Before you mention "Cheap options" and Amazon Glacier, read this: https://news.ycombinator.com/item?id=10921365

"How I ended up paying $150 for a single 60GB download from Amazon Glacier"

> If you’re retrieving more than 5% of your stuff, expect to pay a fee, “starting at $0.011 per gigabyte”.

Key words here is "starting at". as in, no cap on how much they can and will charge you for your data. Cause you know, 60GB is realllly expensive to transfer :(

_______________________________

What Glacier seems to be for, is more like compliance paperwork and other "We are required to save this for X years" kind of paperwork. Because most of those types of documents, you don't care about and only legally have to keep available. And if a court case comes along, the Glacier retrieval price can just be added on top of court fees (which in that case, Glacier is probably a pittance even at the $10k level).

There's a lot of articles about these things that seem to have the same summary: "I didn't read the documentation for how pricing works and then I was surprised by how pricing worked." Some people were shocked by Glacier's egress pricing and others were shocked by the cost per file. Both of these are approximately zero, most of the time, for most people, if they know what they're doing. If you don't know what you're doing and haven't read the docs you shouldn't be using Glacier.

> Cause you know, 60GB is realllly expensive to transfer :(

Network and disk IO capacity planning is hard. You can't just buy GBs of egress at $0.01 and then resell them at $0.01, because you're paying for maximum capacity and turning around and selling average usage. Similarly, when you have a bunch of data shoved into unused sections of disk, you can't just read them back out without affecting whatever else is reading from the same disk. If you want to sell something close to cost, you need to reflect the pricing structure of what you're paying to your customers.

So if you upload a bunch of data to a super-cold storage system that's even named after a geological formation that stays frozen for centuries, to remind you how cold it is, and then you make it hot by trying to download it all at once, you'd expect it to be more expensive.

Just do your regular cost-benefit analysis and it should be fine, based on how long you plan on storing the data and how likely or often you think you'll need to restore it.

The root problem here is , that Amazon Glacier is as opaque as mud about pricing. Or did the words "Starting At" not pique your buzzword bullshit detector? It certainly did mine.

Pricing means sifting through a legalese dense block of text to try to come up with even a way to estimate the costs. That is a problem. I should be given a rough cost expectation up front. Sure, it's not going to be exact and all, but This article is about a cost factor of *180.. That's the difference between an iPad and a Ferrari.

Other places of varying types of backup systems are a lot more clearer to tell how much I should be expected to pay. Or at least, I get within +/- 10%. Glacier? "Yeah, Don't worry about it till we bill ya!" That's the scam with all these cloud services really. Once we have your nuts in a vice, we can squeeze and extract, cause you don't have any other choice, now do you?

I think the information you're working with might be out of date, I don't see the words "starting at" anywhere. Glacier egress pricing is $0.01 per GB and $0.05 per 1,000 requests. That's for standard retrievals. Not so complicated.

Before Feb 2017, it did work with the peak hourly request fees. Those were well documented. Apparently peak transfer rates were "confusing". I didn't think so, after all, that's similar to how I pay for internet access at home (I pay based on capacity, not based on bytes transferred).

Glacier pricing has changed since that post. It's now a much simpler model, around $0.10/GB to download. The original pricing was very complicated though!

https://aws.amazon.com/glacier/pricing/

So how would you categorize CrashPlan? They offer unlimited storage for $60/year. I've had around 8TB backed up there for the last 2-3 years. They have existed since 2001 (no idea how long the unlimited feature has existed).

On the other hand, supporting your point #1, I used to have several TB on StreamLoad which 10ish years ago used to also offer unlimited... until they suffered a critical hardware failure and lost tons of customer data. They then changed to MediaMax, then suddenly disappeared.

Perhaps you know of others, or this list is incomplete, but Wikipedia has a list of file services, of which there aren't many who offer an unlimited plan:

https://en.wikipedia.org/wiki/Comparison_of_file_hosting_ser...

CrashPlan is "unlimited" but it's not a storage service, it's a backup service. From what I understand, you have to use their client software, and that software will use RAM proportional to the total size of the data you're storing. In practice this puts a hard limit on the amount of data which is proportional to the amount of RAM you have.

So heavy users are not only capped, they're also subsidized by the light users.

I've got about 12TB on Crashplan, and the client's never used more than about 2GB. I doubt your "cap" applies in practice.

So about 170mb/TB if this scales linearly for some reason.

Some users were apparently bragging about having 1PB+ on ACD. That would mean the RAM usage would actually be a useful upper bound, as I doubt many of these guys have 170GB of RAM in their media server PCs.

In addition to the comment below about CrashPlan requiring you to use their client software, which limits abusabiliy -- I am also seeing signs of strain from CrashPlan. They have always had a policy that claimed they would delete data from machines that hadn't touched the service in a long time, but they had never implemented this in practice before; just last year they finally started sending out threatening letters to people saying they're going to delete data on short notice. (And they didn't really back down when people got angry.) This tells me the storage expense is starting to hurt them.

Yep, I cancelled my CrashPlan account when they pulled that. IIRC they gave me under a week to either download the data for 5-6 machines, or lose it forever.

> So how would you categorize CrashPlan? They offer unlimited storage for $60/year. I've had around 8TB backed up there for the last 2-3 years.

CrashPlan's upload speed is a joke for people located outside North America. I had an account for ~3 years, but found it difficult to upload more than 6TB. (There are rumours that CrashPlan throttles uploads.)

On the other hand, I've managed to backup most of my ~15TB NAS (and more) on Amazon Cloud Drive.

I'm in the UK and evaluated CrashPlan a couple of years ago - IIRC, it was so slow it was going to take several weeks to upload my 1TB of data.

Crashplan is a backup service. Different category of company.

Ever hear of Sia? That's a new decentralized option gunning for Amazon. Website is sia.tech.

It's very difficult to take Sia seriously. For one thing, Sia only has 25T of data. That's total data, in the entire global network, including parity blocks. And when I checked prices today, it was more expensive than Glacier—but I can only imagine how much the prices would swing if you suddenly bought 2.5T of storage.

Sia's decentralization is not unique. Amazon, Google, and Microsoft cloud storage are also decentralized. The idea of reselling unused bytes on disks that you already own is also not unique—again, Amazon, Google, and Microsoft already do this.

Sia's decentralization is unique. It's unique not because it spreads data all over the world, but because there's no central player that controls prices or decides to shut things down. Amazon, Google, Microsoft, they can change terms or disable services whenever they want. Sia is a lot more robust, because it's governed via a blockchain, and you put your data on 30+ individual parties, making it extremely unlikely that a sufficient number of them all shutdown simultaneously.

This situation with retiring unlimited storage on ACD would not be possible with Sia. Amazon is changing terms and conditions, Sia is a blockchain where no party has the power to do that.

-----

To address your pricing comment, there are more than 1000 TB for sale on the Sia network. Adding 2.5 TB to the network would not move the price at all, that's really not how pricing works on Sia.

The reason that it's more expensive than it used to be is because prices are set entirely in Siacoin. When the siacoin prices rise, hosts need to manually re-adjust their prices to keep the same USD price. As of writing, there are no tools that will let this happen automatically. The siacoin price has doubled 6 times in 6 months. The result is that a lot of the defaults are now grossly expensive, and hosts need to be very familiar with the pricing mechanisms to respond accordingly. Most have not, though the ones that have are seeing much higher utilization than the ones that haven't.

We will be releasing stuff in the next month to help hosts set prices more intelligently, that should move the prices back to the competitive spot that they historically been at. The high prices right now are merely a result of market confusion among the hosts.

Hi Taek, Sia recently came on my radar, and it seems pretty interesting usage of the blockchain, and I've started following it.

Is there a good place to keep an eye on the roadmap for the technology (with info like what you had in the last paragraph)?

> The siacoin price has doubled 6 times in 6 months.

Presumably this represents entirely speculative activity driven by the other fashionable blockchains?

I'm sure that's part of it, but you can't really avoid that. But there are a ton of altcoins out there now, and most of them don't have interesting tech behind them or much to distinguish them from any of the others. Sia seems to have interesting tech behind it, to me at least.

Have to be careful with Glacier: the retrieval costs can be very high and are extremely hard to calculate.

Amazon S3 Infrequent Access has higher retrieval costs, though not as crazy as Glacier, and a delete penalty. Deleting a file after 1 day will cost 30x normal S3 storage fees.

Likewise, Coldline sounds great, but they have higher retrieval fees (not so bad) and a delete penalty. The Coldline delete penalty will cost 90x the regular storage cost if a file is deleted the next day. I don't recommend Coldline unless you know you won't be deleting files before 90 days.

If it is backup, it doesn't need to be super safe. The risk that you lose both your primary data, (optionally your local backup) and your online backup at the same time is pretty insignificant, given that the online backup will be uncorrelated as in a different physical location (primary and local backup very correlated I agree).

> The risk that you lose both your primary data, (optionally your local backup) and your online backup at the same time is pretty insignificant, given that the online backup will be uncorrelated as in a different physical location (primary and local backup very correlated I agree).

Actually, that reasoning is exactly what I wanted to talk about. When the primary is lost, there is a much higher chance than you'd expect that a secondary contains unrecoverable errors. This is why RAID 5 arrays fail to rebuild after a disk failure so often—they're supposed to be able to tolerate a single disk failing, but they can't tolerate a disk failing and any other IO error at the same time. Part of this is due to how short the timeout is for failed reads in RAID setups, but I've still seen a lot of RAID 5 arrays fail, and I've seen a few RAID 6 arrays fail too.

On top of that, there's the high chance of configuration errors in DIY systems.

Agree. Having a script doing a full read of all the data every couple of months and sending a report by email (that you would notice if not sent) is a sort of must have.

I am still looking for a good way of doing this.

Weekly integrity check of all backed up data. E-Mail which informs of result. Web interface which shows overview of results of historical checks. External service which sends E-Mail if integrity check failed to run (e.g. https://deadmanssnitch.com/).

FreeNAS gets fairly close to this out of the box. My home server runs ZFS with RAIDZ2. By default, I think, there's a weekly cronjob to scrub the ZFS pool (integrity check everything, as I understand it), and then the results of that scrub are emailed to me.

I don't believe it has a web interface with historical checks, although I could be wrong. That said, it might be stored in a log file somewhere.

I also don't have an external service that would send me an email if it failed to run. That said, I would get an email if cronjobs had a mysterious error; otherwise, if the server itself was dead, my data would not be accessible on my home network, so I'd notice.

If the home server dies tragically, well, I hope Google Cloud Storage is doing similar integrity checks -- that's where my offsite backups are.

Integrity is harder because you have to sort of maintain a signature of every file on the side. What I do is to just have a script that reads all data. If some data is unreadable an exception will be thrown and a text message will be sent to me. And a confirmation email is sent otherwise.

In .net it's only a few lines of codes. Haven't thought of deadmansnitch but it's a good idea. Would just take one more line of code.

git-annex has fsck command that can test the data and rclone remote so you can use any configuration of "cloud" storages to hold the chunks.

I would argue the opposite. When you reach for your backup, you are in a dire situation. If it fails you then, things will be very gloom.

The problem is sometimes you only find out your backup is corrupted when you need it.

If you don't test your backups, then they're not backups.

yup, the assertion that both failing is highly unlikely is under the assumption that you are checking the status of each copy regularly.

In fact for any disk, it is important to have a script that reads all the data at least every couple of months. It will force the bad sectors to be identified or to be notified you have a bad disk before things get worse.

Yeah, agree. And I am not talking about enterprise level of reliability. More like personnal / small office.

But for anything above a few TB, running your own hardware will be way cheaper if you take say a 5y horizon. Disks are so large today that you don't need a big config. In fact you might even keep two copies to reduce the risk.

Of course there is the occasional trip to the datacentre. I made that mistake. I pay more in uber than I saved by picking a datacentre far away.

Do you need co-location though? I dropped an HP Microserver at a friend's place, which backs up snapshots of my main backup server at home daily. Seems to work well so far, at quite limited costs. Since my main backup server also backs up his NAS we both win. Since all backups are encrypted there is also little privacy risk from theft (or snooping on one another).

With 1Gbps symetrical connections slowly becoming more frequent this would be a much cheaper alternative (assuming you only run it for backups, I also run a mail server and some websites). But in London it is not very practical. I know very few people who have really fast broadband.

My backup contains files and old versions of files I no longer have on the primary storage. It is much more than a 1:1 copy, and can not be recreated.

I need to feel safe that I can reasonably undo my changes on the primary storage, even if it takes me years to realise what I did.

This is why I keep two backups. Until now, the secondary offsite copy was on ACD...

That's not a backup. That's an archive. You need to backup archives the same way you do any other data you care about.

If it is backup, then it needs to be scrubbed regularly to detect and repair bit rot.