Amsterdam to Lyon is what, 15ms? Your latency baseline matters. Dnsperf with -l 30 or -l 60 is fine for throughput but tells you nothing about how that behaves at 95th percentile when the network hiccups.
I want to see traceroute and mtr output from your test host to your query source. Bad hop will invalidate everything.
I sell white-label DNS on all three of these. BIND for conservative clients who want ISC support contract. Knot for the speed demons who read NLnet Labs blog. CoreDNS for Kubernetes cowboys who think etcd is a database.
The honest answer I give customers: at under 10K zones and under 50K qps sustained, you will not feel the difference. Buy the one your team can debug at 3am.
Member OP
This is perhaps the wisdom I needed. My association has 200 members and perhaps 5 queries per second at peak. I am optimizing the wrong thing. But the learning is still interesting to me.
Same here. Ran similar shootout last year on Hetzner CPX11. CoreDNS won on raw numbers, BIND won on "I understand the logs." Stuck with BIND. Sometimes boring is a feature.
This.
Auckland represent. Also packet loss on dnsperf UDP mode is invisible unless you check the final stats line. Everyone looks at the big number, nobody reads the fine print.
Then run unbound and spend your time on literally anything else. Resolver, not authoritative, I know, but still. The optimization treadmill is real.
For heap under 100MB, sure. Scale to million-record zones with DNSSEC and tell me again. I've got graphs. The p99 latency spikes correlate with GC cycles at heap thresholds, not with query complexity. This is observable, reproducible, documented in Go runtime issues.
Your "smells like" is not data. My data is data.
Your data is from Contabo migrations with 12 incidents. Contabo's network stack is not RackNerd's. Their hypervisor scheduling is not RackNerd's. You're generalizing from one budget provider to all CoreDNS deployments.
Riga has four seasons. I have four monitoring systems. They disagree often.
I run Knot on two OVH VPS in Gravelines. No benchmark, just works. The zone transfers to secondary are faster than my old BIND setup, that's all I noticed. For 5 qps, who cares about qps.
