Kubernetes StatefulSet

Links: 110 Kubernetes Index

---

StatefulSet

• It is a k8s component used specifically for stateful applications.
  • Databases are example of stateful applications

Stateless applications are deployed using deployment component whereas stateful applications are deployed using StatefulSet.

Pod Identity

• In StatefulSet all the replica pods are not identical.

• Every pod has its own identifier.

  • Unlike deployment where we have random hashes at the end, StatefulSet gets fixed ordered names (statefulsetname-ordinal).
    • 
  • So if you create a StatefulSet called mysql with 3 replicas you will have pods with the following names
    • 
    • First is the master then comes the slave in the order of startup.
  • Next pod is only created if previous is up & running.
  • Deletion starts from the last pod.
    • 

• Each pod has a persistent identifier which it maintains while rescheduling.

  • This means that if a pod of ID-2 dies then a new pod of the same ID will replace it.

Why it's difficult to scale database applications?

How are database applications scaled?

There can be only 1 master database which takes writes.

• The pods don't share the same physical storage even though they use the same data.
  • 
• They each have their own replica of their storage which they can access.
  • This means to have the same data as other pods the data must be synchronised.
• The worker pods (slaves) must know about each and every changes made to master so that it can be upto date.
  • 
• When a new pod joins it must take care of replicating the data
  • 
  • It first clones the data from the previous pod and then starts synchronisation.

Since all the pods have a copy of the data you can actually get away with only having temporary data storage.

- But this means that all data will be lost when all the pods die. - But in case of persistent storage date will survive even if all the pods die since Persistent Volume lifecycle isn't tied to other component's lifecycle.

Pod State

• Each pod has its own persistent volume plus the pod state.

  • Pod state has information that whether its a master pod or a slave pod.
  • 

• When a pod dies & gets replaced the persistent pod identifiers makes sure that the storage volume gets reattached to the replacement pod.

  • Since the storage has the state of the pod.
  • 

• In case when pods are deleted, pods are deleted but their persistent volume is kept.

  • 

For the reattachment to work it is important to use remote storage.

• Since if the pod gets rescheduled from one node to another then the previous storage must be available on other nodes as well.
• We cannot use local volume storage as it is attached to a particular node.
• 

Pod Endpoints

• Each pod has 2 endpoints.

  • 
  • LoadBalancer Service
  • Individual Service Name
    • 

• We need to have a headless service with a StatefulSet for individual service name.

Having a fixed endpoint & pod name ensures that pods have sticky identity to retain state & retain role.

Stateful applications not perfect for containerised environments

References

• Kubernetes StatefulSet simply explained | Deployment vs StatefulSet - YouTube
• Understanding StatefulSets in Kubernetes - YouTube
• Demos:
  • [ Kube 21 ] How to use Statefulsets in Kubernetes Cluster - YouTube

---

Last updated: 2022-09-06