Hacker News Reader

Comment by azornathogron | Hacker News Reader

I believe they have overestimated the cost of using 3.5" floppies, probably to make themselves look better by comparison.

They give the cost as $211,172, but that's the cost to buy a 1 TiB pack of floppies. Their own storage cost is per-month, so to get the equivalent cost for floppies you need to also divide by the expected useful lifetime of a floppy disk. I did a web search for "floppy disk lifetime" and the internet [1] told me "I’ve seen numbers saying the lifespan of floppy disks is three to five years. But I’ve also seen numbers that claim they can last ten to twenty years or even indefinitely."

If you assume floppy disks have an expected lifetime of 5 years, you can amortise the cost across that time, bringing the cost per TiB-month down to a nice reasonable $3,520.

[1] https://blog.storagecraft.com/data-storage-lifespan/

View on HN

This suddenly reminded me of the floppy-disk raid: https://web.archive.org/web/20080117032102/http://phoenix.cc...

You've got to wonder how it would work if taken to the extreme: rack after rack of floppy drives filling an entire data centre providing a glacially-slow S3 service.

Can you imagine the noise? It would be... glorious.

The linked article used these compact drives: https://www.amazon.com/External-Floppy-1-44MB-FDUSB-M-V1/dp/...

They're 5.75 x 4.25 x 0.75 inches. So if you mount them vertically like in high-density storage arrays, you can fit 25 of them into the width of a rack, about 7 rows from the front to back. So... about 175 per layer that is about 4 RU high including the space for the controller board. You can fit 10 of these layers in a standard rack, for 1,750 floppy drives total per rack.

Let's see... that's 2.52 GB per rack! Seek times are variable depending on the floppy drive model, but 250 ms is approximately correct for the average. So about 7,000 IOPS total per rack. Not too shabby!

A decent sized data centre might have 1,000 racks. So a "cloud-scale" floppy drive object storage system might have 2.52 TB of raw storage. However, you have to divide that by three for the redundant copies, so we're back to 840 GB of usable storage capacity per floor, but with an impressive 7M IOPS.

To put things in perspective, that's directly equivalent to a single modern laptop SSD drive in terms of both capacity and IOPS. Except that the latency of the SSD is 5000x lower.

You can even make music with the sounds. On youtube there are various people doing this, like Paweł with his Floppotron.

https://youtu.be/G-X-p0C0Uas

And pretty easy at that, just just put a PWM signal on the stepper pins (watch the voltage and current though, easily fries and MCU) and there you go

Since this is a write-only service, it doesn't actually matter if the disk stops working, or even ceases to exist. Therefore we may be able to estimate the lifespan as infinite.

There may also be no need to purchase more than one disk. In fact taken to its logical conclusion there is no need to purchase any disk at all - but of course that brings us all the way to the S4 business model itself.

I don't think comparing a service like S4 to a non-existing disk is fair. Especially in enterprise software, I often get the requirement that the data is reliably written. I can easily point an auditor to the S4 service contract and SLAs to prove we fulfill this. A bespoke arrangement of floppy drives and shredders might pass if the auditor's having a good day. But removing the disks entirely? That will never fly.

Since we have multiple points of data for lifespan, perhaps we should be computing a weighted average between them?

Going by gut feel: 3-5 years sounds like "pretty likely", let's give that 95% of the weight, at 4 years for simplicity. 10-20 years would be nice, but really, how much tech lasts that long? Not much. How about 4.9999999% at 15 years. And the remaining can go to "indefinitely", which seems pretty darn unlikely, so it has 0.0000001% weight.

(4*95+15*4.9999999+∞*0.0000001 / 100) = ∞

I guess we can ignore the monthly cost ¯\_(ツ)_/¯ I applied weights to things, so this is Bayesian, and we all know that's always right.

You didn't take into account the heat death of the universe.

It won't be all that hot, so I don't really think the floppies are in any risk of melting because of it. Probably nothing to worry about.

It is the lack of heat that would be the issue.

Just roll them in a nice soft blanket. I have one that keeps me warm even in plain winter. We can add a hot chocolate cup and that should do the trick.

To be fair, less entropy means those floppies will last longer - so the heat death should marginally raise the their expected lifetime.

Given my own experience with 3.5" floppies in the past often yielded lifetimes in single digit months, I would say any estimate of years needs to have a very large error margin, and you need to use RAIF.

My experience is very similar, but I can believe that write-only or even single-read scenario could extend the expected lifetime to said years.

But at 25kbps sustained write (inc optimistic disk juggling) it'd take 463 days to write 1TB. If that's your comparison number, you'd need to factor in 17 drives running in parallel to write 1tb in February. And a robot, or 6 people to handle disks. And 14m³ archive space.

And double all that if you want redundancy.

But the abacus beads could store rotational data. There's at least 8 bits of data there, more if you can put some time in. That's free real estate there.

Assuming a write-only interface, I think I could push the cost of writing 1TB of data to a floppy disk down to the cost of a single disk.

Plus, when you’re done writing your data, you’ll have a fancy drink coaster!

i believe there is an opportunity in the market for an advanced data storage technology to support writing huge volumes of information to a single 3.5" floppy, provided the legacy constraint of offering a read operation can be dispensed with.

Easy! Once you've written to the end just seek back to sector zero and continue writing. It's a like data-layer-cake.

So do they actually provide the newest hottest proof-of-writing even? Otherwise they’re just plain scam!

I think they've underestimated it. Hard to visualise 700,000 floppy disks but I have a feeling stored at home it would cost me 1 marriage.

Why use a Floppy Disk comparison at all. I guess its for comic value.

But a 1 TB HDD and an associated Linux system would probably cost about 300 USD.

Assuming a lifespan of about 7 years, monthly cost comes to about 3 USD per month.

What's more, floppies are actually prior art for their business model. At least my AMIGA days were just endless problems with broken floppies.

Their calculations for S3 is wrong too. If you use glacial deep archive storing a terabyte costs $0.99 per month.

LS240 drives can superformat HD floppies to 32MB. Lots more potential savings to be had there.

If I recall correctly the read/write head of a floppy drive rested directly on the media. This meant the disks wore out faster the more you used them.

Reply on news.ycombinator.com