I have a dedicated vm for things that are crucial to the home network, either latency-critical or network related.

That’d be my dns resolver (I enforce it over VLANs by hijacking anyone trying to do DNS to other resolvers, like random IoT devices), homebridge for less important home automaton and my own matter controller for most important home automaton (controlling the lights).

My router of choice is RouterOS in another VM. I tried opnsense, pfsense, vyatta, and a bunch of others (even a containerized Cisco route), and I settled on ROS, because it was the only one who could do IPv6 properly (apart from Cisco, but that has other issues).

For the less important things I run them on k8s and really, there are only two bits worth mentioning as essential: ArgoCD and nixhelm. Together, they provide effortless and mostly automated software updates with very easy rollbacks. I don’t have to go and manually update every single bit of software and that saves huge amounts of time.

I wonder if NixOS is a vacuum coffee maker for how confusing nix looks when you see it for the first time or instant coffee for how reproducible it is…

That’s just Slackware.

That’s a good point. Mind that in most production environments you’d be firewalled rather hard (especailly when it comes to logs processing which oftentimes ends up having PII). I wouldn’t trust any service that tries to use DoT or DoH in there that I couldn’t snoop on. Many deployments nowadays allow you to “punch” firewall holes based on the outgoing dns requests to an allowlisted domain, so chances are you actually want to use the glibc resolver and not try to be fancy.

That said, smaller images are always good in my book!

You’re nailing your goal then!

I would still steer you slightly towards documenting your architectural decisions more. It’s a good skill to have and will help you in a long run.

You have dozens of crate dependencies and only you know why they are in there. A high-level document on how your system interconnects and how the algorithms under the hood work will be a huge help to anyone who comes looking through your source code. We become better programmers not by reading the source code, but by understanding what it actually does.

Here’s a random trivia: your server depends on trust-dns-resolver. Why? Why wasn’t the stock resolver enough? Is that a design choice or you just wanted to have fun? There is no wrong answer but without the design notes it’s hard to figure your intent.

This looks nice, but there’s plenty free alternatives in this space which warrants a section in the readme with the comparison to other products.

You mention ram usage, but it’s oftentimes a product of event size. Based on your numbers, your average event size is about 800 bytes. Let’s call it 1kb. That’s one million events per day. It’s surely sounds more promising than Elastic, but not reaching Loki numbers, or, if you focus on efficiency, is way behind Victoriametrics Logs (based on peeking at their benches).

I think the important bits you need to add is how you store the logs (i.e. which indices you build) and what are your trade-offs. Grep is an efficient logs processor which barely uses any ram but incurs dramatic I/O costs, after all.

Enterprises will be looking at different numbers and they have lots of SaaS products to choose from. Homelab users are absolutely your target audience and you can have it by making a better UI than the alternative (victoriametrics logs aren’t that comfortable to work with) or making resource usage lower (people run k8s clusters on RPis, they sure wonder about every megabyte of ram lost) or making the deployment easier (fire and forget, and when you come to it, it works).

It sounds like lots of things and I don’t want to be discouraging. What you started there is really nice-looking. Good job!

You can enforce an always-on VPN (for at least ipsec) via an MDM profile. This kind of features isn’t found in the casual user setup options, but there’s plenty of knobs to tune in the enterprise profile configurator.

And yes, you can easily install that profile on your phone after.

It is pretty bad. After this thread I tried using Element X again only to learn that its “favorites” aren’t the same as Element’s “favorites” and more so you can’t set someone a favorite in E-X, at least not of your server is Conduit. It’s just silently ignored.

Local models are really good at tokenizing the text and figuring the intent in the user input. Not perfect, but much better than any possible regexps you can think of. And it’s a trivial operation you can run even on a CPU model.

I would absolutely recommend a file system with snapshot capabilities for a home server. One of btrfs mirror, dm-raid (raid5) with btrfs, or zfs would work. The practical differences would be negligible at this scale and you can just pick whatever you fancy.

I’ve been having sync issues with conduit lately, takes minutes for the mobile app to catch up. No way to purge old media, or to use something S3-compatible for its storage either.

Also, element x doesn’t support spaces, so if you want to bridge other chats into matrix they all are going to be messed up together.

I like matrix as a concept, but both servers and clients are in a bit of a shitshow state (same as xmpp was years ago).

The windows client does, yes. But I’ve found that to be fragile on occasions.

Technically, it does have a windows client. It’s just in various states of being broken.

For the last 10 days tailscale clocked 1% battery on my phone. I honestly didn’t even consider turning it off for battery savings.

If tailscale inside a container allows you to talk to it via “direct” connection and not a derp proxy, then it will offer you better service isolation (can set the tailscale ACLs for this specific service) without sacrificing performance.

Tailscale pushes for it because it just ties you in more. It allows to to utilize the ACLs better, to see your thing in their service mesh, and every service will count against the free node limit.

In practice, I often do both. E.g. I’ll have my http ingress exposed to tailscale and route a bunch of different services through it at a single tailscale node, where the access control is done by services individually. But I’ll also run a pod-to-pod tailscale between two k8s clusters because tailscale ACL is just convenient.

Updates to DNS, yes. Not necessarily to your primary zone. In other words, you don’t need access to the name servers for your highly privileged example.com zone, only the nameservers for inconsequential.example.com. With the challenge delegation you can easily narrow the scope by CNAMEing the relevant _acme-challenge enries in your primary domain once. This not only removes the need for the validator to modify your primary zone, but also scopes what subdomains it can validate, too. So the blast radius decreases.

I, too, maintain several devices that insist on having the certificates (and keys, yuck) being fed to them by hand. I automated it all, because I don’t see why a human should be in a loop of copying the secret material. Automaton is good.

How complicated is it to have a CNAME? /s

You can delegate to isolated nameservers with DNS-01, there’s no need to have control over the primary zone: https://www.eff.org/deeplinks/2018/02/technical-deep-dive-securing-automation-acme-dns-challenge-validation

I don’t think your question relates to the language as much as to the platform. The language of choice is somewhat irrelevant and what you care about is what actually happens under the hood.

For the likes of java and go you want to have some understanding of what runtime does for the memory allocations and how their GCs work. For python you sometimes end up in the spots where you need to understand what limitations the GIL imposes (even more important now that they are trying to get rid of it). When you run C (or C++ or Rust) on the embedded hardware it really helps to understand what exactly bit flipping does in specific registers and what DMA means for how you write your code.

You don’t really have to know it all. You can absolutely write code without caring about anything of that and I know plenty software engineers that do. Some people write amazing functional things in java without ever questioning what it does to the machines and what resources you need to run it.

If you start questioning it, that will only expand your understanding. It’s not a lateral move from e.g. C to Rust where you need to learn a lot to write your code in a memory-safe way, it’s a movement deeper into the stack and what you learn there will be applicable to any language you use for this stack.

Answering your question: I always feel bad about not understanding some low-level concept. I have stacks of MCU reference docs lying around, printed, highlighted; I have archives with sample code, and hand-annotated CMSIS reversing notes. Embedded world is hard because you can’t just know C and be done with it. You have to know the hardware, too.

Here’s my advice for you. Make notes of things that you learn from people smarter than you. Create a web of those notes and see where your gaps are. Then, work on learning something in those gaps in particular and see if you can make a blog post or something of your own. When you share what you learn you become one of those people with deep understanding that others look up to. There’s always someone struggling with something that you either know or know how to figure it out.

ECC is slightly more required for ZFS because its ARC is generally more aggressive than the usual linux caching subsystem. That said, it’s not a hard requirement. My curent NAS was converted from my old windows box (which apparently worked for years with bad ram). Zfs uncovered the problem in the first 2 days by reporting the (recoverable) data corruption in the pool. When I fixed the ram issue and hash-checked against the old backup all the data was good. So, effectively, ZFS uncovered memory corruption and remained resilient against it.