the code.

Now that you are familiar at the code level with what Kubernetes considers as a runtime engine, let's look at the individual runtime engines briefly.

Docker is, of course, the 800 pound gorilla of containers. Kubernetes was originally designed to manage only Docker containers. The multi-runtime capability was first introduced in Kubernetes 1.3. Until then, Kubernetes could only manage Docker containers.

I assume you're very familiar with Docker and what it brings to the table if you are reading this book. Docker enjoys tremendous popularity and growth, but there is also a lot of criticism toward it. Critics often mention the following concerns:

Docker is aware of the criticisms and has addressed some of these concerns. In particular, Docker invested in its Docker swarm product. Docker swarm is a Docker-native orchestration solution that competes with Kubernetes. It is simpler to use than Kubernetes, but it's not as powerful or mature.

Starting with Docker 1.11, released on April 2016, Docker has changed the way it runs containers. The runtime now uses containerd and runC to run Open Container Initiative (OCI) images in containers:

Rkt is a new container manager from CoreOS (developers of the CoreOS Linux distro, etcd, flannel, and more). The rkt runtime prides itself on its simplicity and strong emphasis on security and isolation. It doesn't have a Daemon like the Docker engine and relies on the OS init system, such as systemd, to launch the rkt executable. Rkt can download images (both App Container (appc) images and OCI images), verify them, and run them in containers. Its architecture is much simpler.

Hyper containers are another option. A Hyper container has a lightweight VM (its own guest kernel) and it runs on bare metal. Instead of relying on Linux cgroups for isolation, it relies on a hypervisor. This approach presents an interesting mix compared to standard bare-metal clusters that are difficult to set up and public clouds where containers are deployed on heavyweight VMs.

Hypernetes

Hypernetes is a multi-tenant Kubernetes distribution that uses Hyper containers as well as some OpenStack components for authentication, persistent storage, and networking. Since containers don't share the host kernel, it is safe to run containers of different tenants on the same physical host:

In this section, we've covered the various runtime engines that Kubernetes supports as well as the trend toward standardization and convergence. In the next section, we'll take a step back and look at the big picture, and how Kubernetes fits into the CI/CD pipeline.

A CI/CD pipeline is a set of steps that a set of changes by developers or operators that modify the code, data or configuration of a system, test them and deploys them to production. Some pipelines are fully automated and some are semi-automated with human checks. In large organizations, there may be test and staging environments where changes are deployed to automatically, but release to production requires manual intervention. The following diagram describes a typical pipeline.

It may be worth mentioning that developers can be completely isolated from production infrastructure. Their interface is just a Git workflow, where a good example is Deis Workflow (PaaS on Kubernetes, similar to Heroku):

When your deployment target is a Kubernetes cluster, you should rethink some traditional practices. For starters, packaging is different. You need to bake images for your containers. Reverting code changes is super easy and instantaneous by using smart labeling. It gives you a lot of confidence that, if a bad change slips through the testing net, somehow you'll be able to revert to the previous version immediately. But you want to be careful there. Schema changes and data migrations can't be automatically rolled back. Another unique capability of Kubernetes is that developers can run a whole cluster locally. That takes some work when you design your cluster, but since the microservices that comprise your system run in containers, and those containers interact via APIs, it is possible and practical to do. As always, if your system is very data-driven, you will need to accommodate for that and provide data snapshots and synthetic data that your developers can use.