Re: What exactly is BTRFS Raid 10?

[Phillip Susi <phill@thesusis.net>]

"George Shammas" <btrfs@shamm.as> writes:

>> Btrfs raid10 requires an even number of drives with a minimum of 4.
>
> Is this true? I just experimented with 3 drives and `btrfs device
> usage` ends up showing 'RAID10/2' But the data is equally spread
> across all three drives. Even though the stripe is 2, the data is
> still being placed evenly across three drives. See [1] for an actual
> example I just created.

mkfs.btrfs won't let me create a raid10 with only 3 drives: it says you
need at least 4.  You can do that with mdraid but btrfs doesn't seem to
let you and I thought it was because of:

>> It's pretty much raid 1+0.

> Again, is it? raid 1+0 would imply that two drives are mirrored, and
> hence identical, and the blocks are striped over mirrored sets. This
> would also force devices being in multiple of two, which is not the
> case.

Don't forget that it is done at the chunk level rather than the drive
level.  I just tried 5 drives though and that does seem to work,
however, btrfs filesystem usage shows that the chunk is only allocated
on disks 2,3,4,5 and disk 1 is unused.  Over time you are going to end
up with the average usage about the same but no single chunk is striped
across all 5 drives.

> If I am right, and I don't know that I am, that would make raid1 and
> raid10 have the very similar data loss scenarios that are not
> completely obvious.

Yes, in general raid1 and raid10 have similar redundancy: they can only
handle the loss of a single drive.  Any more than that and you are
relying on peculiarities of the layout and which particular drives
failed, and that's never a good idea.

> And the question remains if that is the case, is there ever a reason to choose raid1 over raid, and vice versa. 

The reason to choose raid10 over raid1 is for the better sequential
performance you can get from the striping.  I ran an mdraid raid10 on 3
disks for years for the good throughput and redundancy qualities.