Tailscale Peer Relays | Hacker News Reader

Tailscale Peer Relays

tailscale.com

364 pointsseemaze114 comments

Great! This feature made a lot of sense, and it took a long time.

It’s like falling back to hub and spoke, except that the traffic is end to end encrypted, and the middle node is used only when direct connection is not possible, and for some clients. It’s also similar to running your own derp server (which works also in TCP), but without the hassle of doing so, and perhaps without having to open ports to the internet (needed in derp) so long as the relay is reachable by peers.

The derp servers have low throughput. Another option could be a pay-as-you-go derp service.

They might also be on their way to remove the need for reverse proxies, with the recent announcement on Tailscale services.

BTW, why could it be paid for more than two relays? You are using just your own devices and bandwidth :)

It actually lower the bandwidth bill for Tailscale by reducing the usage of their own relays. Ideally, by default the software will find whatever nodes could help with direct connection. It’s just routing within your own network.

> It’s also similar to running your own derp server (which works also in TCP), but without the hassle of doing so, and perhaps without having to open ports to the internet (needed in derp) so long as the relay is reachable by peers.

I think most folks will need to open a port to the internet, because otherwise you wouldn't need the tailscale to begin with. e.g. connecting your cloud network to your on premise network etc.

Of course exceptions apply, like both clients can reach the peer relay, but not each other directly.

I could open a port to the internet, but it would be Tailscale’s responsibility to secure the software that listens to the port (subject to an up-to-date software, that is my responsibility).

It’s not a standard Wireguard port. With Wireguard included in Linux, I would not be worried.

This was better solved by tinc about 20 years ago. All tinc nodes can work as relays (but you can disallow that if you want), it does not rely on a centralized server, and works fine without access to the internet. It is a true mesh. The world would be better served by porting tinc to wireguard and some memory safe language instead of reimplementing parts of its functionality from scratch.

Wireguard can do what you're describing.

Wireguard can't punch through NATs or firewalls without third party software like Tailscale. Also I'm pretty sure each peer to peer connection needs to be individually set up in a config file ahead of time

Nebula[0] addresses this and is IMO an improvement over WireGuard. Came out of Slack originally, and it supports peer discovery, NAT hole punching, and some other cool features. Also still uses the Noise Protocol.

In practice, the extra networking features + better first class peer config management baked in is very nice (Nebula’s “lighthouses” are configured with a tool similar to DSNet for Wireguard[1])

[0] https://github.com/slackhq/nebula [1] https://github.com/naggie/dsnet

So now we're back "tailscale but with different steps"

tailscale appears to be a paid product with a free tier, nebula (while DIY) is free

there is an open source control plane called headscale which covers almost all of the features for free (while DIY)

People keep saying that, but haven't we learned already that eventually Tailscale gets bought, then priorities change, then they make incompatible changes because they're need to grow, and headscale either can't keep up, or gets pushed away by Tailscale themselves, and we're back to using $TailscaleCompetitor who promises to not do the same thing.

Just don't rely on centralized for-profit entities, rely on stuff produced by non-profits and foundations, that you know isn't gonna screw you over as soon as they need money.

While I agree in spirit, I find this logic around for profit FOSS projects a little backwards sometimes, because it implies forking Tailscale wouldn't save much time.

What makes you think we'd be better off building a competitor to something open source if it has all the features we want now? The reason we don't see open source competitors to big products is not because people are too dumb to try it. It's because it's way, way harder. It makes way more sense to Fork and work from there while we're still getting this momentum from Tailscale.

If you think Headscale is going to have problems keeping up with a private Tailscale, good luck rebuilding Tailscale.

> Just don't rely on centralized for-profit entities, rely on stuff produced by non-profits and foundations, that you know isn't gonna screw you over as soon as they need money

What do you use that fits that philosophy and offers the basic functionality (NAT traversal, Magic DNS, failover relaying) TS provides?

Nebula has NAT traversal and failover relaying I beleve, but not magic dns

This is correct, and the lack of a MagicDNS solution is definitely felt when using Nebula (esp. when switching from Tailscale)

I am personally happy to use Tailscale directly so I don't know. There isn't anything better out there though.

I use Nebula but the relay config is somewhat clunky and the macOS port is pretty buggy.

What's the story with removing uncooperative endpoints from Nebula?

A $2.5/month vps solves this issue.

It does not. I push hundreds of TB across my private mesh.

When networks get used for unethical or criminal workloads, reliability problems aren’t the tooling's fault. A tiny VPS does the job fine for the rest of us.

What are you talking about? Nothing I am pushing across my mesh is unethical or criminal, from whence these baseless accusations?

I assume the GP assumed that it was pirated media material.

Hundreds of TB over what timeframe? It does seem like quite a bit of data.

Yes. Only use it as rendezvous server for publishing port numbers and establishing direct connections. Do not use for passing traffic

Neither can tinc.

... because its an problem with NAT and not with the protocol.

Yes. I’m replying to a comment thread debating WireGuard vs tinc where someone said that WireGuard couldn’t, as if it was a differentiated thing between them.

Last I checked, tinc supported NAT port punching natively.

As far as I’m aware, that works the same as things like tailscale is doing, where you need at least one node that is publicly reachable to facilitate.

Which is a much smaller requirement than needing extra software on every NATed node.

So yes it is a differentiated thing between wireguard and tinc, as you phrased it in your other comment.

Correct. This is not something Wireguard does, which was the original comparison.

tinc: One public node, thousands of private nodes, with NAT punching. That's fine and typical in my experience.

Folks have solved these problems and open sourced solutions: https://github.com/tonarino/innernet

But all that I've seen are still centralized/federated

(I help maintain innernet)

We have _some_ NAT traversal logic in place, but it's very basic. Tailscale does a much more thorough job on it. It would be cool to add peer relays to innernet but I imagine it's a fair amount of work.

From what I recall, tailscale has their own Wireguard implementation so they have more control over the socket and how things are routed. innernet is just a wrapper around managing wireguard peer lists, and yeah there's a central coordination server which is unfortunate. If the server goes down, you can still connect to peers so thankfully it doesn't bring down your whole network, but you won't be able to learn about new peers or peer endpoints over time until you re-establish connectivity with the coordination server.

It punches through my NAT just fine without third party software.

It's not as simple to make it reliable as it is with Tailscale, but it works

It doesn't universally work without a helper script and a STUN server, though - you need a suitably "friendly" NAT that has reasonably predictable behaviour with respect to port mapping and/or just one side of each pair behind a NAT.

My netgear pr460x supports WireGuard with NAT & firewall; granted it’s not a consumer router.

> Wireguard can't punch through NATs or firewalls without third party software like Tailscale.

That's a false or incorrect statement, I've been using Wireguard and a cheap VPS (actually free on OCI) for several years, and with a cheap VPS at AWS Lightsail before that. No third party software in use at all. The only thing running on the VPS is Wireguard. The only thing running on my peers is Wireguard.

> Also I'm pretty sure each peer to peer connection needs to be individually set up in a config file ahead of time

That's how I do it but there are tools available to make it easy.

Heh. I have a 30-nodes Tinc network over the internet but some hosts are behind a NAT. It keeps randomly losing routes between these nodes. It even has the infuriating behavior that often it loses the route a few seconds after I successfully established a SSH connection.

Also, traffic seems to be decrypted and re-encrypted by relaying nodes. For end-to-end encryption, you need "ExperimentalProtocol = yes" added by Tinc 1.1, which was never formally released.

I'd like to rewrite something like it in a language I'm familiar with (perhaps based on cjdns' protocol which is better documented than Tinc's) but it's not easy.

I remember using Tinc a lot and really liking it. But as 1gbit+ speeds became more common, tinc struggled to maintain that much throughput on lowend hardware.

I also didn't know that tinc was still around, I might give it a try again just to play around with it and see what's changed.

> We believe our new Tailscale Peer Relays connectivity option—unique to Tailscale—gives customers the best performance and flexibility.

Seems pretty similar to some of the stuff ZeroTier was doing years ago. Hard to claim it's unique to Tailscale. Charging for it above and beyond the per user costs seems overboard as well.

I've used ZeroTier in the past and I don't recall a feature similar to this being part of their offering. What I do remember is home brewed crypto, incredibly poor single threaded performance, and a glacial development pace. Any thread about Tailscale inevitably brings up alternatives, but having tried many of them in the past, I've yet to find one that actually competes across all the features and performance. The comparisons to other solutions tend to give me "why do you need Dropbox when FTP exists?" vibes.

It's not clear to me if it's different than Zerotier's routing design or not. Partly because ZT has their "moons" and custom network controllers. Googling about it does seem to indicate ZT doesn't support peers routing traffic directly, but you can setup your own DERP style TCP relays.

To me Tailscale is just a nicer version of Zerotier but not in different genre like ftp vs Dropbox. From a general non-power users perspective they operate almost identically.

ZeroTier supports relaying natively. You create a network with three nodes, the "client" (on the internet), the "gateway" (public subnet) and the "server" (private subnet). ZeroTier will automatically route traffic between the client and server through your gateway with no configuration.

I wonder if the next step could be to have all tailscale clients automatically able to accept forwarding requests between any two machines within the tailnet, so that the mesh seamlessly auto-routes around any breaks within the mesh?

Yeah, such extension is natural and it comes up frequently in the context of overlay networks. The defining question here is if you are OK with relaying other people's CP traffic.

"within the tailnet." So, the "other people" are people you've authorized to be on your tailnet, at least in this conversation.

Ah, my bad. You are right, I missed that bit.

Since the latency would be higher due to the additional hops or distance, I wonder if it's not better to make the request crash so that the issue is clearly detected rather than to continue in a degraded but "invisible" state.

Is it possible to specify the external ipv4 and ipv6 address? There are scenarios where the eggress traffic uses a different address to ingress, or the host has multiple internet-connected addresses but only one has a firewall permitting traffic to the nominated port.

Main downside I see compared to DERP is there's no way for this to work in the browser since it's native UDP. I wonder if it'd be possible to make it work over WebTransport in the future.

(Tailscalar) I'm very bullish on WebTransport in the future, but, it doesn't solve for NAT traversal. I'm watching keenly as well how Iroh gets on with QAD, relatedly.

If it all comes together this may well become more magic :D

But this is specifically for relays, so you shouldn't need NAT traversal right? ie in the browser you already know p2p isn't going to work (unless Tailscale implements WebRTC support) so you just go straight to relays for everything.

I hadn't heard about iroh using QAD. Thanks for that.

True, the relays currently use http and have a web socket path. The web based ssh client in the admin panel uses this transport.

Wow! I just spent a good chunk of time last week setting up headscale and split horizon SSL behind my network, and I expected I was going to just expose a Wireguard UDP port, but discovered no, it’s DERP or nothing. DERP has been OK, but I think just exposing a UDP port on my local network is better.

If we’re really confident in the security of that UDP client, that is. I feel very comfortable exposing a Wireguard bastion, time will tell how secure whatever protocol tailscale is serving, here, will be.

mind developping on the "it's DERP or nothing"? Have you been trying to expose a direct wireguard port of your own, or the Tailscale?

You can use Tailscale to connect services together (not just someone's laptop to a service, replacing OpenVPN), but what if Tailscale has an outage? Will my services not be able to find each other anymore?

The tailscale login servers had an issue last week. My local network had an issue at the same time and all connections dropped. Then none of my stuff could reconnect because I couldnt connect to tailscale :(

Looking into setting up my own headscale instance now. This is the first issue I’ve had with tailscale but seems dumb that my local lan devices couldn’t even talk to each other.

(Tailscalar here) We're taking this kind of outage very seriously. In particular this outage meant newly connected devices couldn't reliably reach our control plane and couldn't get the latest network state. IMO that's not okay.

One of Tailscale's fundamental promises is that we want to try as much as possible to get our control plane and infrastructure as out of the way of your connectivity paths, while still using our infra to "assist" when there's connectivity issues (like difficult to traverse NAT), and maintain trust across the network, and keep everything up to date.

It's a tough balance and this year we're dedicating resources to making sure even small blips in our control plane don't mean temporary losses of connectivity across even your newly woken up devices. In particular we're taking a multi-pronged approach, right now. We're working in parallel to increase client tolerance of control outages (in response to cracks shown in this incident) and have an ongoing effort to make the control plane more resilient and available.

I was literally looking for a solution for this over the weekend and ended up with a very quirky setup for my Kubernetes Operator.

Now I can rip all that out and use this! Bravo!

Man K8s is my nightmare haha. 100% agree

K8s is a hell of a platform but it is a ... hell of a platform.

Hard to parse the networking jargon, but does this enable offline connections?

If I have two devices on my local LAN (both connected to a Tailnet) and my home internet goes, currently the devices disconnect from each other. I have been looking for a way to prevent that, so that the all devices connected to the same WiFi network on a tailnet can find each other even if the internet connection to the wider world is broken.

that already works

Not for me. Maybe some misconfiguration on my part then.

run `tailscale status` and ensure that your local machines are connected to each other `direct` and not using a relay.

> my home internet goes, currently the devices disconnect from each other.

You might be able to solve this by hosting your own control plane such as Headscale. Instead of having Tailscale manage/coordinate all the nodes on tailnet.

https://github.com/juanfont/headscale

I ran headscale for a while and can confirm local devices could still chat when the internet went down.

That'd be easy if only wireguard supported multiple endpoints per peer, then one could add the ULA address of the local devices in addition to the external one.

What's the use case for this? It seems to be for situations where you might have a SaaS product, but there is some data required from a customer system. You'd expose the customer data using this relay and integrate into the SaaS. Is that the gist of it? Integration would still likely involve you giving the customer some software to expose a limited API and handle auth, logging, etc.

They are an alternative to the tailscale operated DERP servers, which are cloud relays.

Even with the much touted NAT punching capabilities of tailscale, there are numerous instances where tailscale cannot establish a true p2p connection. The last fallback is the quite slow DERP relay and from experience it gets used very often.

If you have a peer in your tailscale network that has a good connection and that maybe you can even expose to the internet with a port forward on your router, you now have this relay setting that you can enable to avoid using the congested/shared DERP servers. So there is not really a new use-case for this. It's the same, just faster.

The explanation that I think wasn't entirely clear in the post is how it actually works/why that's better.

From what I can tell, the situation is this:

1. You have a host behind NAT

2. That NAT will not allow you to open ports via e.g. uPnP (because it's a corporate firewall or something, for example) so other tailscale nodes cannot connect to it

3. You have another host which has the same configuration, so neither host can open ports for the other to connect in

The solution is to run a peer relay, which seems to be another (or an existing) tailscale node which both of these hosts can connect to via UDP, so in this circumstance it could be a third node you're already running or a new one you configure on a separate network.

When the two NAT'ed hosts can't connect to each other, they can both opt to connect instead to this peer node allowing them to communicate with each other via the peer node.

Previously this was done via Tailscale's hosted DERP nodes; these nodes would facilitate tailscale nodes to find each other but could also proxy traffic in this hard-NAT circumstance. Now you can use your own node to do so, which means you can position it somewhere that is more efficient for these two nodes to connect to and where you have control over the network, the bandwidth, the traffic, etc.

Is there a way to determine if a particular connection is falling back to DERP today?

I have a pretty basic setup with tailscale setup on an Apple TV behind a bunch of UniFi devices and occasionally tunnelled traffic is incredibly slow.

Wondering if it’s worth setting this up on my Plex server which is behind fewer devices and has a lot of unused network and cpu.

tailscale ping <node IP>

It will tell you how each ping has been answered until a direct connection is established.

Tailscale is a few things. It might be fair to say that it is mostly a software platform with a web frontend that allows orgs (and individual users alike) to easily create secure VPNs, so their various systems can have a secure, unfiltered virtual private network on which to communicate with eachother even if they're individually scattered across the four corners of the Internet.

The usual (traditional) way to do VPN stuff is/was hub-and-spoke: Each system connected to a central hub, and through that hub each system had access to the other systems.

But the way that Tailscale operates is different than that: Ideally, each connected system forms a direct UDP/IP connection with every other system on the VPN. There is no hub. In this way: If node A has data to send to node F, then it can send it directly there without traversing through a central hub.

And that's pretty cool -- this peer-to-peer arrangement is gloriously efficient compared to hub-and-spoke. (It's efficient enough that a person can get quite a lot done with Tailscale for free, with no payment expected ever.)

But we don't live in an ideal world. We instead often live in a world of NAT and firewalls -- sometimes even implemented by the ISPs themselves -- that can make it impossible for two nodes to directly send UDP packets to eachother. This results in unreachable nodes, which is not useful.

So Tailscale's workaround to that Internet problem is to provide Designated Encrypted Relays for Packets (DERP). DERP usually works, and end-to-end encryption is maintained.

DERP is also not at all new. It brings back some aspects of hub-and-spoke, but only for nodes that can't communicate directly; DERP behaves in a way akin to a hub, to help these crippled nodes by relaying traffic between them and the rest of the VPN's nodes.

But DERP is a Tailscale-hosted operation. And it can be pretty slow for some applications. And there was no way, previously, for an individual user to improve the performance of DERP: It simply was whatever it was -- with a collection of DERP servers chewing through bandwidth to provide connectivity for a world of badly-connected VPN nodes.

But today's announcement brings forth Tailscale Peer Relay.

> What's the use case for this?

The primary use case for this is simple: It is an alternative to DERP. A user can now provide their own relay service for their network's badly-connected peers to use. So now, rather than being limited to whatever bandwidth DERP has available, relaying can offer as much bandwidth as a user can afford to pay for and host themselves.

And if a user plans it right, then they can put their Peer Relay somewhere on the network where it can help minimize inter-node latency compared to DERP.

(It's not for everyone. Tailscale isn't for everyone, either -- not everyone needs a VPN at all. I'd never expect a random public customer to use it knowingly and directly.)

Yeah, Tailscale is really cool. The only thing I wish is that they didn't tie auth to either a big tech monopoly (Google, github etc) or running your own IDP service. I would love to use Tailscale for some self hosted stuff I have, but hesitate to start exposing something like an identity management tool because that's a high value target. And of course, I don't really want to let Google et al be in control of my VPN setup either.

That's a valid concern.

I've also used ZeroTier with good success.

They're a competitor that offers VPN with similar idealized P2P topology. Unlike Tailscale, ZT is not based on wireguard (ZT predates wireguard), but they do offer the option to use their own local auth without reliance/potential issues with yet-another party.

ZT also allows a person to create and use their own relay (called a "moon"), if that's something useful: https://rayriffy.com/garden/zerotier-moon

(For my own little purposes I don't really have a preference between either Zerotier or Tailscale.)

Thanks for the tip! I'll check that out and see if it would work for my VPN needs, but it certainly sounds promising.

They support Passkeys. This is exactly how I continue using them after moving away from Google Workspaces.

Oh wow, I had totally missed this[0]! Is it possible to migrate an existing SSO account (with associated tailnet) to a passkey one?

[0]: https://tailscale.com/blog/passkeys

One thing I didn’t understand: it uses an UDP port of my choice. What IP is it using? Everything via the tailnet or do I need to open this port to the internet?

If only available via Tailscale/tailnet - how is connectivity better since if two devices can connect to each other via Tailscale we are already on the direct connection route instead of a relay / derp connection?!

> It allows customers to make just one firewall exception for connections only coming from their tailnet.

You'll need to open a single UDP port on your firewall, so it's your public facing IP address. You don't need an entire VM somewhere, just a single port.

Regarding the speed question. You'd use the derp when it's not possible to make a peer to peer connection, which limits your speed to derp server's speed and load. Which the peer relay, you can practically use the entire bandwidth you have between your devices.

> for connections coming from their tailnet

So instead of whitelisting all ports from IP range 100.64.0.0/10 I would just whitelist e.g. UDP port 12345 coming from IP range 100.64.0.0/10 to my public IP? Or just open up UDP 12345 completely?

I think you need open UDP 12345 to public internet, where the Tailscale Disco protocol runs on that, which is the underlay port.

I would assume you only need to make sure that the other clients can access the UDP port, so not like public internet in the sense of 0.0.0.0/0 but just accessible by other peers, whatever their public facing IP addresses will be.

What I understood is similar to making any device accessible on the internet. You would need to open the UDP 12345 on your router and forward the traffic to your server.

It sounds like each peer would first access the peer relay to coordinate how they could establish a direct point-to-point connection with each other.

(in)convenient timing. I just recently implemented a mesh network using wireguard+bgp so that all nodes can connect to each other one way or another. This would have probably been a lot easier for almost the same result.

Any reason this is not supported on iOS/tvOS devices? Would love to make it work on my Apple TV!

I'm paraphrasing what one of the engineers said when I asked this question to the team (I'm a TS insider):

There's an application size limit associated with NetworkExtension.

https://developer.apple.com/documentation/networkextension

So they would need to get the binary size smaller in order to implement this on iOS or Apple TV.

Yay! Excited to see them building this in public.

I recall that tailscale DERP servers were always slow and made things feel delayed when they had to be used as a relay.

How does this interact with machines that are shared from one Tailnet to another? Is there specific syntax to grant the appropriate permission to a user or device that accesses the destination via sharing?

The docs also say:

> As a rule of thumb, the src devices in the grant policy should typically be devices in a stable physical location behind a strict NAT or firewall that prevents direct connections. This typically includes devices in corporate networks or cloud environments. It usually does not include mobile devices or laptops that frequently change locations and network conditions.

Is there some reason that one should not set up a peer relay to enable a laptop to access a machine that is behind a NAT? (Tailscale regularly fails to establish direct connectivity from a laptop behind a NAT to a machine that's behind a different NAT, at least in my experience.)

My reading of the docs suggests that the relays and both peers must be in the same tailnet. Additionally, both peers must have the correct ACLs set up to access each other

In your example the src would be the "machine that is behind a NAT". That's the one the peer relay enable access to. And then all your other devices (that laptop) can reach it through the peer relay.

I was also a bit confused on the meaning of src/dst in the grants. The naming didn't match my thinking.

Hmm. It would be very nice if this worked when the laptop is on a different tailnet.

One limitation of custom DERP is that across tailnets, they don't share the same DERP maps and don't have access to each others' DERPs.

With Tailscale Peer Relays, the available relay bindings can be seen by the devices on either side of a connection; as such it should work out of the box with a sharing relationship between tailnets.

How to do site-to-site traffic over Tailscale / WG encryption? From preliminary testing, it seems have difficulty to saturate a 10Gbps connection while plain HTTP (nginx) traffic does that fine. Of course it should vary from CPU to CPU, but any tips how to improve that? Ideally I would love to go over with encrypted traffic, although everything is public, just one less thing need to be careful (in case future need to transport some non-public data over).

You'd need to analyze what's bottlenecking when you're doing that.

Even a cursory look at htop on both ends while you're trying to saturate that link would be informative.

I can finally tear down my custom DERP server that I was using to get higher throughput between two NAT'd clients.

Doesn't Libp2p do this relay stuff already? Like, its their core model (and their NAT code is limited.)

Why go through the effort of reimplementing all this instead of using libp2p?

You mean the library that almost got me banned from my VPS provider because it dialed about half the RFC1918 space?

Is there a way to force clients to use a relay? It seems like this is only meant as a fallback, but what if a relayed connection is actually faster (like when direct peering between tailnet members is slow, not rare in consumer connections)

Just set the relay up as an exit node too

Really nice move, just speed and simplicity.

Tailscale playing catchup to netbird...

What is the equivalent feature of netbird? Can you link the documentation please?

I couldn't find a specialised page for it. But here they mention that the relay service can be self hosted: https://docs.netbird.io/about-netbird/how-netbird-works#rela...

It's using a turn server to provide the DERP feature (in the case of the quick setup it's coturn but you can use other turn servers). So to create this you just set up more turn servers and add them. It's fairly simple.

I love Tailwind and use it on all my devices, but I'm not sure I really understand what this is.

OK, another use for port 1337. Anyone got a better, mnemonically memorable choice in the low range?

I'm using 3377 (D-E-R-P) since it's effectively replacing the DERP servers on the specified source devices.

> All customers can use two peer relays, for free, forever. As your needs scale, so will the number of available peer relays. To add even more peer relays to your tailnet, come have a chat with us.

I have to pay to be able to donate my own infra to make tailscale's service better?

Tailscale employee here.

I kinda doubt we'll end up charging for it (as it costs us ~nothing except support costs, which are real), but it's easier to make it free later when it's GA rather than rug pull on people and start charging for it in the future if we start it out free+unlimited.

> to make tailscale's server better?

While I don't think that's accurate, I, too, am surprised that this feature isn't¹ completely free. After all, it will make it easier (or possible at all) for many people to use Tailscale.

¹) s/isn't/might not end up being. See https://news.ycombinator.com/item?id=45751253

Your peer relays don't make Tailscale's infrastructure better.

They instead make your own infrastructure better.

Running a peer relay donates nothing to anyone.

It makes their service better. It fixes slow transfer speeds between my nodes because their derp servers are rate limited.

Reply on news.ycombinator.com