Another user posted the blog where they discuss their speedup techniques: https://tailscale.com/blog/more-throughput/

It’s likely that the kernel version can use similar techniques to surpass the performance of the userspace version that tailscale uses, but no one has put in the work to to make the kernel implementation as sophisticated as the userspace one.

I had a look through the comments on this HN thread the other day and came away more intrigued by https://github.com/openobserve/openobserve than hyperdx. Hyperdx is built on top of clickhouse whereas open observe has it’s own storage engines based on parquet files that can be accessed from local disk, S3, or a few other protocols.

I haven’t tried either option yet… I’m, currently using netdata for metrics and don’t do anything special for logs or tracing, but at tiny self-hosting scale I often find software with it’s own storage engines (often sqlite) to be extra hassle-free. I’m curious to kick the tires on openobserve for that reason.

This is a very strong explanation of what’s going on. And as a follow-up, I believe that ZeroTier present a single Ethernet broadcast domain, and so WoL tricks are more likely to work naturally there than with Wireguard. I haven’t used ZeroTier, and I do use Wireguard via Tailscale/Headscale. I’ve never missed the Ethernet features of ZeroTier and they CAN result in a very chatty wan if you’re not careful. But I think ZT would make this straightforward.

Though as other people note… the simplest/least-disruptive change is probably to expose some scripty thing on the rpi that can be triggered via be triggered over a routed protocol and then have the rpi emit the Ethernet broadcast packets from the physical network.

No no, sorry. I mean can I still have all my network traffic go through some VPN service (mine or a providers) while Tailscale is activated?

Tailscale just partnered with Mullvad so this works out of the box for that setup: https://tailscale.com/blog/mullvad-integration/

For others, it’s a “yes on paper” situation. It will probably often not work out of the box, but it seems likely to be possible as an advanced configuration. At the end of the line of possibilities, it would definitely be possible to set up a couple of docker containers as one-armed routers, one with your VPN and one with Tailscale as an exit node. Then they can each have their own networking stack and you can set up your own routes and DNS delegating only the necessary bits to each one. That’s a pretty advanced setup and you may not have the knowhow for it, but it demonstrates what’s possible.

To a first approximation, Tailscale/Headscale don’t route and traffic.

Ah, well damn. Is there a way to achieve this while using Tailscale as well, or is that even recommended?

Is there a way to achieve what? Force tailscale to route all traffic through the DERP servers? I don’t know, and I don’t know why you’d want to. When my laptop is at home on the same network as my file-server, I certainly don’t want tailscale sending filserver traffic out to my Headscale server on the Internet just to download it back to my laptop on the same network it came from. I want NAT traversal to allow my laptop and file-server to negotiate the most efficient network path that works for them… whether that’s within my home lab when I’m there, across the internet when I’m traveling, or routing through the DERP server when no other option works.

OpenVPN or vanilla Wireguard are commonly setup with simple hub-and-spoke routing topologies that send all VPN traffic through “the VPN server”, but this is generally slower path than a direct connection. It might be imperceptibly slower over the Internet, but it will be MUCH slower than the local network unless you do some split-dns shenanigans to special-case the local-network scenario. With Tailscale, it all more or less works the same wherever you are which is a big benefit. Of course excepting if you have a true multigigabit network at home and the encryption overhead slows you down… Wireguard is pretty fast though and not a problematic throughout limiter for the vast majority of cases.

Have a read through https://tailscale.com/blog/how-nat-traversal-works/

You, and many commenters are pretty confused about out tailscale/Headscale work.

1. To a first approximation, Tailscale/Headscale don’t route and traffic. They perform NAT traversal and data flows directly between nodes on the tailnet, without traversing Headscale/Tailscale directly.
2. If NAT traversal fails badly enough, it’s POSSIBLE that bulk traffic can flow through the headscale/tailscale DERP nodes… but that’s an unusual scenario.
3. You probably can’t run Headscale from your home network and have it perform the NAT traversal functions correctly. Of course, I can’t know that for sure because I don’t know anything about your ISP… but home ISPs preventing Headscale from doing it’s NAT traversal job are the norm… one would be pleasantly surprised to find that a home network can do that properly.
4. Are younreally expecting 10gb/s speeds over your encrypted links? I don’t want to say it’s impossible, people do it… but you’d generally only expect to see this on fairly burly servers that are properly configured. Tailscale just in April bragged about hitting 10gb speeds with recent optimizations: https://tailscale.com/blog/more-throughput/ and on home hardware with novice configd I’d generally expect to see roughly more like single gigabit.

I don’t know what’s up on your case, but I would not jump to the conclusion that it’s impossible to use tailscale with any other VPN in any circumstance.

Rather, tailscale and Mullvad will now work easily and out of the box. For other VPNs, you may need to do understand the topology and routing of virtual devices and have the technical ability and system permissions to make deep networking changes.

So I’d expect one can probably find a way for most things to coexist on a Linux server. On a non-rootrr android phone? I’m less confident.

So I have a question, what can I do to prevent that from happening? Apart from hosting everything on my own hardware of course, for now I prefer to use VPS for different reasons.

Others have mentioned that client-caching can act as a read-only stopgap while you restore Vaultwarden.

But otherwise the solution is backup/restore. If you run Vaultwarden in docker or podman container using volumes to hold state… then you know that as long as you can restart Vaultwarden without losing data that you also know exactly what data needs to be backed up and what needs to be done to restore it. Set up a nightly cron job somewhere (your laptop is fine enough if you don’t have somewhere better) to shut down Vaultwarden, rsync it’s volume dirs, and start it up again. If you VPS explodes, copy these directories to a new VPS at the same DNS name and restart Vaultwarden using the same podman or docker-compose setup.

All that said, keeypass+filesync is a great solution as well. The reason I moved to Vaultwarden was so I could share passwords with others in a controlled way. For single-user, I prefer how keypass folders work and keepass generally has better organization features… I’d still be using it for only myself.

And just today with a comment by a world admin! Hopefully they’ll get it sorted soon.

Yeah, snapshots sent to a separate and often remote pool is an extremely common backup strategy for folks who have long-term settled on ZFS. There’s very nice tooling for this that presents a more traditional schedule/retention based interface to save you scripting snapshots and sends directly.

• Sanoid is an old standby in that space.
• Zrepl is getting a lot of traction lately and seems to be an up-and-coming option.
• I use pyznap, but I don’t recommend it to others as as the maintainer is on a multi-year hiatus which makes it undermaintained. It works great, but isn’t getting active development which makes it a poor bet in a crowded space with many great options. I plan to eval Zrepl when I get around to it.

I don’t know if what you’re suggesting is possible, which as I read it is to split your “live” raid-1 in half and use one drive to rebuild the “live” pool and the other drive to rebuild the “backups” pool. It might be, but I can’t think of any advantage to that approach and it’s not something I would have thought to attempt.

I’d do one of:

• Ship the data over the network using ZFS send or something like syncoid/sanoid (which use ZFS send under the hood). It might be slow, but is that an issue? Waiting a week for the initial sync might be fine.
• But syncing by sneakernet is a good strategy too, and can be faster if your backup site is close or your connectivity is slow. In this case, I’d build the backup pool at the live site… ideally in an external drive bay… but one could plug it in internally as well. Then sync them with a local ZFS send, export the backup pool, detach and transport the backup pool to the backup site, them reattach the backup pool at the backup site and import it. Et Voila, the backup pool is running at the remote site fully populated with data and subsequent ZFS sends will be incremental.

Splitting and rebuilding your live pool might be possible, but I can imagine a lot of that might go wrong and I can’t see any reason to do it that way over export/import.

You connect to Headscale using the tailscale clients, and configuration is exactly the same irrespective of which control server you use… with the exception of having to configure the custom server url with Headscale (which requires navigating some hoops and poor docs for mobile/windows clients).

But to my knowledge there are no client-side configs related to NAT traversal (which is kind of the goal… to work seamlessly everywhere). The configs themselves on the headscale server aren’t so bad either, but the networking concepts involved are extremely advanced, so debugging if anything goes sideways or validating that your server-side NAT traversal setup is working as expected can be a deep dive. With Tailscale, you know any problems are client-side and can focus your attention accordingly… which simplifies initial debugging quite a lot.

… only if you are in the US and get an API key from NCMEC. They are very protective of who gets the keys and require a zoom call as well.

Do you have a source for these statements, because they directly contradict the Cloudflare product announcement at https://blog.cloudflare.com/the-csam-scanning-tool/ which states:

Beginning today, every Cloudflare customer can login to their dashboard and enable access to the CSAM Scanning Tool.

… and shows a screenshot of a config screen with no field for an API key. Some CSAM scanners do have fairly limited access, but Cloudflare’s appears to be broadly available.

Yeah, misread the pricing page. Fixed the post, thanks for the correction.

I use Headscale, but Tailscale is a great service and what I generally recommend to strangers who want to approximate my setup. The tradeoffs are pretty straightforward:

• Tailscale is going to have better uptime than any single-machine Headscale setup, though not better uptime than the single-machine services I use it to access… so not a big deal to me either way.
• Tailscale doesn’t require you to wrestle with certs or the networking setup required to do NAT traversal. And they do it well, you don’t have to wonder whether you’ve screwed something up that’s degrading NAT traversal only in certain conditions. It just works. That said, I’ve been through the wringer already on these topics so Headscale is not painful for me.
• Headscale is self-hosted, for better and worse.
• In the default config (and in any reasonable user-friendly, non professional config), Tailscale can inject a node into your network. They don’t and won’t. They can’t sniff your traffic without adding a node to your tailnet. But they do have the technical capability to join a node to your tailnet without your consent… their policy to not do that protects you… but their technology doesn’t. This isn’t some surveillance power grab though, it’s a risk that’s essential to the service they provide… which is determining what nodes can join your tailnet. IMO, the tailscale security architecture is strong. I’d have no qualms about trusting them with my network.
• Beyond 3 devices users, Tailscale costs money… about $6 US in that geography. It’s a pretty reasonable cost for the service, and proportional in the grand scheme of what most self-hosters spend on their setups annually. IMO, it’s good value and I wouldn’t feel bad paying it.

Tailscale is great, and there’s no compelling reason that should prevent most self-hosters that want it from using it. I use Headscale because I can and I’m comfortable doing so… But they’re both awesome options.

My money is also on IO. Outside of CPU and RAM, it’s the most likely resource to get saturated (especially if using rotational magnetic disks rather than an SSD, magnetic disks are going to be the performance limiter by a lot for many workloads), and also the one that OP said nothing about, suggesting it’s a blind spot for them.

In addition to the excellent command-line approaches suggested above, I recommend installing netdata on the box as it will show you a very comprehensive set of performance metrics without having to learn to collect each one on the CLI. A downside is that it will use RAM proportional to the data retention period, which if you’re swapping hard will be an issue. But even a few hours of data can be very useful and with 16gb of ram I feel like any swapping is likely to be a gross misconfiguration rather than true memory demand… and once that’s sorted dedicating a gig or two to observability will be a good investment.

Tailscale is out, unfortunately. Because the server also runs Plex and I need to use it with Chromecast on remote access…

I rather suspect you already understand this, but for anyone following along… Tailscale can be combined with other networking techniques as well. So one could:

• Access Plex from a Chromecast on your home network using your physical IP, and on your tailnet using the overlay IP.
• Or one could have some services exposed publicly and others exposed on the tailnet. So Immich could be on the tailnet while Plex is exposed differently.

It’s not an all or nothing proposition, but of course the more networking components you have the more complicated everything gets. If one can simplify, it’s often well worth doing so.

Good luck, however you approach it.

So for something like Jellyfin that you are sharing to multiple people you would suggest a VPS running a reverse proxy instead of using DDNS and port forwarding to expose your home IP?

I run my Jellyfin on Tailscale and don’t expose it directly to the internet. This limits remote access to my own devices, or the devices of those I’m willing to help install and configure tailscale on. I don’t really trust Jellyfin on the public internet though. It’s both a bit buggy, which doesn’t bode well for security posture… and also a misconfiguration that exposes your content could generate a lot of copyright liability even if it’s all legitimately licensed since you’re not allowed to redistribute it.

But if you do want it publicly accessible there isn’t a hoge difference between a VPS proxying and a dynamic DNS setup. I have a VPS and like it, but there’s nothing I do with it that couldn’t be done with Cloudflare tunnel or dyndns.

What VPS would you recommend? I would prefer to self host, but if that is too large of a security concern I think there is a real argument for a VPS.

I use linode, or what used to be linode before it was acquired by Akamai. Vultr and Digitalocean are probably what I’d look to if I got dissatisfied. There’s a lot of good options available. I don’t see a VPS proxy as a security improvement over Cloudflare tunnel or dyndns though. Tailscale is the security improvement that matters to me, by removing public internet access to a service entirely, while lettinge continue to use it from my devices.

Do I need to set up NGINX on a VPS (or similar cloud based server) to send the queries to my home box?

A proxy on a VPS is one way to do this, but not the only way and not necessarily the best one… depending on your goals.

• You can also use port-forwarding and dyndns to just expose the port off your home-ip. If your ISP is sucky, this may not work though.
• You can also use Cloudflare’s free tunneling product, which is basically a hosted proxy that acts like a super port-forward that bypasses sucky ISP restrictions.
• If you want to access Immich yourself from your own devices but don’t need to make it available to (many) others on devices you don’t control, I like and use tailscale the best. The advantage of tailscale is that Immich remains on a private network, not directly scannable from the internet. If there’s a preauth exploit published and you don’t pay attention to update promptly, scanners WILL exploit your Immich instance with internet-exposed techniques… whereas tailscale allows you to access services that internet scanners cannot connect to, which is a nice safety net.

Do I need to purchase a domain (randomblahblah.xyz) to use as the main access route from outside my house?

Not for tailscale, and I don’t think for Cloudflare tunnel. Yes for a VPS proxy.

I’ve run a VPS for a long while and use multiple techniques for different services.

• Some services I run directly on the VPS because it’s simple and I want them to be truly publicly accessible.
• Other services I run on a bigger server at home and proxy through the VPS because although I want them to be publicly accessible, they require more resources than my VPS has available. When I get around to installing Immich, there’s a decent chance it will go into this category.
• Still other services, I run wherever and attach them to my tailnet. These I access myself on my own devices (or maybe invite a handful of trusted people into my tailnet), but aren’t visible to the public internet. If I decide not to use immich’s shared gallery features (and so don’t need it publicly accessible) or decide I don’t trust it security-wise… it will go here instead of the proxy-by-vps category.

I use k8s at work and have built a k8s cluster in my homelab… but I did not like it. I tore it down, and currently using podman, and don’t think I would go back to k8s (though I would definitely use docker as an alternative to podman and would probably even recommend it over podman for beginners even though I’ve settled on podman for myself).

1. K8s itself is quite resource-consuming, especially on ram. My homelab is built on old/junk hardware from retired workstations. I don’t want the kubelet itself sucking up half my ram. Things like k3s help with this considerably, but that’s not quite precisely k8s either. If I’m going to start trimming off the parts of k8s I don’t need, I end up going all the way to single-node podman/docker… not the halfway point that is k3s.
2. If you don’t use hostNetworking, the k8s model of traffic routes only with the cluster except for egress is all pure overhead. It’s totally necessary with you have a thousand engineers slinging services around your cluster, but there’s no benefit to this level fo rigor in service management in a homelab. Here again, the networking in podman/docker is more straightforward and maps better to the stuff I want to do in my homelab.
3. Podman accepts a subset of k8s resource-yaml as a docker-compose-like config interface. This lets me use my familiarity with k8s configs iny podman setup.

Overall, the simplicity and lightweight resource consumption of podman/docker are are what I value at home. The extra layers of abstraction and constraints k8s employs are valuable at work, where we have a lot of machines and alot of people that must coordinate effectively… but I don’t have those problems at home and the overhead (compute overhead, conceptual overhead, and config-overhesd) of k8s’ solutions to them is annoying there.