The way I handle this is to run unbound on a server in the public cloud and then tunnel over TLS from my local unbound to the cloud instance. My local clients query a PiHole, which forwards to unbound on localhost:15353, which forwards everything over TLS to the fully recursive instance of unbound in the cloud, which uses root.hints.

Still, your cloud vendor can see the requests.

But someone can see it, but you can rotate upstream resolvers to split requests if you have to.

How is the latency for such a setup? And how is the general browsing experience with said latency (I realize it also depends on which sites you frequently access)? I see some cloud providers/CDNs using a caching TTL as low as five minutes.

You spend most of the time waiting for the actual resolve to complete, not for the "transfer" time between you and VPS.

Source: I'm running Unbound on my notebook, I'm actually queried the stats for some heated discussion on reddit.

For example my current stats_noreset:

    histogram.000000.000512.to.000000.001024=17
    histogram.000000.001024.to.000000.002048=33
    histogram.000000.002048.to.000000.004096=251
    histogram.000000.004096.to.000000.008192=509
    histogram.000000.008192.to.000000.016384=1161
    histogram.000000.016384.to.000000.032768=1891
    histogram.000000.032768.to.000000.065536=2611
    histogram.000000.065536.to.000000.131072=3197
    histogram.000000.131072.to.000000.262144=2502
    histogram.000000.262144.to.000000.524288=1547
    histogram.000000.524288.to.000001.000000=857
    histogram.000001.000000.to.000002.000000=121
    histogram.000002.000000.to.000004.000000=70
    histogram.000004.000000.to.000008.000000=22
    histogram.000008.000000.to.000016.000000=441
    histogram.000016.000000.to.000032.000000=80

As you can see most of queries are completed in a way below 500ms. Adding another 20-40ms on top that doesn't change anything, because caching is a thing and with Unbound you can even ask to actually refresh the expiring records, so you would be served a fresh one from the cache every time, though I never bothered with it, it works fine even without it.

In addition to the root hints, you should also download the DNSSEC anchor key (available on the same site as the root hints). That will let you detect manipulations of records that are DNSSEC-signed.

Otherwise, you could spin up your recursive resolver on your cloud, VPS, or other hosting provider of choice, and then use that.

And make sure it only resolves records for you. Don't leave it wide open or it will be used as an attack vector. https://www.cloudflare.com/learning/ddos/dns-amplification-d...

Right, DNSSEC will solve the "manipulate" problem, but it won't solve the "see" problem. But whether that's a concern is up to you. You could run your resolver on a VPS and speak DoT / DoH to that, which shifts the leak from your ISP to your VPS provider.

It doesn't solve the "manipulate" problem we're talking about here, either: nothing about DNSSEC prevents a DNS server (or middlebox) from denying results to a disfavored domain; it only (situationally) prevents them from redirecting it somewhere else. (And, of course, it only works if you're running your own recursive server; it does nothing whatsoever in the 8.8.8.8-type use case).

> nothing about DNSSEC prevents a DNS server (or middlebox) from denying results to a disfavored domain

But at least it is detectable thanks to NSEC and NSEC3 records.

Kind of. An intermediary can drop packets and the client will never get the response.

It's detectable when the site that the DNS provider is censoring falls off the Internet!

Yes, that's true.

In theory if you got a resolver which could disable UDP queries, it would then default to TCP, and the ISP couldn't manipulate those. Don't know if any resolver supports disabling UDP though.

> would then default to TCP, and the ISP couldn't manipulate those

They can easily manipulate TCP as well. Unless you establish an authenticated session like TLS, TCP can be mitm-ed easily.

Unbound as a client will fall back to TCP automatically if the UDP response is too big. I'm uncertain if there is widespread support for TCP servers though.

Unbound