ZFS + RAID1/10 vs RAID6/60

[jenksdrummer]

I've generally looked at RAID6/60 as the 'correctable' RAID; that if there is a bit-rot/flip that happens in the data or in one of the parity, between the original data and the two parity calculations a corrective action can be performed to put that data back to it's original value (or the parity can be updated). I've viewed RAID10 as not being able to do that...basically it can detect that the data is a mismatch between the mirrors and it's more or less a best-guess as to which is the correct bit value.

That's probably incorrect; but that's just what I've gone with and my preference for RAID6/60 (outside of the N+2 vs N/2 and that you can lose any 2 disks vs any 1 disk per mirror pair parts) - though performance is performance...

** But, my point is with ZFS, does that make that thought moot? Without researching much more on the topic, I believe that on top of all of it, there is a checksum that is also written per block, which if it fails the checksum, then the checksum can be compared to the data either in parity or on the mirror, and the mis-match value corrected. Is this accurate?

 

[danb35]

Is this accurate?

Yes.

 

[PhiloEpisteme]

I've viewed RAID10 as not being able to do that...basically it can detect that the data is a mismatch between the mirrors and it's more or less a best-guess as to which is the correct bit value.

Would you mind clarifying this point? If the bit-flip occurs on a single disk in RAID10 wouldn't the disk's mirror pair be able to correct the error? Or are you referring to a bit flip affecting both disks in a mirror pair?

 

[danb35]

If the bit-flip occurs on a single disk in RAID10 wouldn't the disk's mirror pair be able to correct the error?

In a traditional mirror (i.e., one without checksums), there's no way to know which one is "right"; the system can only say there's an error.

 

[jgreco]

In a traditional mirror (i.e., one without checksums), there's no way to know which one is "right"; the system can only say there's an error.

Incorrect in practice. In order to detect an error, you would need to read both sides of the mirror, which doesn't happen. So the system doesn't even say there's an error.

 

[PhiloEpisteme]

Ah, I suppose I was more thinking of bad sectors causing issues rather than bit flips. Thanks.

 

[jgreco]

Ah, I suppose I was more thinking of bad sectors causing issues rather than bit flips. Thanks.

Yes. And fortunately most errors these days *do* get detected and passed up the chain as bad sectors, so that's pretty good. Not like the old days of crappy-to-nonexistent error detection on magnetic media...