Sometimes it's completely acceptable that a server will run for 10 years with say 1 week or 1 month of downtime spread over those 10 years, yes. That's the sort of uptime you can see with single servers that are rarely changed and over-provisioned as many on Hetzner are. Some examples:
Small businesses where the website is not core to operations and is more of a shop-front or brochure for their business.
Hobby websites too don't really matter if they go down for short periods of time occasionally.
Many forums and blogs just aren't very important too and downtime is no big deal.
There are a lot of these websites, and they are at the lower end of the market for obvious reasons, but probably the majority of websites in fact, the long tail of low-traffic websites.
Not everything has to be high availability and if you do want that, these providers usually provide load balancers etc too. I think people forget here sometimes that there is a huge range in hosting from squarespace to cheap shared hosting to more expensive self-hosted and provisioned clouds like AWS.
But I do agree the poster should think about this. I don't think it's 'off' or misleading, they just haven't encountered a hardware error before. If they had one on this single box with 30 databases and 34 Nginx sites it would probably be a bad time, and yes they should think about that a bit more perhaps.
They describe a db follower for cutover for example but could also have one for backups, plus rolling backups offsite somewhere (perhaps they do and it just didn't make it into this article). That would reduce risk a lot. Then of course they could put all the servers on several boxes behind a load-balancer.
But perhaps if the services aren't really critical it's not worth spending money on that, depends partly what these services/apps are.
Could I take it down for the afternoon? Sure. Or could I wait and do it after hours? Also sure. But would I rather not have to deal with complaints from users that day and still go home by 5pm? Of course!
Also, in general, you can architect your application to be more friendly to migration. It used to be a normal thing to think about and plan for.
VMware has a conversion tool that converts bare metal into images.
One could image, then do regular snapshots, maybe centralize a database being accessed.
Sometimes it's possible to create a migration script that you run over and over to the new environment for each additional step.
Others can put a backup server in between to not put a load on the drive.
Digital Ocean makes it impossible to download your disk image backups which is a grave sin they can never be forgiven for. They used to have some amount of it.
Still, a few commands can back up the running server to an image, and stream it remotely to another server, which in turn can be updated to become bootable.
This is the tip of the iceberg in the number of tasks that can be done.
Someone with experience can even instruct LLMs to do it and build it, and someone skilled with LLMs could probably work to uncover the steps and strategies for their particular use case.
This is a general response to it.
I have run hosting on bare metal for millions of users a day. Tens of thousdands of concurrent connections. It can scale way up by doing the same thing you do in a cloud, provision more resources.
For "downtime" you do the same thing with metal, as you do with digital ocean, just get a second server and have them failover.
You can run hypervisors to split and manage a metal server just like Digital Ocean. Except you're not vulnerable to shared memory and cpu exploits on shared hosting like Digital Ocean. When Intel CPU or memory flaws or kernel exploits come out like they have, one VM user can read the memory and data of all the other processes belonging to other users.
Both Digital Ocean, and IaaS/PaaS are still running similar linux technologies to do the failover. There are tools that even handle it automatically, like Proxmox. This level of production grade fail over and simplicity was point and click, 10 years ago. Except no one's kept up with it.
The cloud is convenient. Convenience can make anyone comfortable. Comfort always costs way more.
It's relatively trivial to put the same web app on a metal server, with a hypervisor/IaaS/Paas behind the same Cloudflare to access "scale".
Digital Ocean and Cloud providers run on metal servers just like Hetzner.
The software to manage it all is becoming more and more trivial.
> This level of production grade fail over and simplicity was point and click, 10 years ago.
While some of the tools are _designed_ for point and click, they don't always work. Mostly because of bugs.
We run Ceph clusters under our product, and have seen a fair share of non-recoveries after temporary connection loss [1], kernel crashes [2], performance degradations on many small files, and so on.
Similarly, we run HA postgres (Stolon), and found bugs in its Go error checking cause failure to recover from crashes and full-disk conditions [3] [4]. This week, we found that full-disk situations will not necessarily trigger failovers. We also found that if DB connections are exhausted, the dameon that's supposed to trigger postgres failover cannot connect to do that (currently testing the fix).
I believe that most of these things will be more figured out with hosted cloud solutions.
I agree that self-hosting HA with open-source software is the way to. These softwares are good, and the more people use them, the less bugs they will have.
But I wouldn't call it "trivial".
If you have large data, it is also brutally cheaper; we could hire 10 full-time sysadmins for the cost of hosting on AWS, vs doing our own Hetzner HA with Free Software, and we only need ~0.2 sysadmins. And it still has higher uptime than AWS.
It is true that Proxomox is easy to setup and operate. For many people it will probably work well for a long time. But when things aren't working, it's not so easy anymore.
[1]: "Ceph does not recover from 5 minute network outage because OSDs exit with code 0" - https://tracker.ceph.com/issues/73136
[2]: "Kernel null pointer derefecence during kernel mount fsync on Linux 5.15" - https://tracker.ceph.com/issues/53819
[3]: https://github.com/sorintlab/stolon/issues/359#issuecomment-...
Even if Amazon was down, if I was planning to buy, I'd wait. heck, I got a bunch of crap in my cart right now I haven't finished out.
Intentional downtime lets everyone plan around it, reduces costs by not needing N layers of marginal utility which are all fragile and prone to weird failures at times you don't intend.
> Intentional downtime lets everyone plan around it, reduces costs by not needing N layers of marginal utility which are all fragile and prone to weird failures at times you don't intend.
Quite frankly, I would manage if things were run "on-supply" with solar and would just go dark at night.
That's a strawman version of what happens.
There have been times when I've tried to visit a webshop to buy something but the site was broken or down, so I gave up and went to Amazon and bought an alternative.
I've also experienced multiple business situations where one of our services went down at an inconvenient time, a VP or CEO got upset, and they mandated that we migrate away from that service even if alternatives cost more.
If you think of your customers or visitors as perfectly loyal with infinite patience then downtime is not a problem.
> Unless you are Amazon and every minute costs you bazillions, you are likely gonna get the better deal not worrying about availability and scalability. That 250€/m root server is a behemoth. Complete overkill for most anything.
You don't need every minute of downtime to cost "bazillions" to justify a little redundancy. If you're spending 250 euros/month on a server, spending a little more to get a load balancer and a pair of servers isn't going to change your spend materially. Having two medium size servers behind a load balancer isn't usually much more expensive than having one oversized server handling it all.
There are additional benefits to having the load balancer set up for future migrations, or to scale up if you get an unexpected traffic spike. If you get a big traffic spike on a single server and it goes over capacity you're stuck. If you have a load balancer and a pair of servers you can easily start a 3rd or 4th to take the extra traffic.
Great. So how much did the webshop lose in that hour of maintenance (which realistically would be in the middle of the night for their main audience) and how much would they have paid for redundancy? Also a bit hard to believe you repeatedly ran into the situation of an item sold at a self-hosted webshop and Amazon alike. Are you sure they haven't just messed up the web dev biz? You could totally do that with AWS too...
> If you're spending 250 euros/month on a server, spending a little more to get a load balancer and a pair of servers isn't going to change your spend materially.
Of course, but that's not the argument. It's implied you can just double the 250€/m server for redundancy, as you would still get an offer at the fraction of cloud prices. But really that server needs no more optimization in terms of hardware diversification. As I said, it's complete overkill. Blogs and forums could easily be run on a 30€/m recycled machine.
Spot on! People still go to Chick-fil-A, even if they are closed on Sundays!
The confusing part about this article is the emphasis on a zero-downtime migration toward a service that isn't really ideal for uptime. It wouldn't be that expensive to add a little bit of architecture on the Hetzner side to help with this. I guess if you're doing a migration and you're paid salary or your time is free-ish, doing the migration in a zero downtime way is smart. It's a little funny to see the emphasis on zero downtime juxtaposed to the architecture they chose where uptime depends on nothing ever failing
Clever architecture will always beat cleverly trying to pick only one cloud.
Being cloud agnostic is best.
This means setting up a private cloud.
Hosted servers, and managed servers are perfectly capable of near zero downtime. this is because it's the same equipment (or often more consumer grade) that the "cloud" works on and plans for even more failure.
Digital Ocean definitely does not guarantee zero downtime. That's a lot of 9's.
It's simple to run well established tools like Proxmox on bare metal that will do everything Digital Ocean promises, and it's not susceptible to attacks, or exploits where the shared memory and CPU usage will leak what customers believe is their private VPS.
Nothing ever failing in the case of a tool like Proxmox is, install it on two servers, one VPS exists on both nodes (you connect both servers as nodes), click high availability, and it's generally up and running. Put cloudflare in front of it like the best preference practices of today.
If you're curious about this, there's some pretty eye opening and short videos on Proxmox available on Youtube that are hard to unsee.
When you have 2 nodes running, both are mirrored and running, one can have hardware break.
Also, hardware can provide failure notifications before it breaks, and experience teaches to just update and upgrade before hard drives break.
Since tools like proxmox just add a node, you add new hardware, mark the VM for that node to mirror, and it is taken care of.
Terraform etc can sit below Proxmox and alleviate what you're speaking about:
Some examples: https://www.youtube.com/watch?v=dvyeoDBUtsU
Look from my perspective, I'll got flying color from my owner because of the cost saving and got my team morale up that I really need them to maintain the system instead of lay them off.
Also in some cases that also mean new jobs opening.
Also, don't underestimate the reliability of simplicity.
I was a Linux sysadmin for many years, and I have never seen as much downtime from simpler systems as I routinely see from the more complicated setups. Somewhere between theory and reality, simpler systems just comes out ahead most of the time.
Usually those articles describe two situations:
- they were "on the cloud" for the wrong reasons and migrating to something more physical is the right approach
- they were "on the cloud" for the right reasons and migrating to something more physical is going to be a disaster
As a bonus, Hetzner is European.
Dealing with over engineered bullshit, that behaved in strange ways that disrupted the service was far more often a problem.
So, yes, redundancy is something that can be left away, if you're comfortable to be responsible for fixing things at a Saturday morning.
They saved money and lost nothing.
Now, if they so wish, they could use a portion of that to increase redundancy - but that wasn't the point of the article.
If someone starts thinking about redundancy and load balancers than DO's solution is rent a second similar sized droplet, and then add their load balancing service. If you do those things with Hetzner instead, you would still be spending less than you did with Digital Ocean.
Personally, what is keeping me on DO is that no single droplet I have is large enough to justify moving on its own, and I'm not prepared to deal with moving everything.
Given the downtimes we saw in the past year(s) (AWS, Cloudflare, Azure - the later even down several times), I would argue moving to any of the big cloud providers give you not much of a better guarantee.
I myself am a Hetzner customer with a dedicated vServer, meaning it is a shared virtual server but with dedicated CPUs (read: still oversubscribed, but some performance guarantee) and had zero hardware-based downtime for years [0]. I would guess their vservers are on similar redundant hardware where the failing components can be hotswapped.
[0] = They once within the last 3 years sent me an email that they had to update a router that would affect network connectivity for the vServer, but the notification came weeks in advance and lasted about 15 minutes. No reboot/hardware failure on my vServer though.
It’s amusing that the US government can shutdown for days/weeks/months over budget reasons and there’s no adult discussions that take place about fixing the cause. Yet the latest HN demo that 100 people will use need all 9’s reliability and hundreds of responses.
Not a bad tradeoff for 99.8% of shops out there.
I know people like FAANG LARPing. Not everyone has budget or need to run four nines with 24/7 and FAANG level traffic.
If you can tolerate few hours of downtime and some data rollback/loss, single server + robust backups can be viable strategy
If your scaling need is not that high, you can get very far with a single server
Like, I know Leetcode tells otherwise, but most companies really don't need full FAANG stack with 99.999% uptime. A day of outage in a few years isn't going affect bottom lines.
LARPing as a FAANG is waste of money and lots of companies doesn’t even need 3 nines let alone 5ve.
People underestimate how far you can go with one or two servers. In fact, what I have seen in ky career is many examples of services that should have been running on one or two servers and instead went for a hugely complex microserviced approach, all in on Cloud providers and crazy requirements of reliability for a scale that never would come.
Deploying a new docker instance or just restoring the app from a snapshot and restoring the latest db in most cases is enough.