rawfile-localpv/README.md
2021-10-02 21:22:25 +03:30

3.5 KiB
Raw Blame History

FOSSA Status

RawFilePV

Kubernetes LocalPVs on Steroids

Prerequisite

  • Kubernetes: 1.21+

Install

helm install -n kube-system rawfile-csi ./deploy/charts/rawfile-csi/

Usage

Create a StorageClass with your desired options:

apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
  name: my-sc
provisioner: rawfile.csi.openebs.io
reclaimPolicy: Delete
volumeBindingMode: WaitForFirstConsumer
allowVolumeExpansion: true

Features

  • Direct I/O: Near-zero disk performance overhead
  • Dynamic provisioning
  • Enforced volume size limit
  • Access Modes
    • ReadWriteOnce
    • ReadOnlyMany
    • ReadWriteMany
  • Volume modes
    • Filesystem mode
    • Block mode
  • Volume metrics
  • Supports fsTypes: ext4, btrfs, xfs
  • Online expansion: If fs supports it (e.g. ext4, btrfs, xfs)
  • Online shrinking: If fs supports it (e.g. btrfs)
  • Offline expansion/shrinking
  • Ephemeral inline volume
  • Snapshots: If the fs supports it (e.g. btrfs)

Motivation

One might have a couple of reasons to consider using node-based (rather than network-based) storage solutions:

  • Performance: Almost no network-based storage solution can keep up with baremetal disk performance in terms of IOPS/latency/throughput combined. And youd like to get the best out of the SSD youve got!
  • On-premise Environment: You might not be able to afford the cost of upgrading all your networking infrastructure, to get the best out of your network-based storage solution.
  • Complexity: Network-based solutions are distributed systems. And distributed systems are not easy! You might want to have a system that is easier to understand and to reason about. Also, with less complexity, you can fix unpredicted issues more easily.

Using node-based storage has come a long way since k8s was born. Right now, OpenEBSs hostPath makes it pretty easy to automatically provision hostPath PVs and use them in your workloads. There are known limitations though:

  • You cant monitor volume usage: There are hacky workarounds to run “du” regularly, but that could prove to be a performance killer, since it could put a lot of burden on your CPU and cause your filesystem cache to fill up. Not really good for a production workload.
  • You cant enforce hard limits on your volumes size: Again, you can hack your way around it, with the same caveats.
  • You are stuck with whatever filesystem your kubelet node is offering
  • You cant customize your filesystem:

All these issues stem from the same root cause: hostPath/LocalPVs are simple bind-mounts from the host filesystem into the pod.

The idea here is to use a single file as the block device, using Linuxs loop, and create a volume based on it. That way:

  • You can monitor volume usage by running df in O(1) since devices are mounted separately.
  • The size limit is enforced by the operating system, based on the backing file size.
  • Since volumes are backed by different files, each file could be formatted using different filesystems, and/or customized with different filesystem options.

License

FOSSA Status