[ale] disk drive diagnostics nirvana - NOT - I have questions

[Phil Turmel]

On 10/23/2012 10:49 AM, Derek Atkins wrote:
> Phil Turmel <philip at turmel.org> writes:
> 
>> My critical servers all use linux software raid in various combinations,
>> and all of the raid arrays are scrubbed weekly.  By scrubbed, I mean a
>> cron job instructs the kernel to read every sector on every member
>> device in the background, compute parity as appropriate, and report any
>> inconsistencies.  Any read errors trigger the corresponding recovery and
>> rewrite functions that would normally occur if an application
>> encountered the sectors.  Any unsuccessful write kicks that device out
>> of the array as usual.
> 
> I thought that the raid scrub did a read/write of every sector, not just
> a read?

No, the "check" operation reads everything and only writes over read
errors.  It will report but not fix parity errors and mirror mismatches.

The "repair" operation reads all raid4/5/6 data blocks and computes and
writes all parity blocks.  For mirrored data, it reads all of mirror #1
and writes to all of the other mirrors.

This operation is only recommended if a "check" operation turns up
mismatches.

>> I have been doing this for about ten years now, with about seven or
>> eight drive failures in that time.  Never lost any data, though I've
>> been nervous a few times when waiting for a replacement disk for a raid5
>> array.  Everything is now raid6 or triple-mirrored, so I sleep well.
> 
> I use RAID-10 personally.

I included raid10,far3 in my "triple-mirrored" statement.  I use it for
highest random I/O performance on my root partitions.  Bulk media and
general purpose volumes are on raid6.

> I look for a new drive as soon as I get the first errors (assuming I
> don't happen to have a cold spare on hand).

Due to the nature of the media, I don't count transient read errors that
disappear when the raid rewrites the bad spot.  Relocations are my
critical indicator.

Phil