> But also, google spent a mountain of money advertising chrome.
That money was also used to increase the user base via drive-by installations, e. g. while installing Adobe Reader you had to deselect the Chrome installation, otherwise you'd find yourself with a new standard browser afterwards.
In theory even a single company service could be distributed, so only a fraction of websites would be affected, thus it's not a necessity to be a single point of failure. So I still don't like this argument "you see what happens when over half of the internet relies on Cloudflare". And yes, I'm writing this as a Cloudflare user whose blog is now down because of this. Cloudflare is still convenient and accessible for many people, no wonder why it's so popular.
But, yeah, it's still a horrible outage, much worse than the Amazon one.
The "omg centralized infra" cries after every such event kind of misses the point. Hosting with smaller companies (shared, vps, dedi, colo whatever) will likely result in far worse downtimes, individually.
Ofc the bigger perception issue here is many services going out at the same time, but why would (most) providers care if their annual downtime does or doesn't coincide with others? Their overall reliability is no better or worse had only their service gone down.
All of this can change ofc if this becomes a regular thing, the absolute hours of downtime does matter.
I think you're being overly dramatic. In practice I've seen complexity (which HA setups often introduce) causing downtimes far more often than a service being hosted only on a single instance.
You'll have planned downtime just for upgrading MongoDB version or rebooting the instance. I don't think that this is sth you'd want to have. Running MongoDB in a replica set is really easy and much easier than running postgres or MySQL in an HA setup.
No need for SREs. Just add 2 more Hetzner servers.
The sad part of that is that 3 Hetzner servers are still less than 20% of the price of equivalent AWS resources. This was already pretty bad when AWS started, but now it's reaching truly ridiculous proportions.
from the "Serverborse": i7-7700 with 64GB ram and 500G disk.
37.5 euros/month
This is ~8 vcpus + 64GB ram + 512G disk.
585 USD/month
It gets a lot worse if you include any non-negligible internet traffic. How many machines before for your company a team of SREs is worth it? I think it's actually dropped to 100.
Sure, I am not against Hetzner, it's great. I just find that running sth in HA mode is important for any service that is vital to customers. I am not saying that you need HA for a website. Also, I run many applications NOT in HA mode but those are single customer applications where it's totally fine to do maintenance at night or on the weekend. But for SaaS this is probably not a very good idea.
Yes, any time someone says "I'm going to make a thing more reliable by adding more things to it" I either want to buy them a copy of Normal Accidents or hit them over the head with mine.
How bad are the effects of an interruption for you? Google has servers running every day, but you with one server can afford to gamble on it, since it probably won't fail for years - no matter the hardware though, keep a backup, because data loss is permanent. Would you lose millions of dollars a minute, or would you just have to send an email to customers saying "oops"?
Risk management is a normal part of business - every business does it. Typically the risk is not brought down all the way to zero, but to an acceptable level. The milk truck may crash and the grocery store will be out of milk that day - they don't send three trucks and use a quorum.
If you want to guarantee above-normal uptime, feel free, but it costs you. Google has servers failing every day just because they have so many, but you are not Google and you most likely won't experience a hardware failure for years. You should have a backup because data loss is permanent, but you might not need redundancy for your online systems. Depending on what your business does.
HA can be hard to get right, sure, but you have to at least have (TESTED) plan for what happens
"Run a script to deploy new node and load last backup" can be enough, but then you have to plan on what to tell customers when last few hours of their data is gone
Theoretically it has slightly better memory bandwidth, (you are supposed to get) the Nvidia AI software ecosystem support out of the box, and you can use the 200G NIC to stick 2 together more efficiently.
Practically, if the goal is 100% about AI and cloud isn't an option for some reason, both options are likely "a great way to waste a couple grand trying to save a couple grand" as you'd get 7x the performance and likely still feel it's a bit slow on larger models using an RTX Pro 6000. I say this as a Ryzen AI Max+ 395 owner, though I got mine because it's the closest thing to an x86 Apple Silicon laptop one can get at the moment.
Because the ML ecosystem is more mature on the NVidia side. Software-wise the cuda platform is more advanced. It will be hard for AMD to catch up. It is good to see competition tho.
But the article shows that the Nvidia ecosystem isn't that mature either on the DGX Spark with ARM64. I wonder if Nvidia is still ahead for such use cases, all things considered.
On the DGX Spark, yes. On ARM64, Nvidia has been shipping drivers for years now. The rest of the Linux ecosystem is going to be the problem, most distros and projects don't have anywhere near the incentive Nvidia does to treat ARM like a first-class citizen.
The complete Framework Desktop with everything working (including said Ryzen AI Max 395+ and 128 GB of RAM) is 2500 EUR. In Europe the DGX Spark listings are at 4000+ EUR.
It's a different animal. Ryzen wins on memory bandwidth and has 'AI' accelerator (my guess matrix multiplication). Spark has times lower bandwidth, but much better and more generic compute. Add to that CUDA ecosystem with libs and tools. I'm not saying Ryzen is bad, actually it's great Mac substitute for poor man. $2K for 128GB version on Amazon now.
Online only at https://frame.work AFAIK. I don't think people shelling out 2-4k for an AI training machine are concerned whether or not they can find it at a hardware store locally or online, but I may be wrong.
The vast majority of Ryzen AI Max+ 395s (by volume at least) are sold as complete system offerings as well. About as far as you can go the other way is getting one without an SSD, as the MB+RAM+CPU are an "all or nothing" bundle anyways.
Fortunately, AMD upstreams its changes so no custom distro is required for Strix Halo boxes. The DGX is the platform more at risk of being left behind on Linux - just like Jetson before it, which also had a custom, now-abandoned distro.
Needing a customized spin of Ubuntu to have working video drivers is an Nvidia thing. One can also choose a Windows option, if they like, and run AI from there as it's just a standard x86 PC. That might actually be the best option for those worried about pre-installed OSs for AI tinkering.
The userspace side is where AI is difficult with AMD. Almost all of the community is build around Nvidia tooling first, others second (if it all).
Does Romc=ROCm, or something else? If the former, ROCm is just a userspace compute library for the in-kernel amdgpu driver. The "kernels" it runs are GPU compute programs, not customized Linux kernels.
1. Proxmox cannot even join a network using DHCP requiring manual IP configuration.
2. Disk encryption is a hell instead of checkbox in installer
3. Wi-Fi - no luck (rarely used for real servers, but frequently for r&d racks)
Of course, it is a Debian core underneath and a lot of things are possible given enough time and persistence, but other solutions have them out of the box.
Proxmox wasn’t considered because of the audience (leadership) and Proxmox’s perceived market (SMBs/homelabs). I couldn’t even get them to take Virtuozzo seriously, so Proxmox was entirely a non-starter, unfortunately.
FWIW, I use Proxmox at home. It’s a bit obtuse at times, but it runs like a champ on my N100-based NUCs.
I believe you're referring to our January 2020 Kafka incident where a logic bug caused data loss for a customer. It was a serious failure and a huge learning moment for us.
The platform we operate today is fundamentally different and far more resilient than it was five years ago. We've scaled significantly (recently passing $100M ARR) because we took those early lessons seriously and continue to prioritize reliability.
They are pretty strict with account sign-ups because they are so cheap, which attracts abuse. Unfortunately this prevents some regular sign-ups too. You could try Netcup, I've also had a good experience with them.
Terraform/OpenTofu is good for infrastructure but it becomes a pain if you use it for k8s deployments. I suggest using GitOps (ArgoCD / flux) for everything inside the cluster and OpenTofu for the cluster itself.
Downside with that is having DNS, managed databases, cloud storage (s3) outside K8s - When deployments are also managed with tf, it's easy to connect a deployment in K8s to resources outside K8s (via tf state).
You can still do that by reading the values you need directly from the state file. I suggest that you define outputs and access them through the state file from other projects, otherwise the external dependency will be hidden.
It depends on your use case what the best option would be. You can put variables into kustomize templates and replace them via envsubst. For helm charts you can just supply the values for the variables during the install / upgrade.
Another possibility would be to create a ConfigMap with the values you read from the state file.
> Using the internet in the UK/EU is such a horrible experience, every cookie pop-up is a reminder how badly thought out these rules are.
Technical cookies don't require any consent so every time you see a cookie banner the website owner wants to gather more data about you than necessary. Furthermore, these rules don't require cookie banners, it's what the industry has chosen as the way to get consent to track their users.
So when I see a tracking cookie dialog on a web site, either 1. the site collects more data than they need to in order to run the site or 2. they don't and the site's management is incompetent. Both are pretty good reasons to avoid that particular web site.
There's no risk, they know what they are doing because the law doesn't just mandate the banner, it mandates you to know which third party service you're sharing the data to.
Check the banner next time, you'll see how many “partners” they do sell your data to.
A lot of websites for smaller businesses will not be run by technical people, they'll be run by business people or otherwise who don't understand cookies beyond "I see cookie banners on every website I visit, therefore to avoid legal trouble I need one too", you can't expect someone like that to understand the difference between tracking cookies and technical cookies.
Ah yes of course. How could I forget about poor Mom & Pop Co. and their 186 business partners that they want to share my personal data with. Surely we can't expect such a small operation to know what they are doing.
That's not the point I'm making, I'm saying whoever in Mom & Pop Co. set up the website may well not understand the difference between the cookie types and even if they are using no tracking cookies and sharing no data, they may well put a cookie notice on their website anyway as they're so common they think they're normal, the law allows for huge fines, and they're doing it out of an abundance of caution.
That money was also used to increase the user base via drive-by installations, e. g. while installing Adobe Reader you had to deselect the Chrome installation, otherwise you'd find yourself with a new standard browser afterwards.
reply