[mythtv-users] Slightly OT - How many People have Video libraries over 8TB?

[Raymond Wagner]

On 7/7/2012 13:34, digid myth wrote:
> I would never build a raid with 2 TB drives. Takes too long to rebuild
> the raid and 60% of the time you will have a 2nd drive fail during the
> rebuild.

Huh? There was a whole lot of FUD several years ago about UBE rates 
making it probable you experience a second bit failure by the time you 
finish a recovery. So a brief lesson in statistics and hard drive 
design, what is going on here?

A UBE (unrecoverable bit error) or URE (unrecoverable read error) is 
simply when a hard drive fails to decode a single bit. You don't 
actually read the data directly. That requires some really expensive 
gear, and is too slow. You read most of a whole bunch of redundant 
information, and decode your data back out of it using CRC recovery 
codes. Many hard drives are rated for a UBE rate or less than one in 
10^14 bits. That means within 100Tbits, or 12.5Tbytes (or 11.37TB), you 
are less than 50% likely to experience an event in which you don't read 
enough information off the platter to decode your data from that sector.

So what happens when you experience a UBE on a RAID array? These things 
are expected due to the design of hard drives. It doesn't indicate a 
failed drive. It merely indicates a single failed read. The RAID array 
detects the fault, decodes the data from the parity block, writes it 
back to the drive the fault occurred on, and life goes on. Some 
companies sell "RAID Edition" drives, and the only difference between 
these and the standard consumer fare is that the standard drive will go 
to Herculean lengths trying to re-read that sector to recover the data, 
while the RAID Edition drive will fault out early, allowing the array to 
recover the data on its own, rather than stall operations for an 
extended period.

What if it's a RAID0 array, or critical RAID5 array? As mentioned, this 
isn't a drive failure, it's only a sector fault. The RAID controller 
isn't going to expel the card and render itself incomplete. That would 
be stupid. It's going to warn the administrator, and may wait for 
confirmation, before going ahead and completing the recovery of the rest 
of the array. Before with a single drive, you just lost the sector, 512B 
or 4KB, while now the worst case is you lose that whole stripe, maybe 
1MB or so. Of course you could always run RAID6, where after the loss of 
that first drive, you still have a second parity set, and your chances 
of two UBEs occurring within the same stripe on two drives is 
infinitesimally small.

There is no RAIDocalypse. We are not all going to lose our data. We will 
just get to the point where an array with no redundancy left guarantees 
some loss of data.

> In the corporate world no drives larger than 500G or so are typicly put
> in a raid group for just this reason.

In the corporate world, no drives larger than 500G or so are typically 
put in a RAID group, for the sole reason that they simply don't exist. 
They just don't make 10K and 15K SAS drives larger than 600GB.

> Everytime I have lost something on a raid it was during a raid rebuild
> of a failed disk.

In all likelihood, you bought a bunch of hard drives at the same time, 
from the same vendor, from the same manufacturer, such that they were 
all built in the same production batch. You then proceeded to operate 
them in the same temperature conditions, under the same use profile. It 
is not at all unreasonable that you would experience two of them failing 
at nearly the same time. That's why you're supposed to buy drives from 
separate lots, and periodically cycle them, so you don't get a bunch 
ready to fail simultaneously.