- Add mcintire.me zone + all 14 DNS records to tofu/cloudflare.tf (A, MX, CNAME, SRV, TXT)
- Include zone settings, HSTS, bot management, firewall for mcintire.me
- Update Porkbun NS from as393837_ns to cloudflare_ns in tofu/dns.tf
- Remove mcintire.me zone from BIND9 master config (group_vars/bind9_servers/)
- Delete mcintire.me.yml zone file
- Remove mcintire.me secondary zone from git.mcintire.me host_vars
- imagePullPolicy: Always so :latest actually tracks upstream (nodes were
running two different cached digests)
- --retries 1: supervisor retries forever; this caps the jittered backoff
at ~2s instead of the exponential 1..32s ramp
- liveness failureThreshold 30: let cloudflared's own fast retry loop
reconnect instead of killing pods into exponential CrashLoopBackOff
(pods had 300+ restarts)
- maxSurge 0 / maxUnavailable 1: required anti-affinity on 3 workers
deadlocked surge-based rollouts
- securityContext for restricted:latest PodSecurity (nonroot 65532,
drop ALL caps, no privilege escalation, RuntimeDefault seccomp)
Co-Authored-By: Claude <noreply@anthropic.com>
- so-reuseport: yes — kernel hashes UDP across the per-thread sockets
instead of waking all threads, balancing load and improving headroom
- edns-buffer-size: 1232 — DNS Flag Day 2020 default, avoids IP
fragmentation (retransmits/TCP fallback)
- serve-expired (RFC 8767) with 1800ms client timeout and 86400s stale
ttl — fast answers from stale cache during upstream slowness/outages
The packaged telegraf.service sets PrivateMounts=true, which requires a
mount namespace that unprivileged LXC containers cannot create, so the
service failed to start on the LXC resolver with status 226/NAMESPACE.
Drop in PrivateMounts=false, gated to container guests so VM hosts keep
the unit's hardening.
Add a telegraf role (applied to the resolvers) that runs unbound-control
via the inputs.unbound plugin and exposes the stats as a Prometheus
/metrics endpoint on :9273, firewalled to the 204.110.188.0/22 scrape
network. Imports the InfluxData signing key (influxdata-archive.key; the
_compat key expired 2026-01-17) so dnf can validate the package.
- prom.w5isp.com: scrape job "unbound" for resolver1/2:9273
- grafana: "Unbound Resolvers" dashboard (queries, cache hit ratio,
rcodes, query types, recursion time, request list, cache memory)
- resolvers: disable unused shm-enable to silence the startup warning
The 'Validate sshd configuration' task had ansible.builtin.command with no
command, so it always errored (one of the following is required:
_raw_params, cmd, argv), tripping the rescue that reverts the drop-in. The
post-quantum KexAlgorithms hardening therefore never applied. Run sshd -t
to validate before restart as intended.
Add an RPZ blocklist on the Unbound resolvers that returns NXDOMAIN for a
set of domains (and subdomains) and logs only those matches under rpz-log
tag "blocklist". Domains are templated from resolver_blocked_domains
(single source of truth) into /etc/unbound/blocklist.rpz; respip module
enabled for RPZ support.
Fix deprecation warnings across the Ansible tree:
- apt_repository -> deb822_repository (monitor, aprsc, uisp, grafana),
removing stale .list files and adding update_cache where deb822 drops it
- community.mysql.* -> ansible.mysql.* (monitor) and requirements.yml
- top-level facts -> ansible_facts[...] (ansible_os_family,
distribution_release, fqdn, architecture, default_ipv4) repo-wide
- Add cloudflare_bot_management for all zones with is_robots_txt_managed=false
so each app's own robots.txt is served instead of Cloudflare's managed one
- Fix zone_setting id zero_rtt -> 0rtt (invalid id rejected by API)
- Fix security_header value: object instead of jsonencode (avoids perpetual diff)
- Fix firewall expression http.request.user_agent -> http.user_agent (invalid field)
- tofu fmt reindentation
- Remove duplicate firewall include_role from base role (#18)
- Move remote_user from [ssh_connection] to [defaults] in ansible.cfg (#20)
- Add set -euo pipefail to run.sh (#21)
- Delete orphaned playbook_handlers.yml (#22)
- Fix changed_when: true in debian tasks (#23)
- Add tailscale_authkey env var lookup with TAILSCALE_KEY fallback (#25)
- Document -K requirement in bootstrap.yml (#26)
- Replace bare ansible_host=mail/git with IP addresses (#27)
- Update findings.md
- Add |default() fallbacks for ntpserver, template_run_date,
deb_mirror, and sshd.* to prevent template crash
- Remove include_role: name=alpine from base tasks (no such role)
- Fix SSH -t flag placement in Makefile reboot targets
- Update findings.md
- Convert ~290 short module names to FQCNs across all roles
- Change become: yes/no to become: true/false, gather_facts: no to false
- Remove deprecated inject_facts_as_vars from ansible.cfg
- Add missing task names to import_tasks/include_role calls
- Add become: yes to librenms clone task (partial-become fix)
- Add pipefail to risky shell command in udp-gro-fix.yml
- Quote all octal mode values
- Harden .ansible-lint with production profile and empty skip_list
- Update findings.md with all fixes
- Remove forgejo role, play, group, host_vars, and DNS references
- Replace hardcoded /Users/graham/.ssh/ paths with env var lookup
- Remove broken setup-user Makefile target
- Add community.mysql collection to requirements.yml
- Add .ansible/ to .gitignore
- git rm --cached .DS_Store files and .vscode/settings.json
- Update CLAUDE.md and README.md to reflect actual state
- Add findings.md with audit results
- unbound: add loopback DoH listener (127.0.0.1@8080, http-notls-downstream);
interface-automatic-ports lists 8080 so interface-automatic does not silently
drop the non-53 listener; access-control driven by resolver_allowed_netblocks
- caddy: add RedHat/COPR install path, caddy_template var, /var/log/caddy dir;
new Caddyfile-resolver.j2 terminates LE TLS and h2c-proxies to unbound,
enforcing the client allow-list via remote_ip
- resolvers: open 80/443 in firewall; prefer post-quantum SSH key exchange
(sntrup761x25519-sha512) via validated sshd_config.d drop-in
- bind9: zone.j2 emits AAAA records; add ipv6 for resolver1/2 (2606:1c80::240/250)
- playbook: run caddy role on the resolvers group
Listen on ::0, set do-ip6: yes for IPv6 upstream resolution, and add
IPv6 access-control: default-deny ::0/0, allow ::1 and the VNTX
2606:1c80::/32 allocation.
Without this, a renamed dashboard left the old file on disk and the
provisioner refused to write any dashboards because of a duplicate UID
(seen with towerops-phoenix-live-view vs towerops-phoenix_live_view).
