Why you shouldn't buy a single model/batch of HDDs.

[marbus90]

Sample Size: 37 Clients, each of them has 1x ST4000DX001-1CE168 plus 1x HDS721010DLE630 or HDS721680PLA380 or ST380815AS - plus 5x ST4000DX001-1CE168 in servers.

Code:

Hitachi HDS721010DLE630         05ReallocatedSectors    027
Hitachi HDS721010DLE630         05ReallocatedSectors    022
Hitachi HDS721010DLE630         05ReallocatedSectors    220
Hitachi HDS721680PLA380         TemperatureAlert        50C
Seagate ST380815AS              05ReallocatedSectors    01A
Hitachi HDS721010DLE630         05ReallocatedSectors    007
Seagate ST4000DX001-1CE168      C5PendingSector         008 && C6UncorrectableSector   008
Hitachi HDS721010DLE630         05ReallocatedSectors    003
Seagate ST4000DX001-1CE168      05ReallocatedSectors    1F0
Hitachi HDS721010DLE630         C5PendingSector         018
Hitachi HDS721010DLE630         05ReallocatedSectors    069
Seagate ST4000DX001-1CE168      C5PendingSector         038 && C6UncorrectableSector   038
Hitachi HDS721010DLE630         05ReallocatedSectors    0DD
Hitachi HDS721010DLE630         05ReallocatedSectors    09B
Hitachi HDS721010DLE630         05ReallocatedSectors    026
Hitachi HDS721010DLE630         05ReallocatedSectors    01B

That somewhat correlates with my experience at home as well. If we follow the Backblaze stats and reactions of the general public to them, we can see that a failure rate of ~10% of a single model leads to the assumption that the whole product line is bad.

 

[vincent orange]

This is good to know as I'm about to buy 16 2.5 inch drives for my own machine build. Would this assumption still hold true for SSD's or does this manufacturing quirk occur only with spinning drives?

 

[marbus90]

With SSDs the controllers are likely to fail - usually around the same time as well. So I'd spread that across multiple vendors as well.

 

[marbus90]

Replaced 9 of these disks with the same model from stock, another 2 dead... sigh.

 

[Arwen]

Yes, I totally agree. I got bit by multiple batch failures;

• On my first home drive, a Seagate 32MB SCSI, (and yes, sigh, that really is not a typo, it's MegaByte :-(). It suffered from the now famous Seagate sticktion, (platter or spindle lubricant acting like glue when cold).
• Later I got bit by a batch of IBM 9GB SCSI drive failures. About 20 ended up being replaced. Perhaps only 4 or so failed quickly. But the disk array vendor worked with us to replace each one from that new envrionment before they rest failed.

So, I designed my new FreeNAS, (bought early this year), with different models and bought at different times.
That's 2 x WD 4TB Reds bought locally, about a month apart. Plus, 2 x WD 4TB Red Pros bought mail order,
also about a month apart. All 4 in a RAID-Z2. Hopefully I won't experience a tripple disk failure... But, that's
what backups are for.

 

[marbus90]

I would rather have gotten HGST NAS or Seagate NAS instead of the WD Red non-pros. Especially since that's 7200rpm vs 5400rpm with the Red Pro vs. Red. The HGST one would be 7200rpm as well...

 

[Arwen]

I would rather have gotten HGST NAS or Seagate NAS instead of the WD Red non-pros. Especially since that's 7200rpm vs 5400rpm with the Red Pro vs. Red. The HGST one would be 7200rpm as well...

Speed was not an issue, as I have seen 110MBps write throughput, (1Gbps network saturation), even with mis-matched
drives. (Of course, I failed to mention earlier I have a pair of SSDs for mirrored SLOG.)

One motivation for single vendor, (and I did consider Seagate NAS drives), is things like size or feature set. For example, if I
tried to make a Zpool from 2 disks of one size and 2 disks of a slightly smaller size, how does that work?
Or the time limited error recovery features?
Same with SMART?

 

[Robert Trevellyan]

One motivation for single vendor, (and I did consider Seagate NAS drives), is things like size or feature set. For example, if I
tried to make a Zpool from 2 disks of one size and 2 disks of a slightly smaller size, how does that work?
Or the time limited error recovery features?
Same with SMART?

If drive sizes are not identical within a vdev, FreeNAS will just treat them all as being the same size as the smallest. For TLER and SMART, it will do the best it can with what you give it, so for example, TLER being missing from one won't stop it from working on another.

 

[marbus90]

Speed was not an issue, as I have seen 110MBps write throughput, (1Gbps network saturation), even with mis-matched
drives. (Of course, I failed to mention earlier I have a pair of SSDs for mirrored SLOG.

I have 4x 7200rpm disks in striped mirrors... and I can't saturate Gbe ethernet at times. If you need fast writes (that's what the SLOG is for) you want striped mirrors and fast disks. not a mix of 5400rpm consumer NAS and extremely expensive 7200rpm enterprise HDDs in a raidz. The money on the Red Pros is *wasted*.

 

[cyberjock]

If you need fast writes (that's what the SLOG is for) ...

Sorry, this is wrong. If you are using CIFS, SLOG doesn't help you at all. All CIFS writes are cached in RAM by default. Having an slog (or even forcing writes with sync=enabled) would only make writes slower.

The right tool for the right job. SLOG is for sync writes, and only for sync writes. In theory, SLOGs can bottleneck your writes. It depends on how fast your SSD is (and no, quoting some BS benchmark doesn't give you the information to argue your saturation speed from GbE is less than your SSD). Those of us with experience know that there is much more to how an SSD performs as an slog than what some benchmarks provide. ;)

 

[jgreco]

This is good to know as I'm about to buy 16 2.5 inch drives for my own machine build. Would this assumption still hold true for SSD's or does this manufacturing quirk occur only with spinning drives?

Back in 2011, when SSD was still "too" expensive, we started pairing two different SSD's with different controller technologies together on hardware RAID controllers for SSD datastores. We have had about the expected rate of SSD failures, most of them Sandforce.