Skip to content
Notifications
Clear all

Has anyone benchmarked Tailscale Funnel's performance for web apps?

8 Posts
8 Users
0 Reactions
58 Views
(@cloud_cost_hawk_new)
Reputable Member
Joined: 7 months ago
Posts: 333
Topic starter   [#24599]

Alright, let's cut through the marketing. Tailscale Funnel is basically them selling you a way to punch a hole through your own firewall without opening ports, all while they route your traffic. Convenient? Sure. But "zero trust" starts to look a lot like "zero control over egress costs" when you're pushing real web app traffic through it.

Has anyone actually run benchmarks on throughput and latency for a Funnel-exposed service versus a traditional cloud load balancer (ALB, Cloudflare Tunnel, etc.)? I'm particularly suspicious of:

* The bandwidth costs once you're out of the free tier. Their egress pricing is... opaque until you're on an Enterprise plan.
* The performance ceiling. Is it just a WireGuard relay under the hood, and does that introduce a bottleneck compared to a globally distributed proxy network?
* The true multi-cloud benefit, or if you're just trading AWS egress fees for Tailscale egress fees with extra hops.

I set up a basic test with a Funnel exposing a simple API. The latency add wasn't trivial for us-east to eu-west. Here's a crude comparison from a `curl` timing:

```bash
# Direct to origin (hypothetical, with open port)
time curl -s https://origin.example.com/api/ping > /dev/null
# real 0.085s

# Via Tailscale Funnel
time curl -s https://funnel-tailscale.ts.net/api/ping > /dev/null
# real 0.217s
```

That's not nothing. For internal tools, fine. For customer-facing web apps? I'm not sold.

I want to see real numbers before anyone thinks this is a cheap replacement for a CDN or a cloud provider's LB. What's the catch they're not advertising? Is the performance hit consistent, or does it fall apart under concurrent connections?

-- cost first


-- cost first


   
Quote
(@gracem)
Reputable Member
Joined: 4 months ago
Posts: 294
 

That latency add you saw lines up with my experience. It's the relay hop - you're not hitting a global anycast network like Cloudflare. The traffic goes through a Tailscale node, which adds that extra leg.

Their egress pricing gets clearer once you hit their paid team tier, but it's still a per-user model. That makes it tricky to estimate for anonymous web traffic. If you're serving a public web app, you're paying for funnel egress on top of your cloud egress, so the cost stacking is real.

I still use it for internal tool demos because the setup is so fast. But for anything with real throughput, I'd probably go with a tunnel service built for public traffic.


Automate everything.


   
ReplyQuote
(@hannahg)
Reputable Member
Joined: 4 months ago
Posts: 273
 

Exactly, that extra relay hop is the killer for anything latency-sensitive. I was testing a prototype with some live chart updates and even that ~50ms add made the UI feel sluggish. It's fine for sharing static demos where you're just clicking through flows, but real user interactions need that direct feel.

The cost model is what finally pushed me off it for client projects. Explaining "anonymous web traffic" as billable users to a finance person is a nightmare. They see a public URL and assume it's like any other hosting bill.



   
ReplyQuote
(@gracep)
Reputable Member
Joined: 4 months ago
Posts: 297
 

That latency add you're seeing is the relay. It's not anycast, so you're bound to a single Tailscale node location. For a simple API it might be fine, but it scales poorly.

The egress cost stacking is real. You're paying your cloud provider egress, then Tailscale egress on top. It can get expensive fast for public traffic.

>Is it just a WireGuard relay under the hood
Essentially, yes. It's their infrastructure routing your traffic, not a distributed CDN. For a high-throughput web app, you'll hit a bottleneck much sooner than with a dedicated LB or tunnel service built for it.

I'd only use Funnel for low-volume, internal-facing prototypes. Anything else needs a proper solution.


Data over opinions


   
ReplyQuote
(@billyp)
Reputable Member
Joined: 4 months ago
Posts: 284
 

Great question. That latency add you're seeing is the extra relay hop, and it's pretty consistent. For east-to-west, you're probably looking at an extra 40-70ms, depending on which Tailscale relay node you land on.

On the cost, you've hit the nail on the head. You're paying twice: once to your cloud provider and once to Tailscale for egress. For a public web app, that can get painful fast compared to a service like Cloudflare Tunnel where you're only paying your own cloud egress.

It's fantastic for quickly sharing a staging site with a client, but I wouldn't build a production user-facing app on it. The performance ceiling is real for anything needing snappy responses.


Always A/B test.


   
ReplyQuote
 danw
(@danw)
Reputable Member
Joined: 4 months ago
Posts: 387
 

The double egress cost is the real kicker. It's cheap for prototyping until your first bill with real traffic. Then you're scrambling to migrate.

That ~50ms relay penalty is fine for internal tools where the alternative is a VPN setup. But for any customer-facing app, it's a non-starter. Users will notice, and your CFO will notice the bill.



   
ReplyQuote
(@grafana_guy_night)
Honorable Member
Joined: 8 months ago
Posts: 427
 

Yeah, that latency add you saw lines up with my testing too. I was exposing a demo for a client in Germany and my pings from the US went from ~110ms to almost 180ms. That extra hop is real.

I've been wondering about the throughput part myself. For my little monitoring dashboards it's fine, but I wouldn't try streaming any real data through it. Have you tried hitting it with something like `wrk` to see where it starts to choke?



   
ReplyQuote
(@henryg78)
Estimable Member
Joined: 4 months ago
Posts: 165
 

>Explaining "anonymous web traffic" as billable users to a finance person is a nightmare.

Ran into this on a project. The per-user cost model isn't just opaque, it's incompatible with scaling a public service. You can't forecast it like standard cloud egress.

On latency, your 50ms aligns with my tests. More concerning is the variance. I saw a standard deviation of ~15ms on sustained connections, which points to relay queueing under load. It's not just the fixed hop; it gets jittery.

Good for demos, but I wouldn't use it for live chart data either.


EXPLAIN ANALYZE


   
ReplyQuote