Iduoad's links

The Goal: Workload Portability

• **Workload portability **: write your app once and deploy it anyware. This is achieved by abstracting away cluster details and decoupling the infra from the apps.
• Pods and replicaSets abstract compute and memory very well, but what about state ?
  • P1: Containers are ephemeral (data is lost if the container crashes).
  • P2: Containers can’t share data between them.

Challenges with abstracting storage:

• So many different types of storage:
  • Object storage
  • SQL databases
  • NoSQL databases
  • Pub/Sub systems
  • Time series databases
  • File Storage
  • Block Storage
  • File on Block
  • …

Solution

• K8s dev focus on the following types because the data path is standardized (Posix, ISCSI)
  • File Storage
  • Block Storage
  • File on Block

Volume and Volume Plugins

• The way K8s abstracts away the storage part is by using Volume plugins. They specify:
  • How the volume is setup in the pod
  • Medium that backs it.
• Volume have a lifetime same as the pod’s or longer.
• There are 5 categories or volume plugins that k8s supports:
  1. Remote storage:
     • AWS ebs
     • GCE PD
     • …
  2. Ephemeral Storage:
     • EmptyDir
     • Expose k8s API
       • Secret
       • ConfigMap
       • Downward API
  3. Local Persistent Volumes (Beta)
  4. Out-of-Tree
     • Flex (exec a Binary)
     • CSI (Beta)
  5. Other
     • Host Path

Ephemeral Storage

• Referenced in-line, not via PV/PVC.
• Volumes that expose k8s API objects are based on emptyDirs

Remote Storage

• Can be referenced using PV/PVC.
• PV/PVC decouples volumes implementation (PV) from their consumption (PVC)
• PVs can be provisioned statically or dynamically using Storage Classes.

Host Paths

• Host paths are to use with caution. Never use them unless you know what you are doing.
• One use case is to use them along with Node Affinity to keep some stateful apps in some nodes.

Local Persistent Volume

• They expose a local block or file as PV.
• 2 use cases:
  • Building distributed storage systems on top of k8s
  • Building high performance caching.

Out-of-Tree

• The volume plugins are in tree i.e. they live inside k8s codebase.
  • Powerful abstraction for block and file storage.
  • Automate Provisioning.
  • Storage portability (PV/PVC)
• There are two iniative of bringing storage out of k8s code tree:
  • CSI (Container Storage Interface): Plugins may be containerized.
  • Flex Volumes: exec based i.e. binaries that need to be installed on the nodes (slave and master). It needs Master access.