11. The basics of stateful applications in Kubernetes

Hi.

Brendan Burns from Microsoft Azure and I'm going to talk about the basics of stateful applications in Kubernetes.

So with any stateful applications, you're going to need to think about how you do data resiliency and data replication.

In some cases, there's applications like MySQL that are actually quite difficult to do data replication.

You can do clustered MySQL, but it's not the standard setup.

With that in mind, in Kubernetes, you probably are going to do the same thing you would do with MySQL in general and run a single instance of your MySQL server associated with a persistent volume.

This persistent volume is going to be attached to Cloud-based storage, like Azure Disk, or on-prem, it might be attached to something like ISCSI.

In either case, that's going to be responsible for maintaining the state of your application even if your MySQL pod happens to move from machine one to machine two.

But when we start thinking about more Cloud-native applications, something like Cassandra or MongoDB, where replication is easier to achieve, then we need to start thinking about the primitives within Kubernetes.

Now when Kubernetes started, the only sort of way that you could do replication was using something called a replica set.

With a replica set, every single replica was treated entirely identically.

They have random hashes on the end of their application names.

And if a scaling event happens, for example a scaled down, a container is chosen at random and deleted.

These characteristics make replica set very hard to map to stateful applications.

In particular, many stateful applications expect their host names to be constant.

Sometimes you need to find a master or a root node to start from, a seed node to start from.

So those complexities of using replica sets and stateful applications led to the eventual development of stateful set.

A stateful set in Kubernetes, it's similar to a replica set, but it adds some guarantees that makes it easier to manage stateful applications inside of Kubernetes.

In particular, with a stateful set, the replicas have indices.

So we know that this is replica zero.

You know that this is replica one.

You know that this is replica two.

They have stable host names.

So something like my server zero and so on.

When Kubernetes decides to scale up or scale down a stateful set, it does it in a well understood way.

For example, when you initially create a stateful set the first replica is created, and Kubernetes waits for it to become healthy and available before creating the second replica.

This means that when the second replica is being created, you can depend on the fact that the zeroth index is available for you to connect to.

The same thing when the first replica becomes healthy the next replica will be created, and it likewise can point back to the original member of the stateful set.

This makes it much easier to rendezvous to declare an initial leader, and many other things that are necessary when you're creating stateful applications.

Likewise, when you scale down Kubernetes will also delete starting at the highest index.

So if you scale this from three replicas to two replicas, it's going to start by depleting this replica index two over here.

This again makes it easier to control the way that an application behaves on a scale down event.

Stateful sets also provide for the ability to develop DNS names that actually target individual replicas.

So with a replica set, you might have a service, and it would target a replica set.

Maybe this is called front end.

This creates a DNS entry, but it only creates a DNS entry for front end.

If you're using a stateful set, you can actually create a service such that you get a DNS entry for maye Cassandra, which would go to any of the replicas of the Cassandra cluster, but you also get cassandra zero.cassandra, and likewise dash one dash two and for each replica.

This means that if you don't care which replica of the stateful application you want to go to, for example, reading from this Cassandra cluster, you can always use this service and you'll get load balancing and everything else that you expect.

But if you need to know a specific replica, you can still use naming to discover a pointer to each specific index in the stateful set.

That also makes it easier for you to configure your application, configure a client that may need an explicit list.

Because this host name stays stable as well, no matter how you scale up or scale down, you can be assured that these DNS names will always stay constant for as long as that stateful set lives.

But again obviously, when you're creating these stateful sets, you're going to be thinking about persistence as well.

In some cases, using local persistence may be okay.

There are Cloud-native storage applications out there, and Cassandra is a great example, where there's actually replication going on between the replicas by the application itself.

In that world, you may not need to use persistent volumes because the data itself is replicated between all of the members of the cluster.

But if you do choose to use persistent volume, you're going to need to use persistent volume claims.

A persistent volume claim in the context of your stateful set, so that when that stateful set is created and each replica is created, Kubernetes will go ahead and create a different disk for each replica in the stateful set.

So hopefully that gives you an illustration of how stateful applications are developed in Kubernetes, how you can use either a singleton pattern for more traditional stateful applications or the Cloud-native stateful set resource that's available in Kubernetes with or without persistent volumes to develop their stateful applications inside of Kubernetes.