Guest and Telework resolve public DNS only, through a second Blocky instance that has no path to the internal zone
Date: 2026-09-25
Status: Superseded in part by docs/adr/2026-10-06-dnscrypt-proxy-refuses-the-internal-zones.md. Since 2026-10-06 the shared dnscrypt-proxy refuses every name under fzymgc.house, so Guest and Telework get REFUSED for the zone, its public names included. The second instance and its other rules stay in force.
Deciders: Sean Brandt
Supersedes: docs/adr/2026-08-28-blocky-per-client-view-withholds-the-whole-apex.md, for the Guest and Telework legs only
Context
Section titled “Context”The resolver pair sandiego/portland serves DNS to every VLAN. Blocky is the client tier, and
PowerDNS is the authoritative tier for fzymgc.house. Two legs are untrusted: Guest (VLAN 3100) and
Telework (VLAN 3200). The trusted: false attribute on their dns_vlan_interfaces entries says so.
The 2026-08-28 ADR gave every untrusted client the default client group of Blocky. That group
carries a denylist for the whole of fzymgc.house and both reverse zones. An untrusted client got
NXDOMAIN for every name under the apex. That included the public names id.fzymgc.house and
cluster-docs.docs.fzymgc.house, which a phone on a cellular network resolves without difficulty.
Issue #2302 asks for those legs to see useful names again.
Three facts limit how a single Blocky instance can do this:
- Only
blocking.clientGroupsBlockandupstreams.groupsvary per client. ThecustomDNSandconditionalsections are global, and both answer before any upstream group. - An allowlisted name therefore continues to
conditional.mapping, and PowerDNS returns the internal address. The Firewalla fence for each untrusted leg blocks that address. The client gets a correct-looking answer and a connection that times out. The 2026-08-28 ADR rejected this for that reason. fzymgc.houseis a real public zone in Cloudflare, with the public answers for its public names.
Decision
Section titled “Decision”Guest and Telework see only what public DNS returns, for every name. A second Blocky instance on each
resolver node, blocky-public, serves those two legs.
- The public instance forwards every name to the same Control D DoT upstream as the main instance,
except the special-use names it answers itself. Its configuration has no
conditionaland nocustomDNSsection, so it has no path to PowerDNS. No edit to the main instance or to PowerDNS can make it return an internal answer. - Its rebinding protection is on. Blocky drops any upstream answer that carries a private, ULA, loopback or link-local address and returns an empty NOERROR. An internal address cannot reach these legs, even if the public zone or the Control D profile points inward.
- The
trustedattribute selects the instance. The main instance listens on the addresses of the trusted legs. The public instance listens on the addresses of the legs withtrusted: false, and on127.0.0.3for its health check. No address is in both lists. An entry without the attribute fails the render. VI_guestandVI_teleworkstay outside thedns_vipssync group, for the reason in the keepalived template: an unauthenticated advert from an untrusted segment must not move the trusted VIPs. Each of the two instances trackschk_public_answeron its own.chk_public_answerasks the public instance forhealth.dns.invalidand requires NXDOMAIN. Blocky answers every name underinvalid.from its special-use table, with no upstream query. The check fails only when the public instance does not answer. An outage of the WAN, Control D or PowerDNS does not fail it.
The main instance keeps its internal-namespace denylist and its deny-by-absence default. No
untrusted leg reaches it through a listener, and a source that the configuration does not name
gets NXDOMAIN for the internal namespace. The other decisions of the 2026-08-28 ADR stay in force for the
main instance: the tailnet exemption, the exemption for the Firewalla forwarding address, and the
exemption and rate-limit allowlist entry for the health client.
Alternatives considered
Section titled “Alternatives considered”| Alternative | Verdict |
|---|---|
| A curated allowlist per leg in the main instance | Rejected. An allowed name reaches PowerDNS and gets the internal address, which the fences block. This is the silent failure the 2026-08-28 ADR named. |
| PowerDNS Views selected by EDNS Client Subnet | Not pursued. It needs Blocky to send client subnets to PowerDNS, and it adds a second internal view to maintain. |
Add VI_guest and VI_telework to dns_vips with chk_public_answer as a group tracker |
Rejected. With this, a device on Guest or Telework can step down every DNS VIP in the estate with one VRRP advert. |
Leave VI_guest and VI_telework untracked |
Rejected. With this, a dead public instance on the holder leaves both legs without DNS, and nothing fails over. |
| A check that resolves a public name | Rejected. It proves the upstream path too, but it fails on both nodes during a WAN or Control D outage, and a failover cannot help then. The accepted cost is that an upstream fault on one node only is not detected by the tracker. |
Consequences
Section titled “Consequences”A Guest or Telework client gets the public answer for every name. For id.fzymgc.house that is the
Cloudflare address. The public zone has wildcard MX records for *.fzymgc.house, so an A query for
an internal-only name returns NOERROR with no records, not NXDOMAIN.
The RFC 1918 reverse zones are answered NXDOMAIN locally by the special-use table. The ULA reverse zone is not in that table, so a ULA PTR query goes to Control D, which holds no internal records.
The Firewalla answers a Guest or Telework query that is not sent to the leg VIP, through its own per-leg dnsmasq. When the resolver file of that dnsmasq names the leg VIP, the query reaches the public instance from the leg gateway address. The DNS runbook gives the command that reads the file. All such clients share the rate-limit bucket of that address. On an untrusted segment the cap is the abuse control, so that address is not allowlisted.
A failure of the public instance on the VIP holder moves only VI_guest and VI_telework to the
peer. The trusted VIPs do not move, so the two nodes can each hold some VIPs. Before this change,
those two instances had no tracker.
Each node now runs two Blocky processes. They share the binary, the unit template, the upstream,
the cache settings and the rate limits. A leg that changes trust moves its addresses from one
instance to the other, so the role restarts the two as one chain. The public instance writes its
query log to the journal as blocky-public.service, and the collector ships it with the main
instance’s log.
References
Section titled “References”docs/engineering/specs/2026-09-25-untrusted-legs-public-dns-design.md: the design.ansible/roles/blocky/templates/blocky-public-config.yml.j2: the public configuration.ansible/roles/blocky/defaults/main.yml: the listener split and the public-instance values.ansible/inventory/group_vars/dns_hosts.yml: the health address and name both roles read.ansible/roles/keepalived/templates/keepalived.conf.j2: the per-instance tracker.ansible/roles/keepalived/templates/keepalived-check-public.sh.j2: the check.- Internal DNS runbook: the operational view.