[Novalug] LVM Snapshots

[James Ewing Cottrell 3rd]

  On 12/18/2010 11:26 PM, Peter Larsen wrote:
> On Sat, 2010-12-18 at 22:34 -0500, James Ewing Cottrell 3rd wrote:
>> Thanks for the answer, but you really didn't TELL me anything.
> I'm pretty sure I did; at least that was the intention.
OK, I'll blame it on the medium then. How can anyone expect anything of 
quality to come out of a phone?
>> When you make a snapshot, the UUIDs are the same too.
> Ahhh - no. But I think I understand what's going on now. I think you've
> been focused on the file-system configuration on the volume. Not the
> volume itself. Try NOT using the file system at all when using LVM to
> identify the volume. Use the volume ID (UUID) instead. This will change
> and be different from the master volume.
My question was about running on a snapshot; Labels vs UUIDs is a side 
issue.
> As American Dave points out, the UUID on the volume does indeed change.
> But you're right too - the UUID of the file system does not. So - don't
> use the file system UUID. Don't use the file system label. Not with LVM
> at the very least.
Yes, I misremembered what I had done. See my followup posting to him.
>> And it's a lot more intuitive to change labels (e2label) than it is to
>> change UUIDs.
> Intuitive maybe. But since you're looking for uniqueness a very bad idea
> to use labels.
Labels are as unique as the person who makes them. And yes, I forgot to 
change them. But ultimately, grub picked one, and booted off that. And 
while that decision may be a surprising one, it doesn't change whether 
it was safe or not.

Remember, my complaint was that I lost files from BOTH filesystems! And 
while its possible that I simply deleted them, I had that other incident 
with snapshots earlier this year. Of course, that one was on top of 
Metadisk Mirrors, and Bryan warned against that, altho I never 
understood why that was bad.
>> I have done it before, but I'd have to look it up. Besides, I
>> generally make my labels the same name as the LV, and since I usually
>> have only one VG, they *are* unique.
> You can definitely define them as such but there's no guarantee by the
> system they're unique. If you follow my advice above, you'll never
> encounter problems when managing volumes vs. how/where to mount them.
UUIDs are ugly...they are Numbers...like MAC addresses. I want NAMES 
that *I* choose.
>> Yes, snapshots are essentially copy-on-write co-dependent filesystems.
> It helps thinking about how the snapshot is implemented to understand
> it's possibilities. In essence you're creating a file for each snapshot
> (not a file system file) which contains a block change list and the
> changed blocks. When you access your snapshot, if the block being
> requested is not found in the snapshot log, the original block is
> returned. Otherwise it's returned from the snapshot log.
>
> This should then bring you to your next question - about when to use
> snapshots. And unfortunately the answer is "it depends". And I thought I
> covered that in my original answer.
You covered very little. Perhaps it all didn't get posted?

And where did you infer anything about Virtual Machines?
>> But my question still remains....Is It OK to run off a snapshot'ed
>> filesystem assuming I wanted to do that?
> Remember, as you create snapshots it requires an additional write IO per
> snapshot. One to access the snapshot list and/or access the original
> block, and one to write the new block. So as you add additional
> snapshots your write performance degrades. If you're mostly RO there's
> really no penalty. But if you have your swap defined as a snapshot
> expect a very slow system indeed.  So you have to consider the nature of
> the system before being able to answer your question.
The question was "Is It Safe?", not "How Efficient is it?".

And why would anyone snapshot swap space?!?
> If you use the snapshot to store the original state, this is indeed what
> will happen. One to store the old block (or a pointer to it) - one write
> to write the new one.
One of the claims of the benefits of LVM goes something like "well, you 
can just make a snapshot and experiment, and if it doesn't go well you 
can simply come back to your original state". Well, it's a bit more 
complex unless you can either [1] switch to the snapshot or [2] delete 
the original and switch to the snapshot later if things go bad. Reality 
is Trickier than the Cool Idea!

JIM