This release fixes the x86_64 Linux binary of 0.25.1, which crashed with "Illegal instruction" on CPUs without AVX2. Forward mode also stops waiting on a slow upstream and learns which primary answers first, blocklists are fetched through the configured upstream, and the dashboard shows each upstream's answers and median time.
⚠️ Check before upgrading
A slow upstream no longer holds the query. After 200 ms without an answer, Numa also asks the next upstream and uses the first answer. On 2,000 names with Quad9 first and Cloudflare second, the p99 went from 869 to 476 ms for 7% more upstream queries. hedge_ms = 0 under [upstream] keeps strict list order. A single upstream gets no second query, so providers with a query quota see no extra queries. (#505)
The primary that answers first is asked first. Upstreams of the same kind as the first one (private address, plain, or encrypted) race: about 1 query in 20, at most one a minute, goes to two of them at once. One that answers first 7 times before it loses twice is asked first. For a fixed order, put the others in fallback. Thanks @bcookatpcsd (#493, #506)
🐛 Fixes
- The x86_64 Linux binary runs on CPUs without AVX2 again. 0.25.1 crashed on start, also in the amd64 Docker image and Homebrew on Linux. Each Linux binary now has to start on a baseline CPU in qemu (x86-64 v1, Cortex-A53, ARM1176) before it is published. Thanks @id4vip (#518, #519, #520)
- Blocklist hosts resolve through the configured upstream and cache. Before, they went to the bootstrap servers in clear text even with a DoH, DoT or ODoH upstream, and never refreshed on a network that blocks port 53. Thanks @Guara92 (#507, #509)
- On Windows, a reply that fails with os error 10022 after a VPN changes routes is sent again, so the client gets its answer. Thanks @HighPriest (#501, #508)
- The API answers 400 to an override or a blocking pause with a duration the clock can't represent. (#511)
📊 Dashboard
- In forward mode the Outbound Wire panel shows each upstream with its share of the answers and its median time, and
/statshas anupstreamslist. Before, every forwarded answer counted against the upstream asked first. Thanks @bcookatpcsd (#289, #517)
Also notable
- Records and DNSSEC canonical data are written by one encoder. In benchmarks, building signed data takes about 45% less time and DS checks 20% less. (#513, #514)
- The release workflow pins its actions to commit SHAs. (#510)
All changes
What's Changed
- perf(forward): Ask the next upstream after 200 ms without an answer by @razvandimescu in #505
- ci(release): Pin the release actions to commit SHAs by @razvandimescu in #510
- fix(api): reject durations the clock can't represent by @razvandimescu in #511
- refactor: reuse is_paused() and saturating_duration_since by @razvandimescu in #512
- fix(udp): resend unpinned then pinned when the reply source is rejected by @razvandimescu in #508
- refactor(record): one header and RDLENGTH path for every record type by @razvandimescu in #513
- refactor(dnssec): encode canonical RDATA with DnsRecord's writer by @razvandimescu in #514
- refactor(system_dns): share the process spawn helpers across platforms by @razvandimescu in #515
- fix(blocklist): resolve list hosts through the configured upstream by @razvandimescu in #509
- feat(dashboard): show the answers and median time of each upstream by @razvandimescu in #517
- perf(forward): Ask first the primary that answers first by @razvandimescu in #506
- fix(release): build x86_64 Linux binary for baseline CPUs by @razvandimescu in #519
- ci(release): smoke-test aarch64 and armv6 on their oldest CPUs by @razvandimescu in #520
Full Changelog: v0.25.1...v0.25.2