Re: How smart is S.M.A.R.T.?

> Wayne,
>    I don't work in a datacenter, but I do (occasionally) do hardware selection for a variety of machines specifying anywhere from 1 to around 2,000 machines at a time.
> These are my thoughts on the topic:
> 1) Google did a fleet analysis a few years back tracking over 100,000 "consumer" drives [1] which shows that they're pretty reliable in large fleets.
>     The Google machine fleet is all "consumer" drives, because an earlier (unpublished) analysis showed insufficient real-world difference in failure rates to justify the much higher cost.
>     Enterprise drives often perform better, use less power, or have better interaction with RAID cards, but I wouldn't buy them just for reliability unless my storage design and data criticality justified the cost increment.
>     Backblaze publishes continuing data on reliability for their entire drive fleet; it's not an ideal sample set, but it is still useful to see failure rates in a real (albeit extremely cost-conscious) environment [2].
> 2) While many corporate datacenters still run things like RAID and enterprise drives, the largest machine fleets are using distributed computing solutions that deliberately pick the least expensive drives and tolerate higher failure rates.
> 3) I've been working with primarily large distributed computing systems for some years.  I generally pick drives in two different ways:
>    a) For desktops, I select medium reliability consumer drives; WD works well for those at the moment, but it changes year-to-year.
>    b) For "Enterprise" servers, where the platform stack cannot recover from failure (e.g. SQL databases, or NFS file servers) I pick the high-end "enterprise" drives, run frequent *verified* backups, and use RAID (10 if possible, 5 or 6 otherwise).
>    c) For laptops, assume the drive will fail at or near 3 years (18 months if running consumer Windows versions).  The laptop environment is just a bit harsh on spinning platters, and rewrite rate tends to be high for consumer SSD.
>    d) For really large fleets in distributed systems (OpenStack, Hadoop, Riak, Ceph, Cassandra, Mesos, etc...) where single cluster purchases are in the thousands of machines (at 12-48 drives each), I pick the cheapest large-volume purchase contract the company can get.
>       Drives will fail frequently, if they fail in less than 3 months they get returned defective; after that they almost always last about 3 years, at which point the entire machine is usually depreciated out and replaced.

>>>>>>> ID1        Read Error Rate: 152141757
>>>>>>> ID5      Reallocated Sector Count: 178 sectors
>>>>>>> ID187     Reported Uncorrectable Errors: 0 sectors
>>>>>>> ID198    Uncorrectable Sector Count: 0 sectors
>>>>>>> ID199    UDMA CRC Error Rate: 0