How to test lb_estimator and lb_strategy, and advice on my configuration?

[manuel-plavsic]

I have set lb_estimator to true and lb_strategy to "first" and now I would like to test if they are working correctly. Unfortunately, I do not know how to test how lb_estimator constantly tries to estimate the latency of all the resolvers. I tried using systemctl status dnscrypt-proxy2.service, but I don't see the new estimations getting reported there. Can someone explain me this? It is not explained in https://github.com/DNSCrypt/dnscrypt-proxy/wiki/Load-Balancing-Options.

Additional information and context

I recently had some issues with my configuration, where almost all websites would load very slowly (especially at night). I know this because I tried disabling dnscrypt, and I could suddenly browse all websites extremely fast. Enabling it again made it very slow again. This happens only occasionally. I would like to find a way to ensure that a fast resolver is always switched to in case of performance decrease.

My configuration

I am using these servers:

server_names = [
  "dnscrypt-ip4-filter-pri"
  "dnscrypt-ip4-filter-alt"
  "dnscrypt-ip4-filter-alt2"
];

with the following public resolver urls:

urls = [
  "https://quad9.net/dnscrypt/quad9-resolvers-dnscrypt.md"
  "https://raw.githubusercontent.com/Quad9DNS/dnscrypt-settings/main/dnscrypt/quad9-resolvers-dnscrypt.md"
];

How possible is it that all Quad9 servers are simultaneously affected? Also, should setting lb_strategy to "p2" (or maybe even "p3") be preferred in my case in your opinion?

NB: I am on NixOS (unstable).

[ghost]

In short the server list is sorted by response time of queries.
With first the 1st server, i.e. the fastest 1, is chosen. With p2 a random 1 of the 1st 2. p3 with only 3 servers means just using a random 1.
Note that response times have very high variance and therefore the sorting might not be perfect.
However the fewer servers there are, the more likely it is that the fastest 1 is chosen.
So having only 3 servers and using first should very likely choose a fast server, i.e. the slowdown you see is probably an actual server problem.

[manuel-plavsic]

Thanks for taking time to answer. I will keep "first" then.
What do you mean exactly with "the slowdown you see is probably an actual server problem"? That the one exact server I am using in a given span of time has a server problem, or that all Quad9 servers share the same problem?
Any more tips what I can do in this situation?

I also still don't exactly understand how lb_estimator=true works. Are the estimates done e.g. at regular intervals, before every request?

[jedisct1]

The resolvers weights are updated Incrementally, after every transaction.

I wouldn't recommend first. Just use the default p2. Google for "the power of two choices load balancing" to see why.

[ghost]

Any more tips what I can do in this situation?

Probably just look for faster/additional servers.
The order in the server list can change with every query made. It's not that you're stuck with 1 server for 10 minutes or so. So it should have picked a fast server quickly if available.

Google for "the power of two choices load balancing" to see why.

We had this discussion in the past. DNSCrypt-Proxy doesn't support this (for this to work 1 needs to 1st randomly choose 2 servers and afterwards choose the better 1 of them). As the server list is sorted by response time the 1st 1 is more likely to be faster than the 2nd 1.