talosctl debug docker.io/library/alpine:3.24 gives you a shell in the host namespaces, with root mounted at /host. I think it works without kubelet, I suppose it might not help with a failed containerd though.
I'd love to know some details on why? I don't have any particular complaints about my k3s on Debian, but if there is a better option, I would love to know!
Re:Talos Linux
by sgt on Wednesday November 18, 2026 @04:24AM
(#50543943)
I do, still. It's often a lot cheaper and modern HW is very robust and reliable. But it comes with more responsibility, of course. Not happy with performance on typical cheap VMs.
Re:Talos Linux
by dada216 on Wednesday November 18, 2026 @04:28AM
(#50543978)
I do :)
alongside cloud of course, hybrid is the way to go in my view.
I run a bunch of openshift clusters for very private container workloads and VMs.
You can't beat a local fiber channel SAN for speed and reliability, especially compared to cloud software defined storage, and even on the networking side of things if you worked out your local connectivity well using real networking vs cloud providers software defined networks is so much better.
CSIs are hit and miss, some vendors are really really good, some are meh (hitachi, even though they are improving a lot with the latest relases), but the moment you get an IBM nvme san and start to tune all layers you get performance that are out of this world, some of our dbms workloads are very io sensitive and they improved 10x with nvme fabric.
With Kubernetes it makes a lot of sense to run bare metal. What advantage is there to running K8s nodes on a hypervisor? It's one more thing that requires maintenance and can go wrong.
Re:Talos Linux
by sgt on Wednesday November 18, 2026 @04:45AM
(#50544138)
And that's actually my concern with K8s still. It still has an unnecessarily high maintenance burden for projects that are limited in scope and few scalability requirements. Which can often still be an important project.
Often I find that a load balancer and a couple of docker compose file sis more rugged, because you'll be able to leave it for a year and it'll still work and easily be understood by just glancing at it.
Kubernetes will still work after a year too, it just won't be supported. I forgot about a K3s setup for well over a year...
If you don't need a cluster then a single node K3s is still better than Docker Compose. You might know how an ad hoc load balancer and Docker Compose works, but I wouldn't say this knowledge is any more given than knowing K8s these days.
But if you do need a cluster then maintaining a bunch of Talos nodes is easier than maintaining a bunch of full fat GNU/Linux boxen tbh.
Re:Talos Linux
by sgt on Wednesday November 18, 2026 @06:56AM
(#50545198)
I've never looked at K3s so this is intriguing. I find docker compose great but it lacks some things that add to the complexity if bolting on, e.g. secrets management, or even load balancing on the single node to achieve zero downtime deployments. Would k3s handle the proxy layer?
Re:Talos Linux
by r_lee on Wednesday November 18, 2026 @07:50AM
(#50545701)
IMO a lot of the "K8s is complex" is just the fact that you _can_ go complex with it, but you can really just declare pods and stuff and avoid the abstractions
I feel like k8s is better as soon as your thing goes from simple self-contained thing to multiple things running on one box
especially with agents it's never been easier to work with k8s.
from my experience, k3s supports basically everything k8s has that you'd need for a single-node setup
not sure what "Would k3s handle the proxy layer" means but if you mean the load-balancing part, then yes, that's like one of the main abstractions that k8s provides. but again, it provides many knobs so you probably just wanna pick the simplest option available
Re:Talos Linux
by r_lee on Wednesday November 18, 2026 @09:28AM
(#50546733)
you can do that, e.g. by using a pod exposed with a NodePort for it where it just uses the host IP/port you configure and forwards to pods (the internal network) using whatever config HAProxy needs
or use the Gateway/Ingress abstractions that basically do this automatically, where you just expose services via them, and they handle the routing
If you want to do L7 routing, aka virtual hosts, then K8s comes with support for that built in via the Ingress resource, or the new Gateway API. To realise it you need an ingress controller installed in your cluster. It's so common that K3s comes with one out of the box by default: Traefik. For a single node that's all you need.
For load balancing across a proper cluster then you can either provision an external load balancer (like keepalived) or you can look into something like MetalLB which runs entirely on-cluster (but you need BGP support for true load balancing).
In a cloud cluster (like AWS EKS) you get load balancers automatically just by provisioning a LoadBalancer type Service in Kubernetes.
Re:Talos Linux
by trollbridge on Wednesday November 18, 2026 @07:52AM
(#50545720)
Obviously, something is still being run on bare metal. One area of consulting for is right now moving people off of VMware who don’t feel like paying the rather stiff Broadcom licence fees.
Re:Talos Linux
by preisschild on Wednesday November 18, 2026 @05:47AM
(#50544599)
I run it on bare metal since i only see downsides from having another VM layer around it. Just forego kvm and use linux containers. You can also use Talos/Kubernetes as a VM hypervisor through stuff like kata containers.
My team ran on-premise with Talos for the last bunch of years. It avoided tons of odd issues I saw when I built k8s clusters on top of Ubuntu that I never quite ironed out.
It gave me a ton of appreciation for Kubernetes and Linux and how k8s is built on linux assumptions: the file system layout’s a little different in Talos to the default k8s assumptions, so there were some small adjustments I’d make every now and then.
Sidero’s professional support is out of the world, too. I had simple questions and got answers back from their devops engineers who are running clusters 100-1000 times the size of anything I had to touch, or from their software engineering team who write Talos.
Definitely worth a look if you’re hosting in your home lab, definitely worth a look if you’re doing self-hosted k8s for your job.
Re:Talos Linux
by raphinou on Wednesday November 18, 2026 @06:51AM
(#50545172)
Isn't k8s overkill for a home lab? I'm happily running my homelab on docker swarm, and don't miss anything from k8s. Tot the contrary, I think it would cost me some efforts only to convert compose files to deploy to kiss.
I'm asking because I haven't touched k8s for some years, maybe the situation has changed and I should re evaluate my choices?
Re:Talos Linux
by mschuster91 on Wednesday November 18, 2026 @04:28AM
(#50543981)
You'd think someone with a "talk to us" enterprise plan and a bunch of well known brands as customers would have at least one human copytexter and a designer on staff. Instead, a landing page reeking of Claudish in text and look.
You might want to check out what's on your network :-(. It works from here and the certificate was issued by LetsEncrypt in late October, expires late January. If you're seeing something different, you should look at what's between your browser and the site.
Re:Talos Linux
by filleokus on Wednesday November 18, 2026 @06:20AM
(#50544912)
Can't give enough praise to Talos. I run it both at home and at work. Depending on what you wanna play with in your home lab / hobby projects I can recommend their Omni Hobby Plan which was made free recently-ish to avoid having to host Omni.
Really looking forward to see their new "Director" product with the hypervisor support for running VM's as well. I hope the pricing is resonable enough that it is a feasible solution for companies I work with. Would be useful to use that for creating virtual k8s nodes, to slice the same physical hardware into more clusters.
(If Siderolabs are here, it would be awesome to be able to use Director as a hobbyist in the future as well - I think it's a great sales pipeline for you to allow enthusiasts play with it at home and pitch it to internal stakeholders)
Re:Talos Linux
by topf on Wednesday November 18, 2026 @07:28AM
(#50545494)
Running it on bare metal. Highly recommended for anyone who wants to own their infrastructure and reduce their exposure to the never-ending stream of security vulnerabilities.
Re:Talos Linux
by matesz on Wednesday November 18, 2026 @08:18AM
(#50545966)
Worth mentioning that Talos Linux features bit-for-bit reproducible builds and a full-source bootstrapped toolchain, which in my opinion should be featured everywhere. And sooner or later probably will.
For those of you who don’t know what that is I highly recommend to look it up.
In short build dag starts with (ai provided steps) a hand-written 181-byte i386 ELF (hex0) that just converts hex to bytes. It builds a hex assembler with labels, then a macro assembler, then a tiny C-subset compiler (M2-Planet), then GNU Mes, then TinyCC, then GCC 4.0.4. GCC then climbs through 4.7.4 (the last GCC written in C), 10 and 13. That GCC builds Go 1.4 (the last Go written in C), each Go release builds the next, and the final toolchain compiles the Linux kernel and the Talos binaries.
Re:Talos Linux
by cwiggs on Wednesday November 18, 2026 @09:45AM
(#50546936)
I was lucky enough to run Talos Linux for our k8s cluster for various clusters in AWS at my previous job. I was very excited to work on Talos because I've heard such good things and I love the idea of it being so minimal and secure. Unfortunately I don't have great things to say.
1. The (paid) support we received was terrible. We constantly got push back and support didn't really ever help us with issues.
2. We used KubeSpan, which at the time leaked the private wireguard key to the logs when debug mode was enabled. The Preshared key was also written to the logs...and the preshared key is also the cluster secret. This has since been fixed in [this PR](https://github.com/siderolabs/talos/commit/d0b50aa76)
3. We later had an issue with KubeSpan where all our worker nodes dropped one node from it's peerlist. We were using Sidero's discovery service, so we didn't have as much visibility into why/if the discovery service failed. By design the kubelet heartbeat doesn't go over KubeSpan/Wireguard, therefore the worker node reported healthy the entire time, even though no traffic inside the cluster could reach the node, which caused issues with our pods in the cluster. Sidero support wasn't helpful and we decided to just rip out KubeSpan.
For all these reasons I'm not as positive about Talos anymore. I went looking for a CNCF backed project and found https://kairos.io/ , which then has their Hadron OS. I have never tested Kairos+Hadron and I don't like that it uses systemd, but it is at least a CNCF backed project. I'd feel a lot more comfortable if Sidero (Yardi?) donated Talos to the CNCF.
I still think Talos is the best option out there but I hope that changes, or I hope Sidero gets better.
I run a bunch of openshift clusters for very private container workloads and VMs.
You can't beat a local fiber channel SAN for speed and reliability, especially compared to cloud software defined storage, and even on the networking side of things if you worked out your local connectivity well using real networking vs cloud providers software defined networks is so much better.
CSIs are hit and miss, some vendors are really really good, some are meh (hitachi, even though they are improving a lot with the latest relases), but the moment you get an IBM nvme san and start to tune all layers you get performance that are out of this world, some of our dbms workloads are very io sensitive and they improved 10x with nvme fabric.
Often I find that a load balancer and a couple of docker compose file sis more rugged, because you'll be able to leave it for a year and it'll still work and easily be understood by just glancing at it.
If you don't need a cluster then a single node K3s is still better than Docker Compose. You might know how an ad hoc load balancer and Docker Compose works, but I wouldn't say this knowledge is any more given than knowing K8s these days.
But if you do need a cluster then maintaining a bunch of Talos nodes is easier than maintaining a bunch of full fat GNU/Linux boxen tbh.
I feel like k8s is better as soon as your thing goes from simple self-contained thing to multiple things running on one box
especially with agents it's never been easier to work with k8s.
from my experience, k3s supports basically everything k8s has that you'd need for a single-node setup
not sure what "Would k3s handle the proxy layer" means but if you mean the load-balancing part, then yes, that's like one of the main abstractions that k8s provides. but again, it provides many knobs so you probably just wanna pick the simplest option available
Yeah that's what I thought. Perhaps just put HAProxy in front and then let k3s know about it?
or use the Gateway/Ingress abstractions that basically do this automatically, where you just expose services via them, and they handle the routing
looks like the above is documented at: https://www.haproxy.com/documentation/kubernetes-ingress/
the nodeport way would basically be what you'd do with compose, where you map the host (node) ports to the container ones
For load balancing across a proper cluster then you can either provision an external load balancer (like keepalived) or you can look into something like MetalLB which runs entirely on-cluster (but you need BGP support for true load balancing).
In a cloud cluster (like AWS EKS) you get load balancers automatically just by provisioning a LoadBalancer type Service in Kubernetes.
It gave me a ton of appreciation for Kubernetes and Linux and how k8s is built on linux assumptions: the file system layout’s a little different in Talos to the default k8s assumptions, so there were some small adjustments I’d make every now and then.
Sidero’s professional support is out of the world, too. I had simple questions and got answers back from their devops engineers who are running clusters 100-1000 times the size of anything I had to touch, or from their software engineering team who write Talos.
Definitely worth a look if you’re hosting in your home lab, definitely worth a look if you’re doing self-hosted k8s for your job.
I'm asking because I haven't touched k8s for some years, maybe the situation has changed and I should re evaluate my choices?
Now if your home lab priority is to run a plex server for your family and you rarely want to touch the lab, then yes k8s is likely overkill.
I run my stuff on Ubuntu minimal with K3S, it was a fun journey to learn about k8s and related software (Argo,Renovate, CNPG etc..)
Jesus.
At this point it’s a reflex and I race to the HN comments section to find something actually worth reading.
Systemd is essentially growing in this direction but surprised there isn’t something out there already.
I'll say it was my employer VPN intercepting something or whatever.
Talos Containers: Running Containers Without Kubernetes | TalosCon Amsterdam 2026 https://youtu.be/xe_-upvfp5s
talos-hypervisor https://www.siderolabs.com/talos-hypervisor
Really looking forward to see their new "Director" product with the hypervisor support for running VM's as well. I hope the pricing is resonable enough that it is a feasible solution for companies I work with. Would be useful to use that for creating virtual k8s nodes, to slice the same physical hardware into more clusters.
(If Siderolabs are here, it would be awesome to be able to use Director as a hobbyist in the future as well - I think it's a great sales pipeline for you to allow enthusiasts play with it at home and pitch it to internal stakeholders)
There’s been a few times on fly.io where assets just disappear with no trace.
Support is… interesting.
For those of you who don’t know what that is I highly recommend to look it up.
In short build dag starts with (ai provided steps) a hand-written 181-byte i386 ELF (hex0) that just converts hex to bytes. It builds a hex assembler with labels, then a macro assembler, then a tiny C-subset compiler (M2-Planet), then GNU Mes, then TinyCC, then GCC 4.0.4. GCC then climbs through 4.7.4 (the last GCC written in C), 10 and 13. That GCC builds Go 1.4 (the last Go written in C), each Go release builds the next, and the final toolchain compiles the Linux kernel and the Talos binaries.
1. The (paid) support we received was terrible. We constantly got push back and support didn't really ever help us with issues.
2. We used KubeSpan, which at the time leaked the private wireguard key to the logs when debug mode was enabled. The Preshared key was also written to the logs...and the preshared key is also the cluster secret. This has since been fixed in [this PR](https://github.com/siderolabs/talos/commit/d0b50aa76)
3. We later had an issue with KubeSpan where all our worker nodes dropped one node from it's peerlist. We were using Sidero's discovery service, so we didn't have as much visibility into why/if the discovery service failed. By design the kubelet heartbeat doesn't go over KubeSpan/Wireguard, therefore the worker node reported healthy the entire time, even though no traffic inside the cluster could reach the node, which caused issues with our pods in the cluster. Sidero support wasn't helpful and we decided to just rip out KubeSpan.
4. Sidero deprecated their cluster-api and push people to use their BSL Omni (A common trend I got from support was being pushed to use their enterprise offerings) https://github.com/siderolabs/cluster-api-bootstrap-provider...
5. Recently Sidero (the company behind Talos) got acquired: https://www.siderolabs.com/blog/sidero-labs-joins-yardi . Time will tell if that is good or bad.
For all these reasons I'm not as positive about Talos anymore. I went looking for a CNCF backed project and found https://kairos.io/ , which then has their Hadron OS. I have never tested Kairos+Hadron and I don't like that it uses systemd, but it is at least a CNCF backed project. I'd feel a lot more comfortable if Sidero (Yardi?) donated Talos to the CNCF.
I still think Talos is the best option out there but I hope that changes, or I hope Sidero gets better.