[Nottingham] Copy without copying? (file copy deduplication)

Martin martin at ml1.co.uk
Mon Jun 10 14:54:03 UTC 2019


Thanks for that. Your link gives a very good explanation for LVM
including writing to the snapshot itself also.

That's also a good reminder that LVM is a good use-case for handling the
disk images for VMs...


On 10/06/2019 14:49, VM via Nottingham wrote:
> A good explanation of different uses for read-write snapshots is at
> https://www.clevernetsystems.com/lvm-snapshots-explained/
> On June 10, 2019 1:27:54 PM UTC, VM via Nottingham <nottingham at mailman.lug.org.uk> wrote:
>> Martin, LVM snapshots can be mounted read-write and keep modifications
>> made. They also can be merged with the origin if changes are desired.
>> For instance, make a snapshot, mount it, upgrade, check everything is
>> fine, merge snapshot with the origin. It's another use of LVM
>> snapshots, different to read-only backups.
>> On June 10, 2019 12:37:55 PM UTC, Martin via Nottingham
>> <nottingham at mailman.lug.org.uk> wrote:
>>> On 10/06/2019 12:06, J J via Nottingham wrote:
>>>> I must be misunderstanding something, I thought the snapshot
>> retained
>>>> the original data as things continued to get written to the source.
>>>> (This seems to be what LVM does at any rate)
>>> On LVM, when you enable a snapshot device:
>>> The source device is used as normal, unaffected, with data read and
>> new
>>> data written as normal.
>>> While the LVM snapshot device is enabled, any disk blocks written to
>>> the
>>> source device first have the old block data copied to the snapshot
>>> device. Hence you get the original LVM source device in use as normal
>>> and a copy of the old data before being overwritten, block by block,
>>> copied across to the LVM snapshot device. I guess that can be called a
>>> sort of CoW but used instead for 'saving' the old data.
>>> When reading from the LVM snapshot device, any blocks that are
>>> unchanged
>>> are read from the source device as normal. Any changed blocks are
>>> instead read from the copy saved on the snapshot device to make the
>>> source device appear to be unchanged.
>>> Hence, you have the read-only snapshot view and additionally, the
>>> snapshot device area can be much smaller than the source device. (Only
>>> the changed data is stored...)
>>>> Which is great, but kinda the wrong way round for my use case.
>>>> I effectively want the changes written into the snapshot, leaving
>> the
>>>> source untouched.
>>>> Thus I can have as many snapshots as I need (parallel tests etc) and
>>>> then chuck them.
>>>> Much more like the differencing you get with Hypervisors (and
>>> OverlayFS,
>>>> just at the block level to reduce space requirements).
>>> And that is exactly what you can do on btrfs with the inherent magic
>> of
>>> CoW:
>>> Simplistically, all a btrfs snapshot is is an atomic snapshot of the
>>> b-tree indexing. The file data stays in place.
>>> You then have, ready for the very next operation, two identical views
>>> of
>>> the same data and attributes. There is no distinction as to which is a
>>> snapshot of whichever. You have taken ' _a_ snapshot' whereby you now
>>> have two identical views with access from two different points on the
>>> filesystem.
>>> Both views are read-writeable in whatever way you wish.
>>> Or you can restrict whichever or both to read-only in whatever way
>>> wished.
>>>> If XFS or something can do that, awesome, and I'll start looking at
>>> it
>>>> as soon as Kubuntu gets itself out of whatever APT fankle it's
>>> managed
>>>> to get into!
>>> No need for XFS (and in any case, I'm a little prejudiced against its
>>> block allocation ways of working and how that must fit your use case -
>>> whereas btrfs is more flexible for extremes of use cases).
>>>> Not sure how viable BTRFS would be as a choice, I thought the
>> project
>>>> was slowly dying off (I know RedHat has dropped it) and any
>> proposals
>>>> will need to be passed by Internal IT, so can't be too exotic.
>>> btrfs is very much alive and is used by and supported by the big cloud
>>> providers including Facebook...
>>> Red Hat only dropped support for some of their special Red Hat kernel
>>> versions for the sake of long term maintainability.
>>> btrfs is still developing apace for new features. However, their main
>>> features and disk structure can be considered long stable.
>>> I've been using btrfs for about a decade to good effect and without
>> any
>>> adverse issues and spanning many TBytes of data.
>>> Hope that's of help,
>>> Cheers,
>>> Martin
>>> ps: Corrections welcomed!
>>>> On Mon, 10 Jun 2019 at 11:24, Martin via Nottingham
>>>> <nottingham at mailman.lug.org.uk
>>> <mailto:nottingham at mailman.lug.org.uk>>
>>>> wrote:
>>>>     Jason,
>>>>     The obvious fix for that is to use btrfs whereby:
>>>>     You have your pristine folders in one btrfs subvolume (can even
>>> be
>>>>     mounted or set to be read-only);
>>>>     Then create a snapshot subvolume from that first subvolume;
>>>>     Your new snapshot subvolume can then be used or mounted
>>> read-write.
>>>>     Aside: Note that you don't even need to mount the subvolumes as
>>> separate
>>>>     mounts: they can even appear as 'normal' directories on a
>>> (single)
>>>>     higher mount point.
>>>>     The snapshot volume is created with zero copying of files and is
>>> very
>>>>     fast and low resource to do. You'll see no increase in disk
>>> usage!
>>>>     During use, the only file writing will be only for whatever new
>>> data is
>>>>     actually written. The magic of the btrfs CoW means that only new
>>> blocks
>>>>     of data get written. Unchanged file fragments remain unchanged
>>> and
>>>>     uncopied...
>>>>     Hope that fits your needs?
>>>>     Cheers,
>>>>     Martin
>>>>     On 10/06/2019 11:04, J J via Nottingham wrote:
>>>>     > I need read/write access to the new folder.
>>>>     > Mucking around with OverlayFS I can effectively have a
>>> "Template"
>>>>     folder
>>>>     > and then a "Working" folder (plus a couple of others for
>>> reasons) and
>>>>     > that works well. Except for the fact that full file copies
>>> still
>>>>     happen
>>>>     > and some of these files are big.
>>>>     > So the situation is definitely better, but far from ideal.
>>>>     >
>>>>     > I am going to see if I can grok XFS enough to do the similar
>>>>     thing, but
>>>>     > do some block-level magics.
>>>>     > LVM snapshots might also do it, but not sure if they would
>>> allow for
>>>>     > multiple, concurrent "Working" folders and AIUI that would
>> mean
>>>>     mucking
>>>>     > around with how the system is deployed, not something I have
>>>>     control over.
>>>>     >
>>>>     > The other wrinkle is that this needs t be done ad hoc.
>>>>     > But first things first...

More information about the Nottingham mailing list