I forgot to mention - stay far away from SSDNodes. Their prices look good, but they massively oversell their RAM. They use ballooning (virtio_balloon) to overallocate RAM, and require you to use their customized Linux distro images that have it enabled. Ballooning means that they take unused RAM from your VM and lend it to other VMs to use. If you try to use all your RAM, they shut down your VPS.

GreenCloud, HostHatch, and RackNerd all let you install whatever OS you want, by mounting an ISO, connecting to the VPS via VNC, and going through the normal setup process just like you would on a regular computer. They do have pre-built images, but I always prefer to install from my own ISO so I know that the installation doesn’t have any non-standard modifications.

Thankfully, most VPS providers don’t oversell RAM or disk space.

My guess would be renting a dedicated server, or a home server. It can end up cheaper than colo.

self host?

Hosting at home, renting a dedicated server, or colocating are all considered self-hosting.

They’re reliable, but overpriced for what you get, especially after the Akamai acquisition. Akamai are not known for affordability :)

I use GreenCloudVPS for a bunch of things. They used to have VPSes starting at $15/year for 2GB RAM and 20GB disk, but their current budget deals are still good (starts at $25/year for 4GB RAM and 35GB disk). These are on older hardware (mostly AMD EPYC Rome Zen 2 systems, ~5 years old) which is why they haven’t been affected by increased RAM and storage prices yet. https://greencloudvps.com/billing/store/budget-kvm-sale

They also have “premium” plans which are newer (AMD EPYC Milan Zen 3 or Genoa Zen 4, and PCIe 4.0 NVMe storage in RAID10): https://greencloudvps.com/billing/store/premium-kvm-sale

HostHatch is good too. Their regular pricing aims to compete with Hetzner, but in 13 locations worldwide. https://hosthatch.com/products. They do sometimes have sales on LowEndTalk with pricing similar to GreenCloud’s “premium” plans.

RackNerd sometimes have good deals, although they’ve discontined quite a few of them recently. You can use https://racknerdtracker.com/ to see currently active deals. I still see a 2GB RAM + 35GB storage deal for $35/year for example.

I suspect all the very cheap ones (<$50/year) will eventually increase pricing. GreenCloud and HostHatch own all their hardware. Right now they use older hardware that’s already reached break-even, but all hardware needs to be replaced eventually, and the replacement cost for things like enterprise SSDs is a lot higher than it was in the past.

If you want to stick to big names, OVHcloud and Hetzner both have pretty good VPS pricing. OVH runs theirs on very old hardware like Haswell (2014/2015 era) Xeon E5 CPUs which is the main reason why they have cheaper pricing that Hetzner.

There’s also AWS Lightsail that’s essentially Amazon’s version of a VPS. Storage and data transfer are included in the monthly price, unlike EC2 where it costs extra. I’d only use it if you like AWS or want the comfort of using a very big company though - performance is notably worse than a provider like GreenCloudVPS or Hetzner.

I’m not affiliated with any of these companies. This is just my personal experience :)

I wish more modern TPUs were available for purchase. Google’s TPU v8i is very powerful, but you can only rent it via Google Cloud. (that and they’re most useful in big clusters, and the pricing for such a cluster is definitely out of reach for anyone that’s not running their own data center)

For small options, there’s mini PCs with the AMD Ryzen AI Max+ 395, and the Nvidia DGX. Their memory bandwidth is quite low compared to a GPU though, so expect slower performance.

Do you want to run TensorFlow Lite / LiteRT models? PyTorch Mobile? TensorRT? onnx? YOLO? vLLM? Something else? The recommendations will vary based on your use case.

Google Coral was decent for TensorFlow Lite, but it’s EOL (end of life) now. I’ve got the dual TPU Mini PCIe version in my home server, via a PCIe adapter board. I use it for object detection with Blue Iris + CodeProject AI and it works pretty well for that use case.

Hailo-8 is supposed to be like a more powerful version of the Coral, but I don’t have experience with it. It supports a bunch of frameworks: TensorFlow, TensorFlow Lite, Keras, PyTorch and ONNX. I’d be interested in hearing other people’s thoughts on it.

I don’t know if any of these work over USB though. They’re usually internal devices. Google marketed the Coral USB as being for development and testing only, pointing people to the M.2 and PCIe versions for production usage.

As for something totally different… There’s the Nvidia Jetson single board computer which supports TensorRT, but I don’t have experience with it either. I also think it’s a bit older too. You could also consider getting a newer mini PC with a AMD Ryzen AI processor in it, or an Nvidia DGX Spark.

Google’s latest TPUs are only available in their cloud - they’re not selling the hardware to end users any more.

And I don’t ever know if it’ll get better because you need to know why you want to build something someway.

The major issue I’m seeing with junior (and even intermediate) developers is that they trust that the AI will always do things the correct way and don’t question its approach, and they don’t develop proper debugging skills and just rely on the AI to attempt it.

To get decent quality output out of an AI model, you need to have critical thinking skills, at least basic knowledge of the overall architecture for whatever you’re trying to build, and enough knowledge to question the model when it does something wrong.

Blindly trusting AI is why so many old security issues are coming back - stored/reflected XSS, SQL injection, exposing databases directly to the internet with no password, things like that. Newer frameworks mostly got rid of them, and now AI is bringing them back. It’s a fun time for red teams at least.

Does Patchmon not have a setting to look for the Docker socket in a different location?

I could be wrong but I don’t think there’s any security issues making a symlink to a socket, since permissions/ACLs on the socket would still apply.

My Epyc 7702 does have onboard TPM, but my supermicro H11DSi-NT doesn’t pass it through to the OS, for some reason

Huh… That’s interesting. At my workplace we have Linux EPYC servers with working TPM (it’s mandated that all computers, both clients and servers, must have TPM 2.0), but I’m not a hardware person and don’t know exactly how they’re configured.

This is good to know. I haven’t had issues with using a USB drive though, since it doesn’t receive many reads or writes - the system is copied to a RAM drive on boot and runs off that rather than the USB.

I assume this means I’d need another drive to boot it from? My current setup is that I have 2 x 22TB drives in a ZFS mirror for data storage, and 2 x 2TB NVMe SSDs in a ZFS mirror for things like VMs, Docker containers, documents, etc.

I had to get a Supermicro AOM-TPM-9665V TPM chip for my motherboard

How old is your CPU that it doesn’t have onboard TPM? It’s been a standard feature for quite a while now

Does it use http or MQTT?

Home Assistant uses HTTP for this. Realistically, you won’t see much difference between HTTP and MQTT for this use case.

MQTT is harder to secure than HTTP, and has some limitations (eg it normally only supports username and password auth - no SSO, no 2FA) so I’d avoid it for anything public-facing unless you have a specific reason to use it. Using it via a VPN is fine, but you’d still need to configure a separate MQTT username and password per user.

iptables should still work, but these days it gets converted to nftables so you may as well just learn nftables.

Having said that, I find it a pain to manually configure iptables or nftables. There might be a better way to do what you want.

The end goal is to have no reliance on tailscale as i am preparing for the eventual enshitification.

Tailscale is mostly open-source. If they do anything bad then someone could fork the project. The coordination server isn’t open-source, but you could self-host Headscale as a replacement.

If it still doesn’t suit your use cases, there’s some alternatives.

I personally wouldn’t directly deal with iptables or nftables rules, and instead use some other software to deal with that.

iptables is deprecated… If you really do want to do your own custom thing you should learn nftables.

All the data gathered by Cambridge Analytica was gathered through the public API though, after users had consented to share it (by logging into a quiz app that requested the permissions). That’s why the API is very locked down now, and the approval process to get any sort of data access is very strict.

The main issue was that they gathered data from people whose profiles were set to be visible only to friends. If someone logged into the quiz and granted permissions, their friends’ data was also accessible via the API.

aggressively guard

tbh it’s a hard balance for any social media company.

Guard content too little and you end up with Cambridge Analytica, which was literally because the public APIs allowed too much access (third-party apps could see any data through the API that you could see through your Facebook account, including friends profiles). You also end up with headlines talking about big data leaks which really just end up being compilations of public data (which has happened to both Facebook and LinkedIn).

Guard content too much and you restrict users’ freedom too much.

It’s not too bad if you use an outbound SMTP relay for sending. SMTP2Go is pretty good, and they have a free plan with 1000 emails per month. I use Mailcow and you can configure relays in their web UI, but it works just as well with the sender_dependent_relayhost_maps setting in Postfix.

Sure, it’s not fully self-hosted, but the interesting part to self-host is the storage of your emails, not the sending (which will just relay through other SMTP servers along the way anyways).

At least it’s open source so anyone can look at the code and figure out why it asks for the permissions.