MacInTouch Amazon link...

storage systems

Channels
Products, Questions
A perfectly reasonable question, but to pick nits, it's probably worth noting that Thunderbolt 3 is essentially a PCIe 3 x4 connection, and this isn't sufficient to support "the fastest SSD drives available".
It's fast enough to support virtually any existing single M.2 NVME SSD blade at full capacity, because these devices are also based on a PCIe 3 x4 connection. But anything faster - a high performance PCIe datacenter SSD, or a striped (RAID 0) array of multiple M.2 blades on a card - will be performance-limited by Thunderbolt. These types of things would need to be plugged directly into a sufficiently fast/wide PCIe slot to hit full speed (and so aren't really worth the cost if your plan is to put them in a Thunderbolt 3 enclosure..)
Thank you very much, John W (and to Ric, who had tried to steer me to the same facts earlier). You've answered my question fully, by pointing out there's no need for me do the search for such an enclosure or drive in the first place.

When the time comes, I'll simply convert my existing Thunderbolt 3-enabled OWC ThunderBlade to JBOD, which will boot Catalina.
 



The takeaway (as it has been for several years) is that all things being equal, folk who want to host stuff at home should be buying HGST or Toshiba drives. SOHO systems are unlikely to feature as many parity checks / backups as Backblaze and, as such, cannot afford the "business as usual" 1-3% failure rate with Seagate products.

Also notable is the complete absence of WDC products in that data set, even though WDC now owns HGST. If you click around their web site, Seagate's 12TB products have apparently been particularly troublesome, so I'd avoid using those drives. Backblaze is apparently working with Seagate to replace their extant cohort.

So even with all the redundancies that Backblaze has, a 3+% failure rate is not acceptable. (Labor may have something to do with it, as HGST 12TB drives/brands literally require 8x fewer technician visits for drive replacement, never mind the other associated tasks like documenting, shipping, etc. All that maintenance does add up, no matter how cheap the initial drive.)
 


I've found it hard to find the exact models in the Backblaze reports. I'm guessing that many models are superseded by the time of the reports.

Just checking Amazon, it looks like Western Digital is now branding Ultrastar drives as WD. I got no hit for HUH721212ALN604, instead getting HUH721212ALE604, and by vendors I'm unfamiliar with.
 


The takeaway (as it has been for several years) is that all things being equal, folk who want to host stuff at home should be buying HGST or Toshiba drives. SOHO systems are unlikely to feature as many parity checks / backups as Backblaze and, as such, cannot afford the "business as usual" 1-3% failure rate with Seagate products.

Also notable is the complete absence of WDC products in that data set, even though WDC now owns HGST. If you click around their web site, Seagate's 12TB products have apparently been particularly troublesome, so I'd avoid using those drives. Backblaze is apparently working with Seagate to replace their extant cohort.

So even with all the redundancies that Backblaze has, a 3+% failure rate is not acceptable. (Labor may have something to do with it, as HGST 12TB drives/brands literally require 8x fewer technician visits for drive replacement, never mind the other associated tasks like documenting, shipping, etc. All that maintenance does add up, no matter how cheap the initial drive.)
I agree with these conclusions, and when I had a choice, when I still worked with a medium-sized data center, I would try to purchase HGST drives for many-drives-per-shelf, many-shelves-per-rack applications.

But it should be pointed out that Backblaze's results are not strictly applicable to desktop, or small-enclosure (1 - 6 drives) applications. They utilize drives in high-density, custom designed shelves ("Pods") with up to 45 drives. And though those enclosures are designed to minimize vibrations, they're stacked in racks with lots of other enclosures, and the drives are desktop drives — that's the key to Backblaze's low-cost approach — and not designed for the higher acoustic noise (vibration), constantly-powered-on environment of a data center.

TL;DR: The same drives might have different relative failure rates in desktop systems/enclosures (that are kept clean) than in the Backblaze stats.
 



I've found it hard to find the exact models in the Backblaze reports. I'm guessing that many models are superseded by the time of the reports. Just checking Amazon, it looks like Western Digital is now branding Ultrastar drives as WD. I got no hit for HUH721212ALN604, instead getting HUH721212ALE604, and by vendors I'm unfamiliar with.
The "E" toward the end of the second part number represents a 512e ("Advanced Format," or "AF") 6 Gbit/s SATA interface that appears to the host system to have 512 byte sectors despite having larger (4096 byte) physical sectors; I believe the "N" is for drives with an actual 512 byte sector structure, though it might represent 4K "native" AF drives.
 


I've found it hard to find the exact models in the Backblaze reports. I'm guessing that many models are superseded by the time of the reports.

Just checking Amazon, it looks like Western Digital is now branding Ultrastar drives as WD. I got no hit for HUH721212ALN604, instead getting HUH721212ALE604, and by vendors I'm unfamiliar with.
The decoder ring: Ultrastar DC HC520 data sheet (PDF). See the end of the second page: “How to Read the Ultrastar Model Number.” The difference here seems to be “N6 = 4Kn SATA 6GB/s, E6 = 512e SATA 6Gb/s.” Web-searching 4Kn vs. 512e is left as an exercise for the reader. (Or see Joe Gurman's reply that was posted as I typed.)
 


But it should be pointed out that Backblaze's results are not strictly applicable to desktop, or small-enclosure (1 - 6 drives) applications. They utilize drives in high-density, custom designed shelves ("Pods") with up to 45 drives. And though those enclosures are designed to minimize vibrations, they're stacked in racks with lots of other enclosures, and the drives are desktop drives — that's the key to Backblaze's low-cost approach — and not designed for the higher acoustic noise (vibration), constantly-powered-on environment of a data center.
While I agree in principle, I'll also note that many consumer-grade desktop enclosures (especially external drive units from OEMs like Seagate, WDC, etc.) do a fabulous job of roasting enclosed hard drives in their own juices. After a multi-hour backup (which 12TB capacity basically ensures), I'd wager the drives in there are significantly hotter than the ones in the storage pods at Backblaze.

My spinners tend to fail in the backup enclosures, not the main server, even though they get used a lot less hours per year than the server drives. I presume that has to do with turning them off and moving the enclosure off-site.

Inside the server, I keep my spinners below 31*C year-round; they rest on silicone grommets inside a large metal cage with plenty of air flow. The solid state stuff, like the Optane p4801x or the LSI 2116 HBA chip, does tend to run hotter.
 


While I agree in principle, I'll also note that many consumer-grade desktop enclosures (especially external drive units from OEMs like Seagate, WDC, etc.) do a fabulous job of roasting enclosed hard drives in their own juices. After a multi-hour backup (which 12TB capacity basically ensures), I'd wager the drives in there are significantly hotter than the ones in the storage pods at Backblaze.

My spinners tend to fail in the backup enclosures, not the main server, even though they get used a lot less hours per year than the server drives. I presume that has to do with turning them off and moving the enclosure off-site.

Inside the server, I keep my spinners below 31*C year-round; they rest on silicone grommets inside a large metal cage with plenty of air flow. The solid state stuff, like the Optane p4801x or the LSI 2116 HBA chip, does tend to run hotter.
The thermal issues are ones Backblaze's Pods are unlikely to have, given the numbers of high-capacity fans in each one, and the fact that they're in a server room environment (though I don't know what temperature is maintained in them).
 


In the era of the 1.5TB Seagate drives, my favorite storage system accessory was a 140 mm desk fan used to cool drive enclosures and bare drives in 'toasters'. Forgetting the fan meant that any backup taking more than about half an hour would simply stop.

Even now I routinely use a fan for any first full Time Machine or clone backup for drives in a 'toaster'. The latter is especially useful for beta test systems which may be backed up/tested using older drives previously removed from service but still showing no SMART errors.
 


'Way back' in 2014, Backblaze ran an analysis on their data set to compare the longevity of drives to the temperature they were operating in. The first-order conclusion was that it didn't, at least for most drives. However, the temperature range was only 17-31*C, which suggests a heavily cooled data center with the colder drives being at the front of the storage pod and the hotter drives being closer to the rear.

Here is a link to Storage Pod 6.0 with good pictures of the 60-drive 4U assembly. Note that the blowers pull air from the front of the unit towards the back and expel it once it's been pulled over the motherboard. There are no fans in the front of the unit and all air has to pass multiple rows of vertically-oriented hard drives. What is notable about this design is its simplicity and how it ensures that the maximum surface area of every drive is exposed to an even air flow.

There is a huge backplane at the bottom of that unit, with each drive being pulled by gravity into its SATA connector. To access a drive, the pod is pulled out of the rack (ball bearing slides help!), the lid is opened, and the offending drive is pulled and replaced. To my limited knowledge, no SOHO server (8-16 drive) cases offer a similar, simple design. Most SOHO server cases are designed around "hot swap" drive holders that allow the hard drives to be pulled and replaced quickly from the front of the unit - even though few of them ever fail.

Hot-swap designs typically feature much more constricted air spaces between drives to accommodate the mechanisms, guides, and so on. That in turn leads to greater static pressure drop, necessitating stronger (and usually louder) fans to get good air flow. I have yet to find a 8-12-drive SOHO server with the simplicity of the backblaze pod. LIAN LI models like the Q26 come close, but the hard drive holders in there are not quite ideal for good air flow over the drives, and the case doesn't allow for a Flex-ATX motherboard.

I have been trying to keep my drives happy at below 31*C, which is the upper end of the Backblaze temperature range. Crummy air flow designs in other cases than the one I use now have driven my drives to exceed 45*C during multi-hour scrubs. Now imagine what the drive temperature would be inside a a typical external plastic case without forced cooling.

My hard drive toaster from OWC may have a known-defective design (it doesn't allow formatting via eSATA), but USB 3 is good enough, and each drive only sees it once – it's used to zero obsolete spinners before they are donated to a local medical center for reuse in research.
 



I recently purchased a 2TB Seagate drive to replace an older 1TB drive in my Mac Pro. I used SuperDuper to clone the old drive to the new drive and was puzzled why it took so long to copy the data. Only after learning about SMR drives last week did I check model numbers. Sure enough, it was an SMR model and and probably explains why the cloning took so long.
 


Shingled Magnetic Recording (SMR) drives are interesting and the underlying technology allows more data to be written into less space on a platter. That in turn allows their manufacturers either offer higher capacity drives and / or to save cost by shipping drives with fewer platters, read/write heads, smaller motors, and so on vs. Conventional Magnetic Recording (CMR) drives.

Because SMR drives overlap the write tracks (hence the term "shingled"), they have to write data in large continuous chunks rather than allowing random writes all over. Because not all files fit a zone perfectly, the data in a SMR drive is first written to a conventional CMR sector and once that cache fills up, the CMR sector is transferred in a large write to one of the SMR zones, filling it.

The problem with SMR drives is twofold. Writes sooner or later fill the CMR cache, whose transfer to a SMR zone then effectively blocks the drive. Desktop OSes can handle such a slowdown to varying degrees, but NAS and RAID controllers are likely to flag an unresponsive drive as broken.

To me, SMR drives are 100% incompatible with NAS / RAID applications, because any array that is built with a SMR drive is effectively blocked while a CMR-SMR transfer is in progress. The busier the array, the more this issue is triggered.

The issue is even worse if a drive fails and the array has to be rebuilt. In the old days, arrays could have explicit parity-data drives (i.e. 4 drives for data, one for parity). Theoretically, one could mix and match SMR and CMR drives to limit the impact of lots of parity writes (SMR for data, CMR for parity).

However, most arrays these days use distributed parity, which distributes the parity data across all drives (see the RAID5 and RAID6 descriptions at Wikipedia). That in turn triggers lots of writes whenever data is written, and hence the probability of an array stumbling as a CMR-SMR transfer happens increases.

This is issue is magnified further should a drive fail and the replacement has to be rebuilt from parity data. The array will become very unresponsive, the rebuild will take forever, and more drives in the array may be stressed into failure, causing the whole array to fail.

This is why it is so disappointing that multiple manufacturers have quietly changed the underlying technology in their drives from CMR to SMR. Unlike the "archive" drives of yesteryear that advertised the use of SMR technology, these drives are sold as if they were CMR drives (i.e. no explicit SMR specification in the data sheet). Hopefully an outcry from buyers will put an end to this nonsense.
 


Unlike the "archive" drives of yesteryear that advertised the use of SMR technology, these drives are sold as if they were CMR drives (i.e. no explicit SMR specification in the data sheet). Hopefully an outcry from buyers will put an end to this nonsense.
Western Digital, at least, heard the outcry and replied in a blog [my cynical summary]: They don't care, but acknowledge that they are doing it and suggest you buy their more expensive CMR models if you care.

On WD Red NAS Drives - Western Digital Corporate Blog

Also note that in the case of the low-end non-"Pro" WD Red drives, the model number may be different for SMR (EFAX) than CMR (EFRX). Web search shows signs that EFRX may be in the process of getting discontinued.
 


It should be pointed out there are at least three different versions of SMR: device managed, host aware, and host managed. WD Red NAS drives (and others) are device managed.

Device managed SMR drives are 'drop in replacement' for standard CMR drives. That is, no change is needed in the file system and/or OS. As Constantin points out, one can run into issues especially when writing large amounts of data (SMR will have to manage the data once the CMR region and buffer are exhausted). Reading from SMR is never a problem. Well, by necessity, the data tracks are narrower. so seek times are slightly affected, but that's usually a small effect.

Host aware and host managed SMR drives are not drop in replacements. They require changes to the file system and/or OS. Hard drive manufacturers make most of their money not on consumer drives but on the big data centers (think Facebook, Google, Dropbox, etc.) Here, most of the data is static (also called 'cold') and so to reduce TCO, it makes sense for them to optimize around SMR, since it provides 20% more storage (in data centers, it's all about the number of GB/$).

I never thought device managed SMR was a great solution for anything – it's very difficult to write the firmware for all use cases. I can see why the manufacturers are going that route, since it streamlines manufacturing processes: they can use the work they need for the data centers (where the real revenue is) and then reuse it for consumer products (what was historically done). As consumer drives are replaced by solid state drives, it makes less sense to commit resources to CMR heads for the consumer market (in SMR, write heads are wider while the read heads are narrower compared to CMR, and this also affects the recording medium requirements). In other words, as the data centers move towards host managed SMR, the manufactures will be less and less likely to make any CMR heads/medium since it makes little business sense.

Of course, heat assisted magnetic recording (HAMR) has been promised for awhile (and it's typically CMR-based). If it comes out, then SMR will lose its interest, since HAMR should be able to achieve storage densities at least 4x larger, making SMR's 20% improvement irrelevant.
 




There is no doubt that SMR drives are here to stay and that the manufacturer oligopoly will use the performance delta vs. CMR as a means of extracting more money from users who want to run a NAS or a RAID. This is nothing new, the industry has been trying this for years.

Just look at the completely nutty price delta between internal and external hard drives. An external drive is basically an internal drive with an enclosure, a bridgeboard (SATA to USB, eSATA, FireWire, Thunderbolt, etc.) and a power supply (if a 3.5" drive; 2.5" drives are usually bus-powered). So there is a whole bunch of stuff in the package. Only marketing / profit maximization hence can explain why an external drive should cost significantly less than a bare internal drive!

Yet, ever since the floods in Thailand (2011) constraining hard drive supplies, external enclosures have been consistently cheaper than bare drive mechanisms. Sometimes, the drives inside an external enclosure were "slower" (5400 RPM vs. 7200 RPM), but many users can't notice that difference and all things being equal the lower-speed drives should run cooler and consume less power. The firmware running the drive may also have some limitations built into it, but that's the kind of stuff clever folk could write their driver code around.

Certain models of external enclosures could even hold premium helium-filled drives, leading to speculative buying of certain types of external enclosures and "shucking" the internal drives from them. Depending on how lucky you felt and / or the warranty tier, you'd keep the external enclosure in case the drive fails (WD at least requires the appearance that the drive failed in the enclosure, so you have to send the drive back in the enclosure it came with).

The manufacturers keep the stuff inside the enclosures vague for a reason - the more vague, the greater their choice re: what to stuff in there without falling afoul of consumer protection laws. I have found Reddit /datahoarder, servethehome, and like forums to be the best sources for the current contents of external drives. Sometimes, it's an element of luck, such as when WD was swapping back and forth between "Red" and "white" helium drives inside consumer-grade WD elements / easystore enclosures.

For now, I'd buy and qualify any spares you need now while you still have a good chance of obtaining a CMR drive, especially if you want to run a NAS or a RAID array. I have had very good experiences with helium-filled drives, in terms of heat, power-consumption, and reliability. I buy used 7,200 RPM HGST mechanisms for my server and use shucked 5,400 RPM drives in my backup arrays. I only stock 7,200 RPM He10 spares, so I can use them in either application.

PS: I'm amazed at all the tricks that go into making a SMR drive work. It's a fascinating application of many SSD technologies to a old platform (Trim, garbage collection, etc.), which, given the latency of drives with spinning platters, is having some really unpleasant side effects.

If only manufacturers were open about the use of SMR, consumers could make a more informed decision.
 


Of course, heat assisted magnetic recording (HAMR) has been promised for awhile (and it's typically CMR-based). If it comes out, then SMR will lose its interest, since HAMR should be able to achieve storage densities at least 4x larger, making SMR's 20% improvement irrelevant.
Indeed, there is also two-dimensional magnetic recording (TDMR) and microwave assisted magnetic recording (MAMR) in addition to HAMR, and it will be very interesting indeed to see which approach manages to break through in volume into our end of the consumer-grade universe first... though HAMR and MAMR allegedly are delayed until 2022.
Blocks & Files said:
I wouldn't count SMR out though. Allegedly, SMR can be used in conjunction with HAMR to goose density another 20%.
AnandTech said:
If it's all about the $/TB then SMR will continue to be used in conjunction with MAMR, HAMR, or whatever type of writing technology being used.
 


Two dimensional magnetic recording (TDMR) simply uses multiple readers when trying to read data from a track. The magnetic read sensor not only detects the data track you are trying to read, but also neighboring tracks. The idea of TDMR is to use the other readers' response to 'cancel out' the interference of the neighboring tracks. While the idea is simple, in practice it has only shown to improve storage density by less than 6% in actual drives. TDMR can be used with CMR and SMR along with different physical recording technologies: HAMR, MAMR and ePMR. Similarly, SMR can be used with HAMR, MAMR and ePRM.

While HAMR has been shown to work at the drive level by multiple companies, the typical problem is write head reliability. Seagate has claimed (multiple times) they have overcome this problem, but they still haven't shipped HAMR drives other than 'evaluation' or samples. It should be noted that each HAMR write head requires its own laser. While the cost of each laser isn't much, incorporating the laser into the write head is not trivial and the supply has always been in question.

Unlike HAMR, MAMR has never been shown to work at the drive level. Back in Oct 2017, Western Digital claimed to have a prototype MAMR drive, but this turned out to be using ePMR and not MAMR. Western Digital has stated publicly last year that "MAMR will not achieve HAMR densities ... HAMR will be required". I find this somewhat misleading since MAMR hasn't achieved any storage density beyond CMR. There are technical details which strongly suggest a MAMR drive will never exist.

ePMR was discovered when trying to implement MAMR. While Western Digital has never officially announced what ePMR is, I've been told by multiple people that it simply runs an electrical current through a conventional CMR write pole. The field generated by this electrical current helps switch the write field faster resulting is improved "jitter" which allows for more bits down a given track length. While it is interesting, running a current directly into a write pole tip should cause local heating and one would naively assume reduce reliability. Western Digital claims they will be shipping ePMR in second half of 2020 (18TB with CMR, 20TB with SMR). Of course, they also claimed they would be shipping MAMR in 2019. Their current roadmap shows both MAMR and HAMR simultaneously which is very different from what they said just last year (MAMR first and perhaps never HAMR).
 




I'll bet an ice-cream sandwich that the high-volume users out there like Backblaze (BB) are using host-aware SMR, since that in turn should allow them to work around the worst issues associated with SMR. For example, if your server is aware of what drives are currently offline doing garbage collection, flushing the cache, whatever, then you can potentially work around that issue by writing the data to other disks / pods / and so on. For example, the average storage pod at BB now uses 60 disks, so the likelihood of all drives being offline is rather low.

Individual pods can also signal to whatever is controlling the allocation of data in the data center that a pod is "too busy" and hence the data can be stored somewhere else. Currently, BB breaks uploaded data into 17 shards (plus 3 for parity), and distributes them all over their data center. Thus, there is likely some freedom re where the data can be sent at any one time. Pods experiencing SMR-related throughput issues are easy enough to identify.

Lastly, all of the above may not matter too much as long as the host is aware and you simply count on the large number of drives to give you very high throughput, even if individual writes may take longer than expected - simply write the server-side software to allow for very long latencies at times, something a host-aware drive should be able to signal to the host.

Hence, I do not see a place for SMR drives (host-aware or not) in smaller arrays unless the owner is OK with very low throughputs at times. Yes, for additional $$$ you might be able to work around the worst aspects of SMR by using a hybrid approach (first write to a SSD-based cache, then write to the SMR-based drives), but the risk is always that once you overwhelm the primary cache (via heavy usage, for example) that array performance will crater.

Use case will always be an important factor, but I prefer using CMR-based arrays due to their more predictable performance.
 


I'll bet a ice-cream sandwich that the high-volume users out there like Backblaze (BB) are using host-aware SMR, since that in turn should allow them to work around the worst issues associated with SMR. For example, if your server is aware of what drives are currently offline doing garbage collection, flushing the cache, whatever, then you can potentially work around that issue by writing the data to other disks / pods / and so on.
I'll take that bet.
Ariel Ellis (Director of Supply Chain at Backblaze) said:
How Backblaze Buys Hard Drives
SMR would give us a 10-15% capacity-to-dollar boost, but it also requires host-level management of sequential data writing. Additionally, the new archive type of drives require a flash-based caching layer. Both of these requirements would mean significant increases in engineering resources to support and thereby even more investment. So all-in-all, SMR isn’t cost-effective in our system.
 


'Way back' in 2014, Backblaze ran an analysis on their data set to compare the longevity of drives to the temperature they were operating in. The first-order conclusion was that it didn't, at least for most drives. However, the temperature range was only 17-31*C, which suggests a heavily cooled data center with the colder drives being at the front of the storage pod and the hotter drives being closer to the rear. Here is a link to Storage Pod 6.0 with good pictures of the 60-drive 4U assembly. Note that the blowers pull air from the front of the unit towards the back and expel it once it's been pulled over the motherboard. There are no fans in the front of the unit and all air has to pass multiple rows of vertically-oriented hard drives. What is notable about this design is its simplicity and how it ensures that the maximum surface area of every drive is exposed to an even air flow.

There is a huge backplane at the bottom of that unit, with each drive being pulled by gravity into its SATA connector. To access a drive, the pod is pulled out of the rack (ball bearing slides help!), the lid is opened, and the offending drive is pulled and replaced. To my limited knowledge, no SOHO server (8-16 drive) cases offer a similar, simple design. Most SOHO server cases are designed around "hot swap" drive holders that allow the hard drives to be pulled and replaced quickly from the front of the unit - even though few of them ever fail.

Hot-swap designs typically feature much more constricted air spaces between drives to accommodate the mechanisms, guides, and so on. That in turn leads to greater static pressure drop, necessitating stronger (and usually louder) fans to get good air flow. I have yet to find a 8-12-drive SOHO server with the simplicity of the backblaze pod. LIAN LI models like the Q26 come close, but the hard drive holders in there are not quite ideal for good air flow over the drives, and the case doesn't allow for a Flex-ATX motherboard.

I have been trying to keep my drives happy at below 31*C, which is the upper end of the Backblaze temperature range. Crummy air flow designs in other cases than the one I use now have driven my drives to exceed 45*C during multi-hour scrubs. Now imagine what the drive temperature would be inside a a typical external plastic case without forced cooling.

My hard drive toaster from OWC may have a known-defective design (it doesn't allow formatting via eSATA), but USB 3 is good enough, and each drive only sees it once – it's used to zero obsolete spinners before they are donated to a local medical center for reuse in research.
I have a Storinator 30 pod that houses 30 12TB hard drives. (NAS4free system.) Storinator was a Backblaze supplier for a while. These pods are very quiet, and I switched to them because they keep the drives coolish without turning on the A/C. I have used tightly packed Expanders before (bought off eBay) and, although they worked great, they were both noisy and the drives ran hot. If you order Storinators, make sure to request pods built for the Corsair PSU. The new Zippy ones they supply seem more resilient but run a lot louder.
 


I'll take that bet.
OK, where do I send that sandwich?

I could have phrased that a lot better. What I should have said is that the high-volume users out there who do use SMR would be using the host-aware variety rather than the device-managed variety. Mea culpa, hat is in hand, forgiveness I ask.

Please note: It is telling that a high-volume buyer of drives for NAS use is avoiding SMR!

The interview with Mr. Ellis that Michael Schmitt referenced is super-interesting and well worth a read. As usual, I would use the data Backblaze publishes in your own buying decisions (i.e. what drive models are reliable, what technologies they are using). Given that Backblaze is adding tens of thousands of drives per year to its pool, you'd think they'd be a relatively big fish in the buyer pond.

However, Mr. Ellis indicates that a "single tier-one data center consumes maybe 500,000 times what Backblaze does in drives." He also states that they consume about 4,800 drives a month at Backblaze, suggesting that a single tier-one data center consumes 2.4 billion drives per month or 28.8 billion drives annually. Does that seem plausible?

Seagate shipped 338 Exabytes in 2018 with the average capacity at 2.2 Terabytes.

Per Google math, that gets me to 153,636,364 drives annually, i.e. not even a billion, and they're a big supplier. Per Trendfocus, ~67 million drives total were shipped in the first quarter of 2020, of which Seagate had 42%, WD had 37% and Toshiba had 21% market share.

Even with 10% added for growth, that only gets me to 295 million drives annually for the whole industry in 2020... and there are multiple tier-one data centers out there.

The 500,000x claim appears to be uncharacteristically pure hyperbole, especially if you consider that the tier-one cloud suppliers are likely to go after the same high-capacity drives as Backblaze does, for the same reasons (power efficiency, $/TB acquisition cost, data center rental / maintenance charges, and so on).

Per the blocks and files article, high-capacity drive shipments totaled 15 million drives in Q1 2020, i.e. ~60-100 million drives for 2020 after whatever compound annual growth rate you want to apply. Of that, Backblaze will make up about 58k drives. Certainly not the biggest player (about 967 drives per million high-capacity drives sold) but certainly not 2 drives per million drives shipped, as claimed by Mr. Ellis either. He's off, maybe, by a factor of 500 for total drive shipments and likely a lot more when it comes to individual tier-one data centers.
 


Amazon disclaimer:
As an Amazon Associate I earn from qualifying purchases.

Latest posts