Fire declared in OVH SBG2 datacentre building
travaux.ovh.net
travaux.ovh.net
So when https://en.wikipedia.org/wiki/May_1998_riots_of_Indonesia started happening, we heard some harrowing stories of US employees being abducted, among other things.
Around that same time, the equipment in the Jakarta store started sending high temperature alerts prior to going offline. Our NOC wasn't able to reach anyone in the store.
The alerts were quite accurate: that was one of the many buildings that had been burned down in the riots. I guess it's slightly surprising that electrical power to the equipment in question lasted long enough to allow temperature alerting. Most of our stores back then used satellite for their permanent network connection, so it's possible telcom died prior to the fire reaching the UPC office.
In a couple of prominent places in the home office, there were large cutouts of all of the countries WalMart was in at the time up on the walls. A couple of weeks after this event, the Indonesia one was taken down over the weekend and the others re-arranged.
I remember during the 2000s and 2010s that WalMart in the USA earned a reputation for it's inventories primarily consisting of Chinese-made goods. I'm not sure if that reputation goes all the way back to 1998, but it makes me wonder if WalMart was especially targeted by the anti-Chinese element of the Indonesian riots because it.
Prior to the early 2000s, WalMart had a strong 'buy American' push. It was even in their advertising at the time, and literally written on the walls at the home office in Bentonville.
Realities changed, though, as whole classes of products were more frequently simply not available from the United States, and that policy and advertising approach were quietly dropped.
Just for the hell of it, I did a quick youtube search: "walmart buy american advertisement" and this came up: https://www.youtube.com/watch?v=XG-GqDeLfI4 "Buy American - Walmart Ad". Description says it's from the 1980s, and that looks about right.
The switch had already begun.
Part of what my team (Network Management) did was write code and tools to automate all of the various things that needed to be done with networking gear. A big piece of that was automatically discovering the network. Prior to our auto discovery work, there was no good data source for or inventory of the routers, hubs, switches, cache engines, access points, load balancers, VOIP controllers...you name it.
On the surface, it seems scandalous that we didn't know what was on our own network, but in reality, short of comprehensive and accurate auto discovery, there was no way to keep track of everything, for a number of reasons.
First was the staggering scope: when I left the team, there were 180,000 network devices handling the traffic for tens of millions of end nodes across nearly 5,000 stores, hundreds of distribution centers and hundreds of home office sites/buildings in well over a dozen countries. The main US Home Office in Bentonville, Arkansas was responsible for managing all of this gear, even as many of the international home offices were responsible for buying and scheduling the installation of the same gear.
At any given time, there were a dozen store network equipment rollouts ongoing, where a 'rollout' is having people visit some large percentage of stores intending to make some kind of physical change: installing new hardware, removing old equipment, adding cards to existing gear, etc.
If store 1234 in Lexington, Kentucky (I remember because it was my favorite unofficial 'test' store :) was to get some new switches installed, we would probably not know what day or time the tech to do the work was going to arrive.
ANYway...all that adds up to thousands of people coming in and messing with our physical network, at all hours of the day and night, all over the world, constantly.
Robust and automated discovery of the network was a must, and my team implemented that. The raw network discovery tool was called Drake, named after this guy: https://en.wikipedia.org/wiki/Francis_Drake and the tool that used many automatic and manual rules and heuristics to map the discovered networking devices to logical sites (ie, Store 1234, US) was called Atlas, named after this guy: https://en.wikipedia.org/wiki/Atlas_(mythology)
All of that background aside, the interesting story.
In the late 90s and early 2000s, Drake and Atlas were doing their thing, generally quite well and with only a fairly small amount of care and feeding required. I was snooping around and noticed that a particular site of type International Home Office had grown enormously over the course of a few years. When I looked, it had hundreds of network devices and tens of thousands of nodes. This was around 2001 or 2002, and at that time, I knew that only US Home Office sites should have that many devices, and thought it likely that Atlas had a 'leak'. That is, as Atlas did its recursive site mapping work, sometimes the recursion would expand much further than it should, and incorrectly map things.
After looking at the data, it all seemed fine. So I made some inquiries, and lo and behold, that particular international home office site had indeed been growing explosively.
The site's mapped name was completely unfamiliar to me, at the time at least. You might have heard of it: https://en.wikipedia.org/wiki/Shenzhen
I was seeing fingerprints in our network of WalMart's whole scale switch to sourcing from China.
You're correct, if Drake (the always running discovery engine) didn't detect a device on a given port over a long enough time, then another program would shut that port down. This was nominally done for PCI compliance, but of course having open, un-used ports especially in the field is just a terrible security hole in general.
In order to support legit equipment moves, we created a number of tools that the NOC and I believe Field Support could use to re-open ports as needed. I think we eventually made something that authorized in-store people could use too.
As an aside, a port being operationally 'up' wasn't by itself sufficient for us mark the port as being legitimately used. We had to see traffic coming from it as well.
You mentioned elsewhere that you're working with a big, legacy Perl application, porting it to Python. 99% of the software my team at WalMart built was in Perl. (: I'd be curious to know, if you can share, what company/product you were working on.
Walmart must have been an interesting place to work during the late 90s, early 2000s - I imagine that most everywhere they had to solve problems at scale before scale was a considered a thing. I'd be very interested to see how the solutions created in that period match to best-practices today, especially since outside of the telecom or perhaps defense worlds there probably wasn't much prior art.
As for the Perl application, I probably shouldn't say since I'm still employed at the same company and I know coworkers who read HN. If you're interested, DM me and I can at least provide the company name and some basic details.
Definitely. (: I didn't hold it against the hands on workers like yourself. Even (and perhaps especially) back then, WalMart was a challenging, difficult and aggressive partner.
> working on networking gear at Walmart meant that you were out at 11pm at night
That sounds about right; the scheduling I was directly aware of was very fast paced. Our Network Engineering Store Team pushed and pushed and pushed, just as they were pushed and pushed and pushed.
> Walmart must have been an interesting place to work during the late 90s, early 2000
Yup. Nowhere I've worked before or since had me learning as much or getting nearly as much done. It was an amazingly positive experience for me and my team, but not so positive for a lot of others.
> I imagine that most everywhere they had to solve problems at scale before scale was a considered a thing.
Sometimes I imagine writing a book about this, because it's absolutely true, all over Information Systems Division.
For a time in the early 2000s, we were, on average, opening up a new store every day, and a typical new store would have two routers, two VOIP routers, two cache engines, between 10 and 20 switches, two or four wireless access point controllers and dozens of AP endpoints. That was managed by one or two people on the Network Engineering side, so my team (Network Management) wrote automation that generated the configs, validated connections, uploaded configs, etc etc etc. (Not one or two people per store: one or two people for ALL of the new stores.)
The networking equipment was managed by a level of automation that is pretty close to what one sees inside of Google or Facebook today, and we were doing it 20 years ago.
> ... telecom ... prior art ...
John Chambers, the long-time CEO of Cisco, was at the time on WalMart's board of directors. He was always a bit of a tech head, and so when he came to Bentonville for board meetings, he'd often come and visit us in Network Engineering.
Around 2001-2002, we were chatting with him and he asked why we weren't using Cisco Works to manage our network. https://en.wikipedia.org/wiki/Cisco_Prime but back then it was mostly focused on network monitoring and to a lesser extent, config management. We chuckled and told him that there's no way that Cisco Works could scale to even a fraction of our network. He asked what we used, and of course we showed him the management system we'd written.
He was so impressed that he went back to San Jose, selected a group of Cisco Works architects, had them sign NDAs, and sent them to Bentonville, Arkansas for a month. The intent was to have them evaluate our software with an eye toward packaging it up and re-selling it.
Those meetings were interesting, but ultimately fruitless. The Cisco Works architects were Ivory Tower Java People. The first thing they wanted to see was our class hierarchy. We laughed and said we had scores of separate and very shallow classes, all written in Perl, C and C++.
Needless to say, they found the very 'rough and ready' way our platform was designed to be shocking and unpalatable. They went back and told Chambers that there was literally no way our products could be tied together.
> ... match to best-practices today ...
Professionally, I've been doing basically the same kinds of things since then, and I'll say that while our particular methods and approaches were extremely unusual, the high level results would meet or perhaps exceed what one gets with 'best practices' seen today.
Not because we were any smarter or better, but because we had no choice but to automate and automate effectively. At that scale, at that rate of change, at those uptime requirements, 'only' automating 99% would be disastrous.
FWIW, my brain was going "book! book! book! book! book!" back at the top-level comment, and the beeper may have got slightly overloaded and broke as I continued reading. :)
Yes please.
As a sidenote, the story about "the CEO vs the architects" was very fascinating: the CEO could see the end-to-end real-world value of what you'd built, but the architects couldn't make everything align. In a sense the CEO was more flexible than the architects, despite the fact that stereotypes might suggest the opposite being more presumable.
Also, the sentiment about your unusual methodology exceeding current best practice makes me wonder whether you achieved so-called "environmental enlightenment" - where everything clicks and just works and makes everyone who touches the system a 5x developer - or whether the environment simply had to just work really really well. Chances are the former is what everyone wishes they'll find one day, while the latter (incredibly complex upstream demands that are not going to go away anytime soon and which require you to simply _deliver_) definitely seems like the likelier explanation for why the system worked, regardless of the language it was written in - it was the product of a set of requirements that would not accept anything else.
Hmm. Now I think about that a bit and try and apply it to "but why is current best practice worse", I was musing the other day about how a lot of non-technical environments don't apply tech in smart ways to increase their efficiency, because their fundamental lack of understanding in technology means they go to a solutions provider, get told "this will cost $x,xxx,xxx", don't haggle because they basically _can't_, and of course don't implement the tech. I wonder if the ubiquitification (that seems to be a word) of so-called "best practices" in an area doesn't function in a similar way, where lack of general awareness/understanding/visibility in an area means methodology and "practices" (best or not) aren't bikeshed to death, and you can just innovate. (Hmm, but then I start wondering about how highly technically competent groups get overtaken by others... I think I'll stop now...)
Some interesting facts to know for those who don't dig into it. Walmart: - has 80+ internal apps, mostly variants but still unique - runs k8s inside of Distribution Centers - maintains a fleet of >180k mobile devices in the US alone - has a half-dozen data centers in the US - has most International infrastructure seperate from US Stores'
I've got some stories of my own, maybe I'll post them in a bit.
Wow, that's a hell of a change! When I left in 2009, there were exactly two datacenters: NDC and EDC. Not surprising really.
From where I was sitting, the best era was definitely 1997-2004 or so. ISD really went down hill, pretty quickly, in my last five years there, for many different reasons.
Really though, I feel truly awful for anyone affected by this. The post recommends implementing a disaster recovery plan. The truth is that most people don't have one. So, let's use this post to talk about Disaster Recovery Plans!
Mine: I have 5 servers at OVH (not at SBG) and they all back up to Amazon S3 or Backblaze B2, and I also have a dedicated server (also OVH/Kimsufi) that gets the backups. I can redeploy in less than a day on fresh hardware, and that's good enough for my purposes. What's YOUR Disaster Recovery Plan?
The SBG fire illustrates the importance of geographical redundancy. Just because the datacenters have different numbers at the end doesn't mean that they won't fail at the same time. Apart from a large fire or power outage, there are lots of things that can take out several datacenters in close vicinity at the same time, such as hurricanes and earthquakes.
Restic has a single binary that takes care of everything. It feels more modern and seems to work really well. Never had any issue restoring from it.
Just one data point. Stick to whatever works for you. But important to test not only your backups, but also restores!
GPG is a bit quirky but I do regularly check my backups and restores (if once every few months counts as regular).
It's brilliant, works like a charm on freebsd windows and a rpi with linux since over 2 years.
I've found duplicity to be a little simplistic and brittle. Purging old backups is also difficult, you basically have to make a full backup (i.e. non-incremental) before you can do that, which increases bandwidth and storage cost.
Restic looks great feature-wise, but still feels like the low-level component you'd use to build a backup system, not a backup system in itself. It's also pre-1.0.
Have moved to using rdiff-backup over SSH.
https://www.gnu.org/software/tar/manual/tar.html#Incremental...
It's very easy to use spare storage in various places to do backups this way, as ssh, gpg and cron are everywhere, and you don't need to install any complicated backup solutions or trust the backup storage machines much.
All you have to manage centrally is private keys for backup encryption, and CA for signing the ssh keys + some occasional monitoring/tests.
pretty much a textbook use-case for zfs with some kind of snapshot-rolling utility. Snap every hour, send backups once a day, prune your backups according to some timetable. Transfer as incrementals against the previous stored snapshot. Plus you get great data integrity checking on top of that.
"but linus said..."
Yes i still don't understand him, a he calls himself a "filesystem guy". Also i don't understand that no one ever mentions NILFS2.
The draw of ZFS is that it's the log-structured filesystem with 10 zillionty hours of production experience that says that it works. And that's why BTRFS is not a direct substitute either. Or Hammer2. There are lots of things that could be cool, the question is are you willing to run them in production.
There is a first-mover advantage in filesystems (that occupy a given design and provide a given set of capabilities). At some point a winner sucks most of the oxygen out of the atmosphere here. There is maybe space for a second place winner (btrfs), there isn't a spot for a fourth-place winner.
If OVH is totally down, and the fail over IP doesn't work, I have a fairly low TTL on the DNS.
I backup the database state to S3 every day.
Since I'm truly paranoid, I have an Intel NUC at my house that also replicates the DB. I like knowing that I have a complete backup of my entire business within arm's reach.
It could also provide a burglar a fantastic opportunity to pivot into career in data breaches.
These solutions are not one-size-fits-all. What works for a small startup isn't appropriate for a 100+ person company.
There is nothing magical about data centers making them safe while your local copy isn't.
Is this a serious comment? My house is not certified as being compliant with any security standards. Here's the list that the 3rd party datacenter we use is certified as complaint with:
https://aws.amazon.com/compliance/programs/
The data centers we operate ourselves are audited against several of those standards too. I guess you're right that there's nothing magic about security controls, but it has nothing to do with trust. Sensitive data should generally never leave a secure facility, outside of particularly controlled circumstances.
You are entierly missing the point by quoting the compliance programs followed by AWS whose sole business is being a third party hoster.
For most business, what you call sensitive data is customers and orders listing, payment history, inventory if you are dealing in physical goods and HR related files. These are not state secrets. Encryption and a modicum of physical security go a long way.
I personally find the idea that you shouldn't store a local backup of this kind of data out of security concern entirely laughable. But that's me.
This attitude that your data isn’t very important, so it’s fine to not be very concerned about it’s security, while not entirely uncommon, is something most organisations try to avoid when choosing vendors. It’s something consumers are generally unconcerned about, until a breach occurs, and The Intercept write an article about it. At which point I’m sure all the people ITT who are saying it’s fine to take your production database home would be piling on with how stupid the company was for doing ridiculous things like taking a copy of their production database home.
I said there was nothing magical about data centers security, a point I stand with.
It's all about proper storage (encryption) and physical security. Obviously, the physical security of an AWS data center will be tighter that your typical SME but in a way which is of no significance to storing backups.
> This attitude that your data isn’t very important
You are once again missing the point.
It's not that your data isn't important. It's that storing it encrypted in a sensible place (and to be clear by that I just mean not lying around - a drawer in an office or your server room seems perfectly adequate to me) is secure enough.
The benefits of having easily available backups by far trump the utterly far fetched idea that someone might break into your office to steal your encrypted backups.
In the SME space some things are "different", and if you've not worked there it can be hard to get one's head around it:
A client of mine was burgled some years ago.
Typical small business, offices on an industrial estate with no residential housing anywhere nearby. Busy in the daytime, quiet as the grave during the night. The attackers came in the wee small hours, broke through the front door (the locks held, the door frame didn't), which must have made quite a bit of noise. The alarm system was faulty and didn't go off (later determined to be a 3rd party alarm installer error...)
All internal doors were unlocked, PCs and laptops were all in plain sight, servers in the "comms room" - that wasn't locked either.
The attacker(s) made a cursory search at every desk, and the only thing that was taken at all was a light commercial vehicle which was parked at the side of the property, its keys had been kept in the top drawer of one of the desks.
The guy who looked after the vehicle - and who'd lost "his" ride - was extremely cross, everyone else (from the MD on downwards) felt like they'd dodged a bullet.
Physical security duly got budget thrown at it - stable doors and horses, the way the world usually turns.
And as noted, encrypted backups would be resistant to casual interdiction, or even strongly-motivated attempts. Data loss being the principle risk mitigated by off-site, on-hand backups.
I also run our entire production system on one beefy machine at OVH, and replicate to a similar machine at Hetzner. In case of a failure, we just change DNS, which has a 1 hour TTL. We've needed to do an unplanned fail-over only once in over 10 years.
And like you, I have an extra replica at the office, because it feels safe having a physical copy of the data literally at hand.
Again, I may be completely wrong but why would you not use AWS/GCP? Even if it's complexity, Amazon have Lightsail, or if it's cost I thought DigitalOcean was one of the only reputable business-grade VPS providers.
I just can't imagine many situations where a VPS would be superior to embracing the cloud and using cloud functions, containers, instances with autoscaling/load balancers etc.
It has twice the revenue, and is the third largest hosting provider in the world.
In any case, they aren't "primarily for personal use".
I've been making this point for a long time. Both of those AS spaces are legendary for the volume of dodgy packets they barf at the rest of us.
drop
drop
drop
Why not use AWS/GCP? From my personal point of view: as a French citizen, I'm more and more convinced that I can't completly trust the (US) big boys for my own safety. Trump showed that "US interest" is far more important than "customer interest" or even "ally interest". And moreover, Google is showing quite regurlaly that it's not a reliable business partner (AWS look better for this).
Interesting, any examples?
As a customer (or maybe an "involontary data provider"), I do as much as I can to avoid Google to be my SPOF, not technically (it's really technically reliable) but on the business side. I had to setup my own mail server just to avoid any risk of google-ban for example... just in case. I won't use Google authentificator for the same reason. I'm happy to have left Google Photos some years ago, to avoid problems of Google shutting it down. And the list could go on...
As a business, I like to program Android apps but the Google Store is really a risk too. Risk to have any Google account blacklisted because some algorithm thought I did something wrong. And no appeal.
Maybe all this doesn't apply to GCP customers. Maybe GCP customers have a human direct line, with someone to really help and the capacity to do it. Or maybe it's just Google: as long as it work, enjoy. If it doesn't, go to (algorithmic) hell.
..and they do support all of those things you just listed, not just API-backed bare metal.
Their sub-brand soyoustart has older servers (that are still perfectly fine), roughly E3 Xeon/16-32GB/3x2TB to 4x2TB for $40/m ex vat.
Their other sub brand kimsufi for personal servers has Atom low-power bare metal with 2TB HDD (in reality it is advertised 500GB/1TB, but they don't really have any of those in stock left, if your drive fails they replace it with a 2T - so far this has been my exp) for $5.
All of this is powered by automation, you don't really get any support and you are expected to be competent. If your server is hacked you get PXE-rebooted into a rescue system and can scp/rsync off your contents before your server is reinstalled. OS installs, reboots, provisioning are all automated, there's essentially no human contact.
PS: Scaleway, in Paris, used to offer $2 bare metal (ultra low voltage, weaker than an Atom, 2GB ram), but pulled all their cheap machines, raised prices on existing users, and rebranded as enterprisey. The offer was called 'kidechire'
--
It is kind of interesting that on the US side everyone is in disbelief, or like "why not use AWS" - while most of the European market knows of OVH, Hetzner, etc.
My own reason for using OVH? It's affordable and I would not have gotten many projects (and the gaming community I help out with) off the ground otherwise. I can rent bare metal with NVMe, and several terabytes of RAM for less than my daily wage for the whole month, and not worry about per-GB billing or attacks. In the gaming world you generally do not ever want to use usage based billing - made the mistake of using Cloudfront and S3 once and banned script kiddies would wget-loop the largest possible file from the most expensive region botnet repeatedly in a money-DoS.
I legitimately wouldn't have been able to do my "for-fun-and-learning" side projects (no funding, no accelerator credits, ...) without someone like them. The equivalent of a digitalocean $1000/m VM is about $100 on OVH.
32c xeon/256GB ECC/500GB SSD 8TB HDD is $100/m at OVH. The difference is amusing.
Yann, CEO at Scalingo
Like kimsufi equivalent brand "isgenug"
(This is 2011, I think it looks fancier now)
https://lafibre.info/ovh-datacenter/data-center-ovh-roubaix-...
Edit: Seems like they stopped publishing videos for that datacenter, but this seems to be a video for the burn down datacenter in 2013: https://www.youtube.com/watch?v=Y47RM9zylFY
OVH is one of the better deals for bare metal, but there are even better ones for bandwidth. You have to shop around a lot.
Also be sure you have a recovery plan... even with the big providers. These days risks include not only physical stuff but some stupid bot shutting you off because it thinks you violated TOS or is reacting to a possibly malicious complaint.
We had a bot at AWS gank some test systems once because it thought we were cryptocurrency mining with free credits. We weren’t, but we were doing very CPU intensive testing. I’ve heard of this and worse happening elsewhere. DDOS detector and IDS bots are particularly notorious.
I think small bussinesses like smaller simple providers instead of bigclouds. Its different philosophy if you are afraid of extreme centralisation of internet it makes sense.
They also don't charge for bandwidth, although some of those other providers have a generous free bandwidth and cheap overage.
At OVH? If so, their US data centers don't seem to have that option.
Not that I need it. The largest database I run could easily fit in RAM on a reasonably sized dedicated box.
I didnt know.
It's a 4 core 8 thread Xeon with 3x 1TB SATA with 32GB of ECC RAM IIRC (E3-SAT-1-32, got it during a sale with a price that is guaranteed as long as I keep renewing it)
The thing is great, I can run a bunch of VM's on it, it runs my websites and email.
Overall to get something comparable elsewhere I would be paying 3 to 4 times as much.
I would consider $50 a month or less low end pricing. ¯\_(ツ)_/¯
But i assume they are less known in US.
What's most relevant, in your case, would you say? Evil maiden attack, or browser zero day, or spear phishing? An insider? Or something else?
(And how did you arrive at that threat model, if i may ask)
If my money and/or job depended on having something running without (or with minimal) disruption I would be as paranoid as you, too.
BTW - Some people call this business recovery plan, not plain paranoia ;-)
I’ve really struggled working with others that either are both ignorant and apathetic about the business’s ability to deal with risk or believe that it’s their job to keep putting duct tape over the duct tape that breaks multiple times a day while users struggle.
I like seeing these comments reminding others to a wear seat belt or have backups for their backups, but I don’t know whether I should care more about reliability. I work in an environment that’s a constant figurative fire.
I also like to spend time with my family. I know it’s just a job, and it would be even if I were the only one responsible for it; that doesn’t negate the importance of reliability, but there is a balance.
If you are dedicated to reliability, don’t let this deter you. Some have a full gas tank, which is great.
Apathetic people though, that can be really tough going. It’s just that way “because”. Or my favourite “oh we don’t have permission to change that”, how about we make the case and get permission? _horrified looks_ sometimes followed by pitch forks.
At the end of the day you have a budget (of any kind) and a list of priorities on which to spend it. It's up to you or your management to set a reasonable budget, and to set the right priorities. If they refuse, leave or you'll just burn the candle at both ends and just fade out.
A backup on its own is of little worth if unused.
When a backup is used to re-enable something, then the amount of time disabled may be decreased. When it is, this is reliability- we keep things usable and in function, more than not.
> ... [F]inance is fundamentally about moving money and risk through a network. [1]
Your employer has taken on many, many risks as part of their enterprise. If every risk is addressed the company likely can’t operate profitably. In this context, your business needs to identify every risk, weigh the likelihood and the potential impact, decide whether to address or accept the risk, and finally, if they decide to address the risk, whether to address it in-house our outsource it.
You’ve identified a risk that is currently being “accepted” by your employer, one that you’d like to address in-house. Perhaps they’ve taken on the risk unintentionally, out of ignorance.
As a professional the best I can do is to make sure that the business isn’t ignorant about the risk they’ve taken on. If the risk is too great I might even leave. Beyond that I accept that life is full of risks.
[1] Gary Gensler, “Blockchain and money”, Introduction https://ocw.mit.edu/courses/sloan-school-of-management/15-s1...
So they might have their primary and replica databases at different DCs from the same hosting provider, and only their nightly backup to a different provider. Four copies to four different providers is a step above three copies with two providers!
A large enterprise would probably be using a filesystem with periodic snapshots, or streaming their redo log to a backup, to protect against a fat-fingered DBA deleting the wrong thing. Of course, filesystem snapshots provide no protection against loss of DC or wrongful hosting account termination, so you might not count them as true backup copies.
e.g., if you are an AWS customer, have your back ups in S3 and use simple replication to sync that to either GCS or Azure, where you can get the same level of compliance attestation as from AWS.
(My servers aren't in SBG either - phew!)
There's a whole class of (mostly non-technical) risks that you solve for when you do this.
If anything happens with your payment method (fails and you don't notice in time; all accounts frozen for investigation), OVH account (hacked, suspended), OVH itself (sudden bankruptcy?), etc, then at least you have _one_ other copy. It's not stuff that's likely to happen, but the cost of planning for it at least as far as "haven't completely lost all my data even if it's going to be a pain to restore" here is relatively minimal.
Currently trying to reduce that "less than a day" though. Recently discovered "ReaR" (Relax and Recover) from RedHat and sounds really nice for bare-metal servers. Not everybody runs on virtualized/cloud (being able to recover from VM snapshots is really a plus). Let's share experiencies :)
Hot spare on a different continent with replicated data along with a third box just for backups. The backup box gets offsite backups held in a safe with another redundant copy in another site in another safe.
Restores are tested quarterly.
Keep backups of backups. Once bitten, twice shy.
Just be cautious about data locality laws (not likely to affect you as joe average, more for businesses)
For example, the Russian data residency law states that a copy of the data must be stored domestically, not that it can't be replicated outside the country.
The UAE has poorly written laws that have different regulations for different types of data - including fun stuff like only being subject to specific requirements if the data enters a 270 acre business park in Dubai.
Don't even get me started on storing encrypted data in one country and the keys in another...
This was so we remained legal in all of the countries BT worked in which required a lot of behind the scenes work to make sure we didn't serve "illegaly encypted" Data.
Probably this is the most important part of your plan. It's not the backup that matters; it's the restore. And if you don't practice it from time to time, it's probably not going to work when you need it.
Everything is managed with Ansible and Terraform (on DO side), so I could probably get everything back up and running in less than an hour if needed.
That makes it sound like you didn't try/practice. I imagine that in a real-life scenario things will be a little more painful than in one's imagination.
Even if it possible to trigger and test, actually using the recovery mechanism may have some high cost either monetarily or maybe losing some small amount of data. These mechanisms should almost always be an additional layer of defense and only be invoked in case of true catastrophe.
In both cases, the mechanisms should be tested as thoroughly as possibly, either through artificial environments that can simulate improbable scenarios or in the latter case on a small test environment to minimize cost.
Lots of people plan for specific elements they can imagine and forget other equally or even more important things they are going to need in a disaster. Check out how many organisations that doubtless have 24/7 IT support in case a web server goes down somehow had no plan for what happens if it's unsafe for their 500 call centre employees to sit in tiny cubicles answering phones all day even though pandemic respiratory viruses are so famously likely that Gates listed them consistently as the #1 threat.
Fast Fashion for example often employs workers in more or less sweatshop conditions close to the customers (this makes commercial sense, if you make the hot new items in Bangladesh you either need to expensively air freight them to customers or they're going to take weeks to arrive after they're first ordered - there's a reason it isn't called "Slow fashion"). These jobs are poorly paid, many workers have dubious right-to-work status, weak local language skills, may even be paid in cash - and so if you tell them they must come in, none of them are going to say "No".
In fact the slackening off in R for the area where my sister lives (today the towering chimneys and cavernous brick factories are just for tourists, your new dress was made in an anonymous single story building on an industrial estate) might be driven more by people not needing to own new frocks every week when they've been no further than their kitchen in a month than because it would actually be illegal to staff their business - if nobody's buying what you make then suddenly it makes sense to take a handout from the government and actually shut rather than pretend making mauve turtleneck sweaters or whatever is "essential".
For these major routes, there are typically at least bi-weekly voyages scheduled, so for this kind of distance, you can expect about 11 days pretty uniformly distributed +-2 days, if you pay to get on the next ship.
This may lead to (committing to) paying for the spot on the ship when your pallet is ready for pickup at the factory, not when it arrives at the port) and use low-delay overland trucking services. Which operate e.g. in lockstep with the port processing to get your pallet on the move within half a day of the container being unloaded from the ship, ideally having containers pre-sorted at the origin to match truck routes at the destination. So they can go on a trailer directly from the ship and rotate drivers on the delivery tour, spending only a few minutes at each drop-off.
Because those can't rely on customers to be there and get you unloaded in less than 5 minutes, they need locations they can unload at with on-board equipment. They'd notify the customer with a GPS-based ETA display, so the customer can be ready and immediately move the delivery inside. Rely on 360-degree "dashcam" coverage and encourage the customer to have the drop-off point under video surveillance, just to easily handle potential disputes. Have the delivery person use some suitable high-res camera with a built-in light to get some full-surface-coverage photographic evidence of the condition it was delivered in.
I'd guess with a hydraulic lift on the trailer's back and some kind of folding manual pallet jack stuck on that (fold-up) lift, so they drive up to the location, unlock the pallet jack, un-fold the lift, lower the lift almost to the ground, detach the pallet jack to drop it the last inch/few cm to the ground, pull the jack out, lower the lift the rest of the way, drive it on to the lift, open the container, get up with the pallet jack, drive the pallets (one-by-one) for this drop-off out of the container and leave them on the ground, close and lock the container, re-arm the jack's hooks, shove it jack back under the slightly-lowered folding lift, make it hook back in, fold it up, lock the hooking mechanism (against theft at a rest stop (short meal and toilet breaks exist, but showering can be delayed for the up to 2 nights)), fold it all the way up, and go on to drive to their next drop-off point.
So yes, getting insurance can be a good idea to offset some losses you may have, as long as they are somewhat limited compared to your companies overall assets and income. But as soon as the insurance payout matches a significant part of your net worth, the insurance might not save you.
As such, it makes sense to make the level of risk you plan to accept (by not being insured against it and not mitigating) a conscious economic decision rather than pretending you've covered everything.
The big example of this which springs to mind is business interruption cover - it's ruinously expensive so it's extremely unusual to have the max cover the market might be prepared to offer. It's a pure economic decision.
Usually you'd have to show your homework (offers from insurance companies proving that it really is unaffordable). I totally get the trade-off, and the fact that if the business could not exist if it was properly insured that plenty of companies will simply take their chances.
We also both know that in case something like that does go wrong everybody will be looking for a scapegoat, so for the CEO's own protection it is quite important to play such things by the book, on the off chance the risk one day does materialize.
This has killed quite a few otherwise very viable companies, it is fine to take risks as long as you do so consciously and with full approval of all stakeholders (or at least: a majority of all stakeholders). Interesting effects can result: a smaller investor may demand indemnification, then one by one the others also want that indemnification and ultimately the decision is made that the risk is unacceptable anyway (I've seen this play out), other variations are that one shareholder ends up being bought out because they have a different risk appetite than the others.
Let's take an example which might lead to bankruptcy. A typical answer to a major disaster (let's say your main and sole building burning as a typical case) for an SME would be to cease activity, furlough employes and stop or defer every payments you can while you claim insurance and assess your options. Well, none of these things are obvious to do especially if all your archive and documents just burnt. If you think about it (which you should), you will quickly realise that you at least need a way to contact all your employes, your bank and your counsel (which would most likely be the accountant certifying your results rather than a lawyer if you are an SME in my country) offsite. That's the heart of disaster planning: having solutions at the ready for what was easy to foresee so you can better focus on what wasn't.
Bankruptcy when dealt with correctly is a process not an end.
If everything else fail it's better to fill for bankruptcy when there is still something to recover with help of others than to burn everything to ashes because of your vanity.
At least that's how I understood parent's comment.
Yes, that's why "Go bankrupt" is not a plan which was the entire point of my reply. That's like saying that your disaster recovery plan is "solve the disaster".
There are two types of bankruptcies in the US used most often by insolvent businesses: Chapter 7, and Chapter 11.
A Chapter 7 bankruptcy is what most people in other countries think of when they hear "bankruptcy" - it's the total dissolution of a business and liquidiation of its assets to satisfy its creditors. A business does not survive a Chapter 7. This is often referred to as a "bankruptcy" or "liquidation" in other countries.
A Chapter 11 bankruptcy, on the other hand, is a process by which a business is given court protection from its creditors and allowed to restructure. If the creditors are satisfied with the reorganisation plan (which may include agreeing to change the terms of outstanding debts), the business emerges from Chapter 11 protection and is allowed to continue operating. Otherwise, if an agreement can't be reached, the business may end up in Chapter 7 and get liquidated. Most countries have an equivalent to a Chapter 11, but the name for it varies widely. For example, Canada calls it a "Division 1 Proposal," Australia and the UK call it "administation," and Ireland calls it "examinership."
Since there's a lot of international visitors to HN I just thought I'd jump in and provide a bit of clarity so we can all ensure we're using the same definition of "bankruptcy." A US Chapter 7 bankruptcy is not a plan, it's the game over state. A US Chapter 11 bankruptcy, on the other hand, can definitely be a strategic maneuver when you're in serious trouble, so it can be part of the plan (hopefully far down the list).
I wondered why you should plan an event which will "destroy" your company anyways.
Yes it is. (Though it's better, as GP suggested, as a final layer of a plan and not the only layer.)
> Becoming insolvent might be the end result of a situation but it's not going to help you deal with it.
Insolvency isn't bankruptcy. Becoming insolvent is a consequence, sure. Bankruptcy absolutely does help you deal with that impact, that's rather the point of it.
Documentation / SOPs that have been tested thoroughly by various team members are really important. It helps work out any kinks in interpretation, syntax errors etc.
It does feel a little ridiculous at the time for all the effort involved, but incidents like this show why it's so important.
Our goal and the reason we have a lot of stuff backed up on-prem is to have our most time-critical operations back up within a couple of hours - unless the building is destroyed, in which case that's a moot point and we'll take what we can get.
A dev wiped our almost-monolithic sales/manufacturing/billing/etc MySQL database a month or two ago. (I have been repeatedly overruled on the topic of taking access to prod away from devs) We were down for around an hour. Most of that time was spent pulling gigs of data out of the binlog without also wiping it all again. Because our nightly backups had failed a couple weeks prior - after our most recent monthly "glance at it".
https://www.dna.fr/amp/faits-divers-justice/2021/03/10/stras...
You can see more details here: https://baxtel.com/data-center/ovh-strasbourg-campus
Container DCs were a big thing for a while. Even Google did a whole PR thing about how they used them.
https://twitter.com/xgarreau/status/1369559995491172354
Looks glowing red to me.
Fires in DCs aren't rare at all, I know of at least three, one of those in a building where I had servers. This one seems to be worse than the other two. Datacenters tend to concentrate a lot of flammable stuff, throws a ton of current through them and does so 24x7. The risk of a fire is definitely not imaginary, which is why most DCs have fire suppression mechanisms. Whether those work as advertised depends on the nature of the fire. An exploding on prem transformer took out a good chunk of EV1's datacenter in the early 2000's, and it wasn't so much the fire that caused problems for their customers, but the fact that someone got injured (or even died, I don't recall exactly), and before the investigation was completed and the DC released to the owners again took a long time.
Being paranoid and having off-site backups is what allowed us to be back online before the fire was out. If not for that I don't know if our company would have survived.
It's not the best in terms of Disaster Recovery Plan but we accept that level of risk.
My employer has a single rack of servers at HQ. It's positioned at a very specific angle with an AC unit facing it, their exact positions are marked out on the floor in tape. The servers contain VMs that most employees work on, our git repository, issue trackers, and probably customer admin as well. They say they do off-site backups, but honestly, when (not if) that thing goes it'll be a pretty serious impact on the business. They don't like people keeping their code on their take-home laptop either (I can't fathom how my colleagues work and how they can stand working in a large codebase using barebones vim over ssh), but I've employed some professional disobedience there.
I use S3 API compatible object storage platforms for remote backup. E.g. BackBlaze B2. I wrote about my backup scripts for FreeNAS (jail that runs s3cmd to copy files to B2) here: https://www.shogan.co.uk/cloud-2/cheap-s3-cloud-backup-with-...
For Kuberenetes I use velero which can be configured with an S3 storage backend target: https://www.shogan.co.uk/kubernetes/kubernetes-backup-on-ras...
I personally have almost all of the software running in containers with an orchestrator on top (Docker Swarm in my case, others may also use Nomad, Kubernetes or something else). That way, rescheduling services on different nodes becomes less of a hassle in case of any one of them failing, since i know what should be running and what configuration i expect it to have, as well as what data needs to be persisted.
At the moment i'm using Time4VPS ( affiliate link: https://www.time4vps.com/?affid=5294 ) for the stuff that needs decent availability and because they're cheaper than almost all of the alternatives i've looked at (DigitalOcean, Vultr, Scaleway, AWS, Azure) and that matters to me.
Now, in case the entire data centre disappears, all of my data would still be available on a few HDDs under my desk (which are then replicated to other HDDs with rsync locally), given that i use BackupPC for incremental scheduled backups with rsync: https://backuppc.github.io/backuppc/
For simplicity, the containers also use bind mounts, so all of the data is readable directly from the file system, for example, under /docker (not really following some of the *nix file system layout practices, but this works for me because it's really easy to tell where the data that i want is).
I actually had to migrate over to a new node a while back, took around 30 minutes in total (updating DNS records included). Ansible can also really help with configuring new nodes. I'm not saying that my setup would work for most people or even anything past startups, but it seems sufficient for my homelab/VPS needs.
My conclusions:
- containers are pretty useful for reproducing software across servers
- knowing exactly which data you want to preserve (such as /var/lib/postgresql/data/pgdata) is also pretty useful, even though a lot of software doesn't really play nicely with the idea
- backups and incremental backups are pretty doable even without relying on a particular platform's offerings, BackupPC is more than competent and buying HDDs is far more cost effective than renting that space
- automatic failover (both DNS and moving the data to a new node) seems complicated, as does using distributed file systems; those are probably useful but far beyond what i actually want to spend time on in my homelab
- you should still check your backups* All my services are dockerized and have gitlab pipelines to deploy on a kubernetes cluster (RKE/K3s/baremetal-k8s)
* git repo's containing the build scripts/pipelines are replicated on my gitlab instance and multiple work computers (laptop & desktop)
* Data and databases are regularly dumped and stored in S3 and my home server
* Most of the infrastructure setup (AWS/DO/Azure, installing kubernetes) is in Terraform git repositories. And a bit of Ansible for some older projects.
Because of the above, if anything happens all I need to restore a service is a fresh blank VM/dedicated machine or a cloud account with a hosted Kubernetes offering. From there it's just configuring terraform/ansible variables with the new hosts and executing the scripts.
They both run postfix + dovecot, so mail is synced via dovecot replication. Data is rsync-ed daily, and everything has ZFS snapshots. MySQL is not set into replication - my home internet breaks often enough to have serious issues, so instead I drop everything every day import a full dump from the main server, and do a local dump as backup on both sides.
I don't have automatic failover set up.
But when you already use ZFS you can do a very speedy full backup with:
mysql << EOF
FLUSH TABLES WITH READ LOCK;
system zfs snapshot data/db@snapname
UNLOCK TABLES;
EOF
Turns out, our disaster recovery plan is pretty good.
Datacenter burned down and I still was up 4 hours later in another data center with zero data loss. Good times.
I'm not sure how normal this strategy is outside of container land but I like just using scripts, they are simple and transparent - if you take time and care to write them well.
They basically install and configure packages using sed or heredocs with a few user prompts here and there for setting up domains etc.
If you are constantly tweaking stuff this might not suit you, but if you know what you need and only occasionally do light changes (which you must ensure the scripts reflect) then this could be an option for you.
It does take some care to write reliable clear bash scripts, and there are some critical choices like `set -e` so that you can walk away and have it hit the end and know that it didn't just error in the middle without you noticing.
Rolling backups with a month retention to box using rsync.
It creates a network drive to box by default when I boot my desktop.
I have some scripts for putting production db's in test and when I went them locally.
As an immediate plan, the 2-3 business critical systems are replicating their primary storages to systems in a different datacenter. This allows us to kick off the configuration management in a disaster, and we need something in between 1-4 hours to setup the necessary application servers and middlewares to get critical production running again.
Regarding backups, backups are archived daily to 2 different borg repo hosts on different cloud providers. We could lose an entire hoster to shenanigans and the damage would be limited to ~2 days of data loss at worst. Later this year, we're also considering to export some of these archives to our sister team, so they can place a monthly or weekly backup on tape in a safe in order to have a proper offline backup.
Regarding restores - there are daily automated restore tests for our prod databases, which are then used for a bunch of other tests after anonymization. On top, we've built most database handling on top of the backup/restore infra in order to force us to test these restores during normal business processes.
As I keep saying, installing a database is not hard. Making backups also isn't hard. Ensuring you can restore backups, and ensuring you are not losing backups almost regardless of what happens... that's hard and expensive.
You do have backup servers.
I was affected, my personal VPS is safe but down and other VPS I was managing I don't know anything about. I have the backups and right now I'd love for them to just set me up a new VPS so I can restore the backups and restore the services.
But i have a friend who potentially lost important uni work hosted on his nextcloud instance... On SBG2.
A rough reminder that backups are really important, even if you are just an individual
Prayer and hope, usually.
> Printer flammability
Eerie timing: do they possibly suspect some bad cables?
[0]: http://travaux.ovh.net/?do=details&id=49016
[1]: http://travaux.ovh.net/?do=details&id=49017
Why not? Cables with ratings lower than the load they are carrying is a prime cause for electrical fires. If the load is too high for long enough, the shielding melts away, and if it is close enough for other material to catch fire then that's the ball game. It's a common cause for home electrical fires. Some lamp with poor wiring catches the drapes on fire, etc. Wouldn't think a data center would have flammable curtains though.
http://travaux.ovh.net/?do=details&id=47840 earliest one that I found was back in December
https://www.dna.fr/faits-divers-justice/2021/03/10/strasbour...
Edit:
Video at https://www.youtube.com/watch?v=a9jL_THG58U
Satellite view of the site on Google Maps https://goo.gl/maps/L2T6YNFCtiyDdiNv7
it was an interesting proof of concept, but finding the right people to maintain the infra with both IT and Scuba skills was a narrow niche to nail down ;)
I used to work for a provider whose 2 main datacenters of 8k+ sq ft could pull all oxygen out of the building in 60 seconds.
In the case of a fire, it would dump a lot of argonite gas in and consume a large amount of the oxygen in the room, depriving the fire of fuel. It's also safe and leaves minimal clean-up work afterwards, doesn't harm electronics etc. unlike sprinklers and the like.
The amount of oxygen left is sufficient for human life, but not for fires, though my understanding is that it can be quite unpleasant when it happens. You won't want to hang around.
Flames are extinguished by this cooling effect (identical to water in that regard), but humans rely on catalytic processes that aren't affected by the cooling effect.
If you could keep the existing oxygen inside, while adding Novec 1230, humans could continue to breathe while the flames would still be extinguished, but this would require the building/room to be a pressure chamber that holds about half an atmosphere of extra pressure. I'm pretty sure just blowing in some extra oxygen with the Novec 1230 would be far cheaper to do safely and reliably.
I mean, in principle, if you gave re-breathers to the workers and have some airlocks, you could afford to keep that atmosphere permanently, but it'd have to be a bit warm (~30 C I'd guess). Don't worry, the air would be breathable, but long-term it'd probably be unhealthy to breathe in such high concentrations and humans breathing would slightly pollute the atmosphere (CO2 can't stay if it's supposed to remain breathable).
Just to be clear: in an effective argonite extinguishing system you'd have about a minute or two until you pass out and need to be dragged out, ideally get oxygen, get ventilated (no brain, no breathing) and potentially also be resuscitated (the heart stops shortly after your brain from a lack of oxygen, so if you're ventilated fast enough, it never stops and you wake up a few externally-forced breaths later). Having an oxygen bottle to supplement your breaths would fix that problem for as long as it's not empty.
At this point I feel like it would be cheaper just to not have workers go there. Fill the place completely full of nitrogen with an onsite nitrogen generator (and only 1atm pressure). Have 100% of regular maintenance and as much irregular maintenance as possible be done by robots. If something happens that requires strength/dexterity beyond the robots (e.g. a heavy object falling over), either have humans go in in some form of scuba gear, or if you can work around it just don't fix it.
For exit, basically just the reverse, with the pump pumping the air around the person to wherever the person came from.
The general issue with unbreathable atmospheres is that a failure in their SCBA gear easily kills them.
And re-breathers that are only there so you don't have to scrub the atmosphere in the room as often shouldn't be particularly expensive. You may even get away with just putting a CO2 scrubber on the exhaust path, and giving them slightly oxygen-enriched bottled air so you can keep e.g. a 50:50 oxygen:nitrogen ratio inside (so e.g. 20% O2, 20% N2, 60% Novec 1230). And it doesn't even need to be particularly effective, as you can breathe in quite a bit of the ambient air without being harmed, and the environment can tolerate some of your CO2. Like, as long as it scrubs half of your exhausted CO2 it won't even feel stuffy in there (you could handle the ambient air you'd have without re-breathers being used, as it'd be just 1.6% CO2, but you'd almost immediately get a headache).
They'd have an exhaust vent for pressure equalization, which would chill the air to condense and re-cycle the Novec 1230. For pressure equalization in the other direction, they'd probably just boil off some of that recycled Novec 1230.
So yeah, re-breather not needed, if you just get a mouth+nose mask to breathe bottled 50:50 oxygen:nitrogen mix. That 50% oxygen limit (actually 500 mBar) is due to long-term toxicity, btw. Prolonged exposure to higher levels causes lung scarring and myopia/retina detachment, so not really fun.
...then I read in some other comments here that they used wood in the interior construction.
It takes a good while to start burning, and even when significantly charred it still retains most of its strength.
The issue is you'll have lots of plastic (cabling) in a DC, and plastic will burn
I can appreciate an honest mistake though, like the status page server cron is hosted in the same cluster that caught fire and hence it burnt down and can't update the page anymore.
> The whole site has been isolated, which impacts all our services on SBG1, SBG2, SBG3 and SBG4. If your production is in Strasbourg, we recommend to activate your Disaster Recovery Plan
What more could you want?
What's the point of a status page then if it does not show you the status? I don't want to be chasing down twitter handles and support pages during an outage.
I wonder if a server fire would cause Amazon to go to status red. So far anything and everything has fallen under yellow.
At best you might be missing out on some SLA refunds, but at worst it could be disasterous for a business. I've been on the wrong side of a update-by-hand status system from a hosting provider before and it wasn't fun.
> Unable to display this content to due missing consent. By law, we are required to ask your consent to show the content that is normally displayed here.
I've usually seen this with embedded videos rather than comments.
Agreed, though. A fake status page is worse than no status page. I don't mind if the status page states that it's manually updated every few hours as long as it's honest. But don't make it look like it's automated when it's not.
My gut says yes.
I think this is probably linked to a manual reporting system and they got bigger fish to fry at the moment than updating this status page.
https://news.ycombinator.com/item?id=15661218
They are not the only ones though. All too common. Well, it's tricky to set this up properly. The only proper way would be to use external infra for the status page.
So nobody chooses to make an honest status page.
But, they're usually manual affairs because sometimes the system is broken even when the healthcheck looks ok, and sometimes writing the healthcheck is tricky, and always you want the status page disconnected from the rest of the system as much as possible.
It is a challenge to get 'update the status page' into the runbook. Especially for runbooks you don't review often (like the one for the building is on fire, probably).
Luckily my status page was not quite public; we could show a note when people were trying to write a customer service email in the app; if you forget to update that, you get more email, but nobody posts the system is down and the status page says everything is ok.
No traffic whatsoever between sbg-g1 and sbg-g2 ant their peers.
If they don't have any servers anymore, how can they be down ;)
Probably the best they could do at the moment.
"Update 5:20pm. Everybody is safe. Fire destroyed SBG2. A part of SBG1 is destroyed. Firefighters are protecting SBG3. no impact SBG4."
“SBG1, the first Strasbourg data center, consisting of twelve containers, came online in 2012. The 12 containers had a capacity of 12,000 servers.
SBG2 is a non-container data center in 2016 using its “Tower” design with a capacity of 30,000 servers.
SBG3 tower was built in 2017 with a capacity of 30,000 servers.
SBG4 was built in 2013 as several containers to augment capacity, but was decommissioned in 2018 and moved to SB3”
> OVH to Disassemble Container Data Centers after Epic Outage in Europe
> “This is probably the worst-case scenario that could have happened to us.”
> OVH [...] is planning to shut down and disassemble two of the three data centers on its campus in Strasbourg, France, following a power outage that brought down the entire campus Friday, causing prolonged disruption to customer applications that lasted throughout the day and well into the evening.
https://www.datacenterknowledge.com/uptime/ovh-disassemble-c...
This is crazy.
SBG2 was HUGE and if this isn't a translation error on the part of Octave (which I could understand given the stress and ESL) I have a hard time fathoming what kind of fire could destroy a whole facility with nearly 1000 racks of equipment spread out across separated halls.
I'm really hoping "destroyed" here means "we lost all power and network core and there's smoke/fire/physical damage to SOME of that"
I can't even fathom a worst-case scenario of a transformer explosion (which does occur and I've seen the aftermath of) having this big of an impact. Datacenters are built to contain and mitigate these kinds of issues. Fire breaks, dry-pipe sprinkler systems and fire-extinguishing gas systems are all designed to prevent a fire from becoming large-scale.
Really glad nobody was hurt. OVH is gonna have a bad time cleaning all this up.
The couple datacenters I've been inside were small, old and used halon gas which wasn't supposed to destroy the machines. No idea how it works in big places these days.
1. https://www.theregister.com/2018/04/26/decibels_destroy_disk...
Also, fire suppression system do fail. There was an infamous incident in LA for one of the studios. They built a warehouse to be a tape vault with tapes going back to the 80s. A fire started, but the suppression system failed because there was not enough pressure in the system. Total loss. Got to keep your safety equipment properly tested!
I can find surprisingly few English-language resources for this (only Dutch ones); guess it's a combination of a small country + before the internet really took off.
Losing parts of history like that is tragic.
Want to test the system? Just turn off that water valve, hook up your air compressor and pressure test to your heart’s content.
EDIT a picture from earlier this night https://pbs.twimg.com/media/EwGqV17XMAMF_wa?format=jpg&name=...
https://www.ovh.com/world/solutions/centres-de-donnees.xml
and here
"Every data center room is fitted with a fire detection and extinction system, as well as fire doors. OVHcloud complies with the APSAD R4 rule for the installation of mobile and portable fire extinguishers and has the N4 certificate of conformity for all our data centers." https://us.ovhcloud.com/about/company/security
Source in French: https://lafibre.info/ovh-datacenter/ovh-et-la-protection-inc...
I think the real problem, per that post, is this:
>> They are simple sprinklers that spray with water. It has nothing to do with very high pressure misting systems, where water evaporates instantly, and which save servers. Here, it's watering, and all the servers are dead. It's designed like that. Astonishing, isn't it?
>> Obviously, they rely above all on humans to extinguish a possible fire, unlike all conventional data centers.
(all thanks to Google Translate)
This strikes me as a terrible safety system because even if a human managed to detect the fire, they have to make a big call: is the risk of flooding the facility and destroying a ton of gear worth putting out a fire? By the time the human decides "yes, it is", it may well be too late for the sprinklers.
> and at least sprinklers don't accidentally kill the technicians.
Not a real risk with modern 1:1 argon:nitrogen systems – the goal is to pump in inert gases and reduce oxygen content to around 13%, a point where the fire is suppressed and people can survive. You wouldn't want to be in a room breathing 13% oxygen for a long time, but it won't kill you.
All in all, it looks like this was a "normal accident"[1] for a hosting company that aggressively competes on price. The data center was therefore built with less expensive safeties and carried a higher risk of catastrophic failure.
I find it very hard to believe that that would pass code anywhere in the US/EU or most of the world. They may not have had sprinklers but that doesn't mean there isn't fire suppression.
https://lafibre.info/ovh-datacenter/ovh-et-la-protection-inc...
I retract all my prior surprise that this place burned down.
These days, fire suppression systems need to be non-lethal, so inert gasses are out. Water is too, for obvious reasons. Last I checked, they flooded the inside of the DC with airborne powder that coats everything (especially the inside of anything with a running fan). Once that deploys, the machines in the data center are a write off even if the fire was minor.
> Finally, we have other photos of the floor of the OVH "Roubaix 4" tower. It is clearly wood! Hope it's fireproof wood! A wooden datacenter ... is still original, we must admit.
> In France, data centers are mainly regulated by the labor code, by ICPE recommendations (with authorization or declaration) and by insurers. At the purely regulatory level, the only things that are required are:
> - Mechanical or natural smoke extraction for blind premises or those covering more than 300m2
> - The fire compartmentalization beyond a certain volume / m2
> - Emergency exits accessible with a certain width
> - Ventilation giving a minimum of fresh air per occupant
> - Access to the firefighter from the facade for the premises are the low floor of the last level is more than 8 meters
> - 1 toilet for 10 people (occupying a position considered "fixed")
https://lafibre.info/ovh-datacenter/ovh-et-la-protection-inc...
This puts an image on the sentence "SBG2 is destroyed". Do not expect any recovery from SBG2.
http://status.ovh.com/vms/index_sbg2.html
Am I looking at the wrong thing or am I right to wonder why we still bother with public status pages if it never shows the real status?
Edit: nvm just saw another comment pointing out the same further down the thread (I randomly came across this page while looking for the physical location of another DC)
> Due to a fire at one of our data centres, a few of our servers are down and may be down permanently. We are restoring these servers from backups and will enable puzzles and storm as soon as possible. > > We hope that everyone who is dealing with the fire is safe, including the firefighters and everyone at OVH. <3.
I didn't know what OVH was before 10 minutes ago, but this seems really impactful. I hope everyone there is safe and that the immediate disaster gets resolved quickly.
e.g. They have a 4core, 16gb ram server for $22/mo which is 25% of what my preferred provider, Linode, charges.
Now, it comes with older consumer hardware (that one is a sandy bridge i5), and about as much support as the price tag suggests, as well as a dated management interface, but when I used to run a modded minecraft server as a college student, which needed excessive amounts of RAM and could be easily upset by load spikes on other clients, then it was a no-brainer, even if I would expect the modern-ish Xeons Linode uses to win on a core for core basis.
This is the interface they had when I used them last:
It is still a mess of separate accounts, but you can use email addr to log in instead of random-numbers generated handle.
The OVHCloud US is completely separated for legal purposes, from what I remember. No account sharing, different staff, OVH EU cannot help you at all with US accounts.
Naturally, I didn't do regular backups. Most of the data I have from a 2019 backup, but there's a metadata database that I don't have a copy of, and will need to reconstruct from scratch. Thankfully for my case reconstructing will be possible - I know that's not the case for everyone.
Right now I'm feeling pretty stupid, but I only have myself to blame. For their part, OVH have been really good at keeping everyone updated (particularly the Founder and CEO, Octave Klaba).
I believe that when I signed up back in 2017, the Strasbourg location was advertised as one of the cheapest, so I can imagine a lot of people with a ~$4 / month OVH VPS are in the same situation, desperately scrambling to find a backup.
(For those that have a OVH VPS that's down right now, you can find what location it is in by logging onto the OVH control panel.)
[1] https://twitter.com/olesovhcom/status/1369598998441558016
I can remember several San Diego fires that threatened the original JohnCompanies datacenter[1] circa mid 2000s and thinking about all of the assets and invested time and care that went into every rack in the facility.
Very interested to read the post-mortem here ... even more interested in any actionable takeaways from what is a very rare event ...
[1] Castle Access, as it was known, at the intersection of Aero Dr. and the 15 ... was later bought by Redit, then Kio ...
Yeeeeeahhhhhhhhhh!!!!!
Interestingly in this case, having a very compact data center was probably an aggravating factor. This shows how complex these technical choices are, you have to think of operating savings, with a trade off on the gravity of black swan events...
They are so close to each other that they are basically the same physical datacenter with 4 logical partitions.
I don‘t get why people don‘t do offsite backup today as it‘s basically for free. AWS Glacier basically costs nothing at all.
A fire detection system using eg. lasers and Inergen(or Argonite) for putting the fire out is commonly used in datacenters. The gas fills the room and reduces the amount of oxygen in the room so most fires are put out within a minute.
The cool thing is that the gas is designed to be used in rooms with people, so that is can be triggered any time. It is however quite loud, and some setups have been known to be too loud, even destroying harddrives.
* overall high energy density (lots of current flowing everywhere)
* the batteries for backup power are dangerous, and can easily(ish) overheat when activated.
Maybe there is a concentration of oxygen that is high enough for humans to survive, yet too low to sustain combustion?
Yes, supposedly (12% ish?). Can't say I'd be thrilled at the idea of testing it.
Source: I've been in an FM-200 discharge
Inergen consists of 52 % nitrogen, 40 % argon and 8 % carbon dioxide.
Carbon dioxide might sound strange, but it causes the heart to beat faster to compensate for the lowered amounts of oxygen.
The whole point of Inergen is to quickly put out fires where water/foam/powder isn't usable and the room might contain people.
It's cool stuff, and I was under the impression that basically all datacenters used it.
So loud that it destroys hard drives... That's scary, are people's eardrums much more resistant?
But Inergen (and other gas) is more or less useless if you allow it to escape too quickly. So the cooling system should be a fairly closed circut.
Edit: I'm also a Norwegian dude. :)
I can't wait for the conspiracy theories about how the fire is a "cover up" to "hide" bad COVID-19 numbers...
[1] https://twitter.com/GuillaumeRozier/status/13695724905996902...
[3] https://www.meteo-covid.com/trouillocarte (Just wanted to share the "trouillocarte" - which roughly translates to "'how badly is shit hitting the fan today' map" ;) )
- https://www.usine-digitale.fr/article/le-health-data-hub-heb...
- https://www.genethique.org/la-cnam-refuse-le-transfert-globa...
- https://france3-regions.francetvinfo.fr/bretagne/donnees-med...
Here is more plausible conspiracy theory: USA/microsoft is behind the cyberattacks and the fire
https://mobile.twitter.com/abonin_DNA/status/136953802824345...
Not sure how good were the fire suppression systems of the building.
They only mention smoke detection: https://www.ovh.com/world/us/about-us/datacenters.xml
Anti-fire systems didn't kick-in. The reason? For couple of days the system was detecting a little bit of smoke from one of the devices. Operators weren't able to pinpoint exact location, considered it a false alarm and manually switched anti-fire system off.
Ouf.
https://mobile.twitter.com/olesovhcom/status/136953578757072...
I don't really know how FM200 systems work in data centres, but I'm guessing that if the fire didn't start from within the actual server room, FM200 might not save you? e.g. if a fire started elsewhere and went out of control, it would be able to burn through the walls/ceiling/floor of the server room, in which case no amount of FM200 gas can save you, right?
Another possibility, of course, is that the FM200 system simply failed to trigger even though the fire started from within the server room.
There is no published investigation details about this incident yet, I believe. Can somebody chime in about past incidents where FM200 failed to save the day?
FM200 is something else that is often used in modern builds (not just datacenters).
https://lafibre.info/ovh-datacenter/ovh-et-la-protection-inc...
It doesn't really surprise me that the machines are dead, but the whole place being destroyed is much more surreal.
https://www.dna.fr/faits-divers-justice/2021/03/10/strasbour...
Tout le monde est sain et sauf. Le feu a détruit SBG2. Une partie de SBG1 est détruite. Les pompiers protègent actuellement SBG3. Pas d’impact sur SBG4 », a tweeté Octave Klaba, le fondateur d’OVHcloud, en désignant les différentes parties du site. « Tous nos clients sur ce centre de données sont susceptibles d’être impactés », a précisé l’entreprise sur Twitter.
If your primary is set up with credentials to automatically transfer a copy to the backup destination over the network, what happens if your primary gets pwned and the access is used to encrypt or delete the backup?
Secondly, test doing restores of your backups, and have methods/procedures in place for exactly what a restore looks like.
Append-only permissions. We do this in S3 for that specific reason. S3 lifecycle rules take care of pruning old backups.
You can also build a pull-based system where auth resides on the backup system, not the production system.
If they have a good insurance I'm confident this will have little impact on their operations, I really hope they do. I host a few components on OVH/SoYouStart dedicated servers, luckily not mission critical, but still had rather good experience with them, especially in terms of price to performance.
The timing is so bad that it becomes almost suspicious. When is the best time to sabotage your competitor? When they are the most visible.
Honest question: which publicity damage?
A fire in a datacenter is very much part of the things you should expect to see happen when you operate a large number of datacenters and will obviously cause some disruption to your customers hosting physical servers there.
Provided the disruption doesn't significantly extend to their cloud customers and doesn't affect people paying for guaranteed availability (which it shouldn't - OVH operates datacenters throughout the world), this seems to me to be an unfortunate incident but not a business threatening one.
But they seem to be back up
Companies want quick money, they push people to skip important IT operations, like disaster recovery plans.
And backups are the least monitored systems.
It's all connected.
https://twitter.com/olesovhcom/status/1369535787570724864?s=...
"Update 7:20am Fire is over. Firefighters continue to cool the buildings with the water. We don’t have the access to the site. That is why SBG1, SBG3, SBG4 won’t be restarted today."
Is the situation "Halon is legal in Australia" or "Halon isn't actually illegal per se if you don't use a lot of it"?
Other gases like argon or FM200 are not Halons.
https://i.imgur.com/epj1Lue.png
Translation :
We lost our Gitlab and backups...
And the automatic backups that had been put in place no longer worked so a priori we lost everything...
Ouch https://twitter.com/Onepamopa/status/1369484420982407173
Double check those backups, folks.
If it's business critical should you even trust at all?
Nothing fancy, but it's the kind of bread and butter intuition you need to avoid walking straight off a cliff.
He also let me sit at the back writing a compiler instead of learning VB.Net, top dude
What percentage of organizations have these?
There is NO evidence that is what happened here, and I don't think OVH allows customers to bring their own equipment making even less likely. Still I wait and hope to hear a root cause from this one.
> Firefighters immediately intervened to protect our teams and prevent the spread of the fire. At 2:54 am they isolated the site and closed off its perimeter.
> By 4:09 am, the fire had destroyed SBG2 and continued to present risks to the nearby datacenters until the fire brigade brought the fire under control.
> From 5:30 am, the site has been unavailable to our teams for obvious security reasons, under the direction of the prefecture. The fire is now contained.
How many of us here plan Fire Drills within our teams and larger organizations?
I have backup snapshots but its stored in ovh itself :( hoping for a miracle!
Hopefully you had the code in GitHub but that still leaves the DB. It looks like your has something to do with command line or Linux lessons so not sure how much user data is critical? Maybe you can get this up and running again to some extent.
So far I haven't got any communication from OVH alerting me about this. I think that's the first thing they should do, alerting their customers that something it's happening.
Anyway I was running a service that I was about to close, so this may be it. I do have a recovery plan, but I don't know it it is worth it at this point.
I'm never using OVH again. The fire can happen, but don't ask me about my recovery plan, what about yours?
It says they've alerted customers, but I expect some have been missed, through inaccurate email records, email hosted on the systems that have been destroyed etc.
Containers would seem to be at a disadvantage when it comes to dissipating, rather than containing, heat. I hope improved thermal management and fire suppression designs can be implemented.
As another comment says, that density probably exacerbated the issue here.
Anyone know of any casualties?
My impression is that they tried very hard to maintain uptime, which was probably a bad idea when we see the extent of the damages. This VPS just hosts external facing services and is easy to set back up.
The 2nd worst thing, IF you happen to catch it soon and control it, is that the temp raise triggers many alerts and automatic controls.
So when it's controlled, it's still a real nightmare. Here, firemen could not even control it...
The only good thing is that my backup was offsite.
Does OVH offer automatic snapshots for VPS? I know Hetzner does it for 20% addidation of the cost of the server. If they do the next question would be if they are destroyed too?
Here's more information on them: https://docs.ovh.com/gb/en/vps/using-automated-backups-on-a-...
With OVH the price can depending on the situation double your cost, if you've got a 3€ VPS and a 3€ backup configuration (as the price depends on size)
Many customers don't need a backup, so it's up to each customer to arrange their own backups — perhaps with tools and services provided by the hosting company, or their own solution.
Running a company with no backup (for cost or any other reason) is very risky, as some people will have found out today.
Also, if you’re hosting clients’ static websites, you were burning your money, there are way cheaper options out there (and fully managed).
Does anyone recommend a mail provider that implements IMAP failover and replication across different sites ?
My mail account is hosted by OVH, it has been down for hours and from what I read I may have to wait for another 1 or 2 days.
Thanks !
You can see from pictures of the OVH fire aftermath that the outer plates have melted holes in them characteristic of metal.
I have used them for years and I don't believe they've ever said anything like deploy in both SBG1 and SBG2 for safety or availability, because you don't get that choice.
When you provision a machine (eg via API) they tell you "SBG 5 min, LON 5 min, BHS 72h" and you pick SBG and get assigned first-available. There is no "I want to be in SBG4" generally.
Local as in your home/office. While your application may run in AWS or whatever remote server, its necessary to have copies of your data that you can physically touch and access.
One main deployment, one remote backup and one onsite physically accessible backup.
nvm not this dc!
scenario 1: site a is down plan: recover to site b by most convinent means
scenario 2: site b is down plan: restore services, operate without redundancy out of site a
scenario 3: site c is down plan: restore services, catch up later. continue operating out of site a
scenario 4: site b and c down plan: restore services, operate without redundancy out of site a
scenario 5: site a and b down plan: cross fingers, restore to new site from cold storage on expensive cloud VM instances
scenario 6: data archive corrupted ransomware plan: restore from air gapped data archive, hope ransomware was identified within 90 days
scenario 7: site b and c down, then site a down plan: quit
scenario 8: staff hates job and all quit plan: outsource
scenario 9: and so on...
Wait for the misinformation hose to dry up, and decide in a few weeks.
After this sorry episode, I dont think any CTO or CIO of any public company will be able to even consider using the other guys.
edit: I am not implying that we put all eggs in one basket with no failover and dr. I am implying the big cos will pay 2x premium on infrastructure to project reliability.
Also, cost matters if the AWS bill is one of the company's top expenses.
What?
Or you could use t4g.2xlarge, which is cheaper. There's no situation where OVH is 3x cheaper (I mean maybe if bandwidth is your thing, but IDK).
AWS a1.metal: $293 per month for 16 ARM cores, 16GB memory
OVH Rise-4 [1]: $131 per month for 16 Intel Xeon cores, 128GB memory, 500Mbit/s
And the ARM CPUs on the AWS instances are nowhere near the intel cores here, so maybe this configuration is comparable:
OVH Rise-APAC [2]: $57 per month for 4 Intel Xeon cores, 32GB memory
They do have APAC, Canadian, and US data centers as well.
Largest hosting provider in Europe, probably top10 or top5 in the world.
The laugh track and the writing is honestly dated even by the standards of Dads Army.
Comedies that came later ditched the laugh track. They had to work harder to get viewers at home to laugh, but ultimately a bunch of them (starting with The UK Office) hold up much better as a result.
(Having said that, I think "Fire" was a memorable episode that is still amusing. The 0118911881999119 song, "it's off, so I'll turn it on... AND JUST WALK AWAY".)
It might have been ahead of its time. Silicon Valley was well received and is as nerdy and intricately detailed as Season 1 of the IT Crowd. "Normal people" thought it was far out and zany. People that work in tech have been to all those meetings. And, a major character was named PG!
That sounds like a pretty realistic allegory of the last two decades in Free Software (or software in general, or the web...)
Keep in mind I’m talking not of “downtime” but of actual data loss which might affect business continuity.
This is really tragic. I’m hoping they have some kind of multi regional backup/replication and not just multi zones (although from the twitters it appears that only one of the zones was destroyed however the others don’t seem to be operational atm).
They don't automatically geo-replicate things. You still need a backup for the torched EC2 instance to be able to relaunch in another AZ/region.
"The whole site has been isolated, which impacts all our services on SBG1, SBG2, SBG3 and SBG4. "
Wonder if those SBGx were advertised as being the same as "Availability Zones" - when other cloud providers ensure zones are distanced enough from each other (~1km at least) to likely survive events such as fire.
https://cdn.baxtel.com/data-center/ovh-strasbourg-campus/pho...
SBG is for Strasbourg. That's not a region. It's a city. Obviously, SBG1 to 4 are in the radius of one fire. It's four different buildings on the same site.
When I say “heroic efforts”, that’s in contrast to “ordinary error recovery and failover”, which is the way you’d want to handle a DC fire, because DC fires happen often enough.
The thing is, while these big companies have a much larger base of expertise to draw on and simply more staff time to throw at problems, there are factors which incentivize these employees to *increase risk* rather than reduce it.
These big companies put pressure on all their engineers to figure out ways to drive down costs. So, while a big cloud provider won’t make a rookie mistake—they won’t forget to run disaster recovery drills, they won’t forget to make backups and run test restores—they *will* do a bunch of calculations to figure out how close to disaster they can run in order to save money. The real disaster will then reveal some false, hidden assumption in their error recovery models.
Or in other words, the big companies solve all the easy problems and then create new, hard problems.
What we call “progress” in humanity is just putting our best efforts into reducing or eliminating the problems we know how to solve without realizing the problems they may create further down the line. The only way to know for sure is to try it, see how it goes, and then re-evaluate later.
California had issues with many forest fires. They put out all fires. Turns out, that solution creates a bigger problem down the line with humongous uncontrollable fires which would not have happened if the smaller fires had not been put out so frequently. Oops.
Sure, the amount of expertise, redundancy and breadth of service offerings they provide is worth a markup, but they are also significantly more expensive than they need to be.
Thanks to being the leader in an oligopoly, and due to patterns like making network egress unjustifiably expensive to keep you (/your data) from leaving.
AWS may charge more for egress, but that’s not high enough for it to be a concern for most clients.
A bigger, independent concern is probably that there should be sufficient redundancy, backups and such that allows for business continuity. (Note again that I’m not saying that all companies make full use of these features, but those that care for such things do. Additionally, I’ve honestly never heard of an AWS DC burning down. Either it doesn’t happen frequently or it doesn’t have enough of an effect on regular customers, both of situations are equivalent for my case).
Most businesses choose to prioritize the second aspect. Even if they have to pay extra for egress sometimes, it’s just not big enough of a concern as compared to businesses continuity.
Put another way: one is guided towards making operational good choices rather than being left to discover them yourself. This is a value proposition of public clouds since it commoditises that specialist knowledge.
I use a single server at OVH, and I'm not in the affected DC, but if this DID happen to me I could get back up and running fairly quickly. All our data is mirrored on S3 and off site backups are made frequently enough it wouldn't be an issue.
Plus, you still need to plan for a scenario like this even with AWS or any other cloud provider. It is less likely to happen with those, given the redundancy, but there is still a chance you lose it all without a backup plan.
Personally I'll continue to use these third world cloud providers. But I like to live on the edge.
Their solar panels on the roof caught on fire. Presumably there was no actual disruption in service computing-wise.
> Google fire
2006, still a young company at that point. Also this was way before GCP where they were responsible for other people's services.
> AWS fire
Data center caught on fire while it was under construction. Drunk construction workers messed up, not AWS. Presumably no computers were even inside yet. This is completely different from OVH completely burning down while all their customers' data went up in flames.
If these are the best examples of Big Cloud messing up, this is quite the endorsement.
It's funny how the OVH CEO tells everyone to go activate their DRPs while the company didn't have enough foresight to install fire suppression systems.
As you can see below, a developer expert is only making 24K-63K Euro per year at OVH (in US dollar it's almost the same amount):
https://www.glassdoor.com/Salary/OVHcloud-France-Salaries-EI...
after paying taxes you will only get a half of that amount.