4 ms·

I've watched Backblaze closely, and I love their detailed reports of drive stats. I've always been curious about their choices of manufacturers though. In my

I've watched Backblaze closely, and I love their detailed reports of drive stats. I've always been curious about their choices of manufacturers though. In my past I found that Hitachi/HGST drives were far more reliable than others. We had a couple people doing hardware and software support on ~120 machines, and hard drive replacements were expensive (in man-hours) for us, so whenever possible we got Hitachi drives. The Backblaze numbers seem to support this as well, far fewer failures of HGST drives vs Seagate, but they seem to buy mostly Seagate. I guess when you have tens of thousands of drives, you're always going to need to be replacing them. So maybe a 10x reduction in failures isn't worth it? Or maybe they just couldn't get HGST drives in quantity?

cyberpunk0 9y ago
I've never had a Seagate drive last any reasonable amount of time
dspillett 9y ago
They don't buy direct from manufacturers so it might be partly down to availability numbers from their local distributors. Also you want to buy from a wide range - if most of the drives come from one place and they have a manufacturing disaster that causes everything made in one month to fail a few months into use then that might take you too close to redundancy requirements for comfort. I would expect the range picked will be based on individual design (platter design, platter layout, controller type, ...) rather than per manufacturer - maybe HGST has fewer designs available at any one time and that tips the balance a bit.
atYevP 9y ago
Yev from Backblaze here -> We have relationships with manufacturers but don't directly buy from them (yet - fingers-crossed). It really is a combination of price/reliability. At our scale we can afford slightly higher failure rates if it means less expensive drives, so our #1 data point for purchases is the cost per gigabyte. After that we do look at failure rates, but unless it's something wildly out of "normal" we tend to live with the occasional failures. Thanks to Vaults the back-end is designed in such a way that if a drive fails we don't have to run and replace it - so individual failures tend not to stress us out too much anymore!
- dzdt 9y ago
  If a 10's of petabytes isn't enough scale to talk to manufacturers directly, I wonder what is!
  - MattSteelblade 9y ago
    After the order mentioned in the article, I wonder if it will be. Since they mention a mix of 10 an 12 TB hard drives, we can assume the order was for around 9000 drives. At a cost of $0.03/GB (not taking into account shipping or taxes), you're looking at a 3 million dollar order.
volkl48 9y ago
It is worth considering that the use-case of Backblaze is different from the average end-user. No power on/off cycles, no physical impacts to worry about (even most desktops occasionally get bumped into), but non-stop operation and presumably near continuous read/write operations. You'll likely get different failure patterns as a result.
- 15155 9y ago
  > but non-stop operation and presumably near continuous read/write operations. I thought it would be more like: continuous writes until full, and then likely a handful of reads for the rest of that drive's life.
  - metalliqaz 9y ago
    Not if they're in service of B2.
ensignavenger 9y ago
The last I heard, I understood that Backblaze doesn't go around replacing single drives that fail in a pod, rather, their software marks that drive bad and only once a certain percentage of the capacity of the entire pod is gone, do they yank it our and rebuild the pod. That probably helps a lot versus an environment where had to have technicians replacing single drive failures randomly throughout your data center.
- notatoad 9y ago
  they operate at datacenter scale: https://xkcd.com/1737/ https://xkcd.com/1737/

cyberpunk0 9y ago

I've never had a Seagate drive last any reasonable amount of time

dspillett 9y ago

They don't buy direct from manufacturers so it might be partly down to availability numbers from their local distributors. Also you want to buy from a wide range - if most of the drives come from one place and they have a manufacturing disaster that causes everything made in one month to fail a few months into use then that might take you too close to redundancy requirements for comfort. I would expect the range picked will be based on individual design (platter design, platter layout, controller type, ...) rather than per manufacturer - maybe HGST has fewer designs available at any one time and that tips the balance a bit.

atYevP 9y ago

Yev from Backblaze here -> We have relationships with manufacturers but don't directly buy from them (yet - fingers-crossed). It really is a combination of price/reliability. At our scale we can afford slightly higher failure rates if it means less expensive drives, so our #1 data point for purchases is the cost per gigabyte. After that we do look at failure rates, but unless it's something wildly out of "normal" we tend to live with the occasional failures. Thanks to Vaults the back-end is designed in such a way that if a drive fails we don't have to run and replace it - so individual failures tend not to stress us out too much anymore!

dzdt 9y ago
If a 10's of petabytes isn't enough scale to talk to manufacturers directly, I wonder what is!
- MattSteelblade 9y ago
  After the order mentioned in the article, I wonder if it will be. Since they mention a mix of 10 an 12 TB hard drives, we can assume the order was for around 9000 drives. At a cost of $0.03/GB (not taking into account shipping or taxes), you're looking at a 3 million dollar order.

volkl48 9y ago

It is worth considering that the use-case of Backblaze is different from the average end-user. No power on/off cycles, no physical impacts to worry about (even most desktops occasionally get bumped into), but non-stop operation and presumably near continuous read/write operations. You'll likely get different failure patterns as a result.

15155 9y ago
> but non-stop operation and presumably near continuous read/write operations. I thought it would be more like: continuous writes until full, and then likely a handful of reads for the rest of that drive's life.
- metalliqaz 9y ago
  Not if they're in service of B2.

ensignavenger 9y ago

The last I heard, I understood that Backblaze doesn't go around replacing single drives that fail in a pod, rather, their software marks that drive bad and only once a certain percentage of the capacity of the entire pod is gone, do they yank it our and rebuild the pod. That probably helps a lot versus an environment where had to have technicians replacing single drive failures randomly throughout your data center.

notatoad 9y ago
they operate at datacenter scale: https://xkcd.com/1737/ https://xkcd.com/1737/