[Gluster-devel] Introducing Heketi: Storage Management Framework with Plugins for GlusterFS volumes

[Jeff Darcy]

> LVM or Volume Manager Dependencies:
> 1) SNAPSHOTS: Gluster snapshots are LVM based

The current implementation is LVM-centric, which is one reason uptake has
been so low.  The intent was always to make it more generic, so that other
mechanisms could be used as well.

> 2) PROVISIONING and ENFORCEMENT:
> As of today Gluster does not have any control on the size of the brick. It
> will consume the brick (xfs)mount point given
> to it without checking on how much it needs to consume. LVM (or any other
> volume manager) will be required to do space provisioning per brick and
> enforce
> limits on size of bricks.

Some file systems have quota, or we can enforce our own.

> 3) STORAGE SEGREGATION:
> LVM pools can be used to have storage segregation i.e having primary storage
> pools and secondary(for Gluster replica) pools,
> So that we can crave out proper space from the physical disks attached to
> each node.
> At a high level (i.e Heketi's User) disk space can be viewed as storage
> pools,(i.e by aggreating disk space per pool per node using glusterd)
> To start with we can have Primary pool and secondary pool(for Gluster
> replica) , where each file serving node in the cluster participates in
> these pools via the local LVM pools.

This functionality in no way depends on LVM.  In many cases, mere
subdirectories are sufficient.

> 4) DATA PROTECTION:
>    Further data protection using LVM RAID. Pools can be marked to have RAID
>    support on them, courtesy of LVM RAID.
>    https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/6/html/Logical_Volume_Manager_Administration/raid_volumes.html

Given that we already have replication, erasure coding, etc. many users would
prefer not to reduce storage utilization even further with RAID.  Others
would prefer to get the same functionality without LVM, e.g. with ZFS.  That's
why RAID has always been - and should remain - optional.

It's fine that we *can* use LVM features when and where they're available.
Building in *dependencies* on it has been a mistake every time, and repeating
a mistake doesn't make it anything else.