personal_infra/ansible/services/fulcrum/deploy_fulcrum_playbook.yml

34 lines
1.1 KiB
YAML
Raw Normal View History

fulcrum: convert to a role, de-Uptime-Kuma the health check 685-line playbook becomes 33 lines plus a 426-line role (install/service/healthcheck phases, six templates, one handler). fulcrum_vars.yml is deleted; its content is the role's defaults. Verified: fulcrum untouched - active since 2026-07-29 (no restart), 192G datadir, height 966842, bitcoind and db_mem unchanged on disk. Second run changed=1 (the arming run, changed_when: false aside). vipy changed=0. THREE PRE-EXISTING LANDMINES the check-mode diff caught, any of which a faithful extraction would have detonated: - bitcoin_rpc_host was "192.168.1.140", commented "IP of knots_box_local". But .140 is fulcrum-box ITSELF; knots-box is .135. The DHCP leases had reshuffled - the fifth instance of this same disease in this estate. The live config had been hand-corrected to knots-box; running the playbook would have reverted it and pointed Fulcrum at itself. Now addressed by Tailscale name. - The `Restart fulcrum` handler was guarded by uptime_kuma_enabled, so the three tasks that notify it (SSL cert, fulcrum.conf, systemd unit) could not restart anything. A config change applied to disk, reported success, and silently never took effect. That is worse than the other banner casualties: it makes the deployment itself lie. Ungated. - db_mem was about to go 2048 -> 4448 (75% of 5931MB RAM), leaving ~1.4GB for the OS and Fulcrum's non-cache memory. The live value had been hand-tuned down. fulcrum_db_mem_mb_override pins it. Note set_fact outranks role defaults, so the calculation itself has to honour the override. MY OWN ERROR, third instance: retyping `copy:` as `template:` lost `owner:` on the banner and on fulcrum.conf. Rather than keep catching these by eye, every managed path's owner/group/mode is now compared against `git show HEAD:` mechanically - 12/12 match. The health check timer had not fired since 2026-02-17 while reporting `active` and `enabled`. It is OnBootSec + OnUnitActiveSec with no OnCalendar: OnBootSec elapses once, and OnUnitActiveSec needs the SERVICE to have run this boot to have anything to schedule from. Restarting the timer does not supply that; running the service does, so the role now runs the check once after enabling. Also dropped `Requires=fulcrum.service` from the timer - on a timer that means "stop watching when the watched thing stops". Diagnostic note: NextElapseUSecRealtime is always empty for a monotonic timer, so it reads as broken even when healthy. I misread it once and wrongly called the timer dead. Use NextElapseUSecMonotonic or systemctl list-timers. fulcrum_ssl_port and fulcrum_tailscale_hostname moved to services_config.yml - the socket-proxy play on the edge host needs them and a role default cannot reach a second play. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-13 18:02:00 +02:00
---
# Fulcrum: Electrum server indexing the Bitcoin Knots node.
# The index takes days to rebuild, so nothing here touches its data directory
# beyond asserting that it exists.
2025-12-14 18:52:36 +01:00
- name: Deploy Fulcrum Electrum Server
hosts: electrum
2025-12-14 18:52:36 +01:00
become: yes
vars_files:
- ../../infra_vars.yml
- ../../services_config.yml
- ../../infra_secrets.yml
vars:
fulcrum: convert to a role, de-Uptime-Kuma the health check 685-line playbook becomes 33 lines plus a 426-line role (install/service/healthcheck phases, six templates, one handler). fulcrum_vars.yml is deleted; its content is the role's defaults. Verified: fulcrum untouched - active since 2026-07-29 (no restart), 192G datadir, height 966842, bitcoind and db_mem unchanged on disk. Second run changed=1 (the arming run, changed_when: false aside). vipy changed=0. THREE PRE-EXISTING LANDMINES the check-mode diff caught, any of which a faithful extraction would have detonated: - bitcoin_rpc_host was "192.168.1.140", commented "IP of knots_box_local". But .140 is fulcrum-box ITSELF; knots-box is .135. The DHCP leases had reshuffled - the fifth instance of this same disease in this estate. The live config had been hand-corrected to knots-box; running the playbook would have reverted it and pointed Fulcrum at itself. Now addressed by Tailscale name. - The `Restart fulcrum` handler was guarded by uptime_kuma_enabled, so the three tasks that notify it (SSL cert, fulcrum.conf, systemd unit) could not restart anything. A config change applied to disk, reported success, and silently never took effect. That is worse than the other banner casualties: it makes the deployment itself lie. Ungated. - db_mem was about to go 2048 -> 4448 (75% of 5931MB RAM), leaving ~1.4GB for the OS and Fulcrum's non-cache memory. The live value had been hand-tuned down. fulcrum_db_mem_mb_override pins it. Note set_fact outranks role defaults, so the calculation itself has to honour the override. MY OWN ERROR, third instance: retyping `copy:` as `template:` lost `owner:` on the banner and on fulcrum.conf. Rather than keep catching these by eye, every managed path's owner/group/mode is now compared against `git show HEAD:` mechanically - 12/12 match. The health check timer had not fired since 2026-02-17 while reporting `active` and `enabled`. It is OnBootSec + OnUnitActiveSec with no OnCalendar: OnBootSec elapses once, and OnUnitActiveSec needs the SERVICE to have run this boot to have anything to schedule from. Restarting the timer does not supply that; running the service does, so the role now runs the check once after enabling. Also dropped `Requires=fulcrum.service` from the timer - on a timer that means "stop watching when the watched thing stops". Diagnostic note: NextElapseUSecRealtime is always empty for a monotonic timer, so it reads as broken even when healthy. I misread it once and wrongly called the timer dead. Use NextElapseUSecMonotonic or systemctl list-timers. fulcrum_ssl_port and fulcrum_tailscale_hostname moved to services_config.yml - the socket-proxy play on the edge host needs them and a role default cannot reach a second play. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-13 18:02:00 +02:00
# Preserves the push URL this check has been configured with. The role knows
# nothing about Uptime Kuma — this is just "a URL that accepts a ping".
healthcheck_push_url: "{{ healthcheck_push_urls.fulcrum | default('') }}"
roles:
- fulcrum
2025-12-24 10:27:35 +01:00
- name: Setup public Fulcrum SSL forwarding on the edge host
hosts: edge
2025-12-24 10:27:35 +01:00
become: yes
vars_files:
- ../../infra_vars.yml
- ../../services_config.yml
tasks:
bitcoin-knots, fulcrum, datum-gateway: add and use the socket_proxy role Three near-identical hosts: edge plays become one role plus three short calls. 183 lines removed, 34 added, plus a 111-line role. Verified before touching any playbook: all six live units on vipy reproduced byte-identically. Then --limit edge --check per playbook - bitcoin-knots and fulcrum changed=0; datum-gateway changed=2, both attributable to the already known caddy_site comment line and the Reload caddy handler it triggers. The 6 units and 14 Caddy files on the hosts are byte-identical afterwards. PLAN_4 claimed these three plays had "no behavioural drift at all". That was wrong - it came from a diff truncated by head -60. The live bitcoin-p2p-proxy units carry four settings this playbook never wrote: .socket Documentation=, FreeBind=true .service Documentation=, TimeoutStopSec=5, StandardOutput=journal, StandardError=journal FreeBind is the one that matters: it lets the socket bind to an address that is not up yet, so without it the socket can fail to start on boot. Running the bitcoin-knots playbook would have stripped it. Same class of hazard as headscale. The role expresses all four; bitcoin-p2p is the only caller that passes any. Also: UFW treats the rule comment as part of the rule. datum-stratum's live comment is "DATUM Gateway Stratum public access" but the role's derived default produced "DATUM Stratum public access", which rewrote the rule. Caught in the dry-run; datum now passes the comment explicitly. Two deliberate differences from the original, both documented in the README: ignore_errors: yes on the upstream check became failed_when: false, and the handler restarts the .socket, which drops connections open through it - it fires only when a unit file actually changes. The inert Uptime Kuma TCP monitor blocks stay in the playbooks rather than being pulled into a new role (12/12/18 guarded tasks). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-11 23:50:15 +02:00
- name: Expose Fulcrum SSL through a socket proxy
ansible.builtin.include_role:
name: socket_proxy
vars:
socket_proxy_name: fulcrum-ssl
socket_proxy_description: "Fulcrum SSL"
fulcrum: convert to a role, de-Uptime-Kuma the health check 685-line playbook becomes 33 lines plus a 426-line role (install/service/healthcheck phases, six templates, one handler). fulcrum_vars.yml is deleted; its content is the role's defaults. Verified: fulcrum untouched - active since 2026-07-29 (no restart), 192G datadir, height 966842, bitcoind and db_mem unchanged on disk. Second run changed=1 (the arming run, changed_when: false aside). vipy changed=0. THREE PRE-EXISTING LANDMINES the check-mode diff caught, any of which a faithful extraction would have detonated: - bitcoin_rpc_host was "192.168.1.140", commented "IP of knots_box_local". But .140 is fulcrum-box ITSELF; knots-box is .135. The DHCP leases had reshuffled - the fifth instance of this same disease in this estate. The live config had been hand-corrected to knots-box; running the playbook would have reverted it and pointed Fulcrum at itself. Now addressed by Tailscale name. - The `Restart fulcrum` handler was guarded by uptime_kuma_enabled, so the three tasks that notify it (SSL cert, fulcrum.conf, systemd unit) could not restart anything. A config change applied to disk, reported success, and silently never took effect. That is worse than the other banner casualties: it makes the deployment itself lie. Ungated. - db_mem was about to go 2048 -> 4448 (75% of 5931MB RAM), leaving ~1.4GB for the OS and Fulcrum's non-cache memory. The live value had been hand-tuned down. fulcrum_db_mem_mb_override pins it. Note set_fact outranks role defaults, so the calculation itself has to honour the override. MY OWN ERROR, third instance: retyping `copy:` as `template:` lost `owner:` on the banner and on fulcrum.conf. Rather than keep catching these by eye, every managed path's owner/group/mode is now compared against `git show HEAD:` mechanically - 12/12 match. The health check timer had not fired since 2026-02-17 while reporting `active` and `enabled`. It is OnBootSec + OnUnitActiveSec with no OnCalendar: OnBootSec elapses once, and OnUnitActiveSec needs the SERVICE to have run this boot to have anything to schedule from. Restarting the timer does not supply that; running the service does, so the role now runs the check once after enabling. Also dropped `Requires=fulcrum.service` from the timer - on a timer that means "stop watching when the watched thing stops". Diagnostic note: NextElapseUSecRealtime is always empty for a monotonic timer, so it reads as broken even when healthy. I misread it once and wrongly called the timer dead. Use NextElapseUSecMonotonic or systemctl list-timers. fulcrum_ssl_port and fulcrum_tailscale_hostname moved to services_config.yml - the socket-proxy play on the edge host needs them and a role default cannot reach a second play. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-13 18:02:00 +02:00
socket_proxy_listen_port: "{{ service_settings.fulcrum.ssl_port }}"
socket_proxy_upstream_host: "{{ service_settings.fulcrum.tailscale_hostname }}"