What are you doing for home NAS?

[Michael Bilow]

On 2013-12-30 at 10:18 -0500, Ken D'Ambrosio wrote:

> On 2013-12-30 09:41, John Abreau wrote:
>> After trying FreeNAS, I'd no longer consider the consumer-level drives
>> such the MyBook Live as serious options.
>
> I think this stance is a little overly cautious; there is  data showing
> that consumer drives don't fail at rates significantly different than
> "server-grade" drives -- e.g.,
> http://blog.backblaze.com/2013/12/04/enterprise-drive-reliability/
> (though I also remember studies done on significantly larger datasets a
> couple years ago, but they aren't leaping at me from Google).  What I
> *have* found to be troublesome is that some RAID solutions don't handle
> drives that spin down very well.

Server-class drives are very different in at least two major ways.

First, a basic operational failure, such as a bad sector read, will 
cause a consumer-class drive to retry pretty much forever, while 
such an error will be reported relatively rapidly to the controller 
in a server-class drive. This allows the RAID manager (whether 
hardware or software) to handle the error appropriately, usually by 
computing what the sector should contain and writing it, thereby 
causing a reallocation of the failed sector from a reserve of 
spares. Software RAID on Linux, for example, provides for a periodic 
"scrubbing" operation that test reads every sector; Debian codes 
this as a monthly "checkarray" cron job.

Second, server-class drives are tested for 24x7 operation and rated 
accordingly with longer warranties that take this into account. 
Seagate consumer drives are specified with MTBF based on operation 
about 9 hours per day Monday through Friday, a typical work week. 
Western Digital specifies error performance an order of magnitude 
better in their top server-class drives, 1 in 10^16, compared to 
their consumer-class drivers, 1 in 10^15 -- which may not seem like 
much, but actually does have practical effect when dealing with 
multi-terabyte arrays. Not everything that goes wrong in a RAID 
system takes the spindle off-line, which is important to remember.

I obviously cannot dispute the Backblaze data, but their application 
is relatively quiescent and therefore unusual. Most server-class 
drives are operated under heavy demand in something like an e-mail 
server or a database server, and even the Backblaze article notes 
that this could significantly affect their results. Increased 
activity caused increased heat, the great enemy of the drive. To a 
large extent, though, you are paying for the longer warranty.

> For this reason, I tend to either go
> with "server-grade" drives, or really do my homework, and find drives
> that work with the solution

Lately there has been a trend among manufacturers to introduce 
price-differentiated grades of server-class drives, notably the new 
Western Digital "Red" line costing less than their "Se" enterprise 
value-priced line and their "Re" enterprise full-priced line. At 4TB 
a "Red" (WD40EFRX) is about $190, a "Se" (WD4000F9YZ) is about $280, 
and a "Re" (WD4000FYYZ) is about $320. Much of the increased 
reliability of the "Red" comes from simply spinning it at 5400RPM 
instead of 7200RPM, and much of the decreased cost comes from 
providing a 3-year warranty instead of a 5-year warranty. By 
comparison, a consumer-class "Green" (WD40EZRX) is about $170, 
spinning at 5400RPM and the warranty cut to 2 years.

> (e.g., 3Ware has -- or, at least, had -- an
> approved hardware list that I find useful).  But I think that, with a
> suitable amount of caution, there's money to be saved here without loss
> of functionality or increased risk of data loss.

I've been too badly burned on 3Ware RAID controllers to use them 
ever again. Regardless, in the vast majority of cases these days, I 
recommend using software RAID rather than hardware RAID on Linux 
servers unless there are special requirements, such as for hot-swap 
capability or for absolute maximum performance.

> P.S.  One thing I should add here, just from a hoo-boy-did-I-stub-my-toe
> perspective: as a rule, I usually have my arrays use just a leeeeetle
> bit less than the whole disk.  I had a large RAID-5 array once, and one
> of the drives failed.  I got it RMA'd *with the same model number* from
> the manufacturer... and it was one sector smaller.  THAT was annoying.

Although not generally known and hardly recommended practice, Linux 
software RAID can handle that case.

-- Mike