Nothing watched the servers: tonight's asxpio crash loop was found by
hand. On asxpio-ops, Prometheus (30 days), Loki (14 days) and Grafana
at gf.asxp.io (WireGuard only), with blackbox_exporter probing every
service from outside and postgres_exporter. On both hosts node_exporter
and cAdvisor on the VPC address, and Alloy shipping container logs, the
journal and Traefik's access log to Loki over the VPC.
Traefik and Forgejo expose their metrics; restic and the bucket replica
record their last success for node_exporter. Dashboards: an overview as
home, plus hosts, containers, Traefik, probes, Postgres and logs.
Alerts by email to me@asxp.io: host or service down, certificate under
14 days, disk or memory nearly full, a container restart-looping,
Postgres down, backup or replica stale.
Assisted-by: Claude:opus-5-5
A dry run changes nothing to verify, and for a service that is not on
the host yet the check waited a minute and failed the dry run.
Assisted-by: Claude:opus-5-5
Same pattern as traefik.asxp.io: public records to asxpio-ops, so
Traefik gets a Let's Encrypt certificate over HTTP-01, while an IP
allowlist limits it to WireGuard.
Assisted-by: Claude:opus-5-5
The interop project kept a cx23 (claude-test-interop-runner, about
EUR 6/month with its IPv4) only to run its Forgejo runner. Run it here
instead, with the same registration (UUID + token, token in the vault)
and the same three labels, pinned.
Its PR checks are untrusted, so unlike our own runner its jobs get
neither the Docker socket nor host volumes. nix-host, which on the old
host used the host's Nix store as a cache, is a throwaway container for
now; a second, trusted-only registration can bring the cache back.
Rollback: remove the role from site.yaml (or stop the container) and
start the old runner VM again.
Assisted-by: Claude:opus-5-5
Under runner 13 the run that recreates the runner ends as failed, not
cancelled (run 450); say so, and where its log stops, so it is not
mistaken for a real failure.
Assisted-by: Claude:opus-5-5
Floating tags (17-alpine, 3.4-slim, nginx:alpine, :latest, :debug, @v7)
change under a deploy without anyone deciding it: the storage deploy
moved Postgres to 17.12 on its own because of postgres:17-alpine. Pin
every image and action to its exact current version, so each change
arrives as a reviewed Renovate PR and a rollback is a revert.
Assisted-by: Claude:opus-5-5
Renovate now runs from its own repository for every repo, with the
shared rules in its preset. Extend that preset here and keep only what
is specific to this repo (the Ansible var manager, the Paper JRE, the
runner and nix notes); drop the workflow that ran it for iac alone.
Assisted-by: Claude:opus-5-5
Both recreate the runner mid-job: the applying run shows as cancelled,
leaves its job container behind, and needs a later run to confirm.
Write the follow-up steps down where the Renovate PRs send you.
Assisted-by: Claude:opus-5-5
The unquoted value contains ": ", which YAML reads as a nested mapping,
so Forgejo rejected the whole workflow and the last two pushes ran
nothing. Quote it.
Assisted-by: Claude:opus-5-5
The server is stopped and its data deleted (archived to object storage
as minecraft-fossils-final-20261112.tar.zst), so the role that kept it
stopped has nothing left to do.
Assisted-by: Claude:opus-5-5
Every run was listed as "deploy", so dry runs on Renovate branches
looked like deployments. Name each run after what it does and the
branch it ran on.
Assisted-by: Claude:opus-5-5
The repository was renamed from infrastructure to iac. Renovate looks
repositories up by name, and with push-to-create on, the old name now
belongs to a different, empty repository.
Assisted-by: Claude:opus-5-5
restic writes both regions itself, but the ad-hoc backups under
asxpio-ops/, the Terraform state and the k8s state only ever went to
hel1. Sync each hel1 bucket into its fsn1 twin every morning from
asxpio-services, with rclone pinned like everything else.
restic/ is excluded, so the independent fsn1 restic repositories are
never touched. A run that would delete more than 1000 objects stops and
mails, so a mistake on the hel1 side cannot empty the replica the next
morning. The first dry run copied 23 objects and deleted none.
Assisted-by: Claude:opus-5-5
The mounts were only ever read by one-off migration scripts, and their
cache had grown to 9.7 GB. restic and the bucket replica talk to object
storage directly, so nothing needs a filesystem view of the buckets.
Disable s3fs like on asxpio-ops; the bootstrap playbook unmounts and
drops the fstab lines (applied by hand, cache removed).
The playbook still created Immich directories under the mount and its
summary needed the cache size, which a disabled host no longer sets.
Assisted-by: Claude:opus-5-5
The zones were edited by hand in the console, and it showed: dozens of
records for servers that no longer existed, a CNAME and a TXT record
missing their trailing dot, and a 24 h default TTL that made today's
XMPP move linger in resolver caches.
Now both zones live in the asxpio project, so describe them here,
adopted with import blocks (the first apply imported 87 and changed
nothing but protection). Our servers' records reference hcloud_server
instead of literal IPs. Every record set is change-protected except
those of the test XMPP servers, and both zones are delete-protected.
Assisted-by: Claude:opus-5-5
Immich has not run for months and its photos are now in Google Photos.
asxpio-immich-main and asxpio-immich-repl (about 140 GB each) are
deleted; remove them from Terraform (state too) and from the s3fs
mounts, so nothing recreates or remounts them. The mount on
asxpio-services was taken down and its fstab line removed by hand.
Assisted-by: Claude:opus-5-5
/opt/vhpbot-prod/bot.db is an empty directory that Docker created for a
bind mount whose file never existed; the bot keeps no database there.
The hook could only fail and mark every nightly run as failed.
Assisted-by: Claude:opus-5-5
Versions are pinned across the Ansible vars, compose files, Terraform
providers and workflows, and nothing says when one is behind. Run
Renovate daily from Forgejo Actions: one PR per available bump, with
the upstream changelog in the body, which the deploy pipeline then
dry-runs like any other branch.
Variables are tracked through a "# renovate: datasource=... depName=..."
comment above them; Forgejo, Traefik, the runner and its nix image get
one here. The Paper JRE tag is left alone, the runner and nix bumps
carry their manual follow-up steps in the PR, and Renovate's own image
is bumped monthly.
Needs the repository secrets RENOVATE_TOKEN (bot user with write
access) and GH_COM_TOKEN (read-only github.com token for release
notes).
Assisted-by: Claude:opus-5-5
Describe the new roles, how to list and restore restic snapshots, and
that the restic password must be kept outside the servers.
Assisted-by: Claude:opus-5-5
Open the ports of the XMPP server and Xray, which move to
asxpio-services, and close 25566 with Minecraft Fossils gone.
Turn off Hetzner's server backups on both hosts: restic now backs up
the data to object storage in two regions, and the server backups cost
20% of each server's price. Disabling them deletes the existing backup
images, so apply that part only after restic has a verified snapshot.
Assisted-by: Claude:opus-5-5
The VLESS Reality proxy ran from a hand-installed binary on the
temporary server. Run the upstream image instead, pinned and on the
host network, with only CAP_NET_BIND_SERVICE; its config (keys and
client id) is in the vault.
Assisted-by: Claude:opus-5-5
ejabberd, apushd and the tackygram gateway ran on the temporary
claude-test-xmpp server, a separate CX23 billed on its own. Move them
to asxpio-services, which has the room now that Fossils is gone, and
manage them like the other stacks.
Compose follows anotherim-server's compose.yaml and
compose.tackygram.yaml with CERTS=files; image versions are variables
so they can be tracked. The TURN and call addresses come from the
host's public IP instead of a hand-edited env file. Secrets, keys and
the certificate stay on the host under /opt/xmpp/state as data and are
covered by the restic backup.
Assisted-by: Claude:opus-5-5
In check mode nothing is written, so for a service that is not on the
host yet the restart task finds no project directory and fails the
whole dry run. Ignore errors there in check mode, as the compose task
already does.
Assisted-by: Claude:opus-5-5
The Fabric server holds 8.8 GB of asxpio-services' 15 GB of RAM, more
than everything else on the host together, and the XMPP server moves
here next. Save the world, remove its containers and the daily restart
timer.
The world stays in /opt/minecraft-fossils/data, and in the restic
backups, until it is deleted by hand.
Assisted-by: Claude:opus-5-5
Hetzner's server backups cost 20% of each server's price and only
restore whole disks. Back up what matters with restic instead, into the
backup buckets we already pay for: one repository per host in hel1
(asxpio-backups-main) and one in fsn1 (asxpio-backups-repl), written
independently so either region alone is a full copy.
Postgres goes in as a dump, Forgejo's and vhpbot's SQLite databases
through the online backup API, Garage with a fresh metadata snapshot,
ejabberd through its own backup command. Keeps 7 daily, 4 weekly and
6 monthly snapshots, prunes on Sundays, and mails the journal when a
run fails.
restic is the pinned upstream release, so its version is tracked in
the repo like everything else. The vault gains the restic password
and the Xray config for the role that follows.
Assisted-by: Claude:opus-5-5
The cost table was a year old. Use the amounts billed now.
Both servers still cost less than the list price for their type
(3.99 against 5.49, 11.99 against 15.99). That price belongs to the
existing server, so say plainly that a reinstall has to happen in
place.
Assisted-by: Claude:fable-5-2
With provider 1.54 the plan replaces both servers. Hetzner stopped
returning a server's datacenter, the old provider derived the location
from it, and a location that reads back empty differs from "hel1",
which forces replacement. Only prevent_destroy stood in the way.
Provider 1.70.2 reads the location properly: the same plan then shows
the two firewall updates and nothing else. The lock file is not
tracked, so raise the version constraint to keep the old provider out.
Run terraform init -upgrade once in existing checkouts.
Assisted-by: Claude:fable-5-2
Checked against the Hetzner project:
- Both servers have daily backups and delete and rebuild protection
switched on, while the code said backups = false and set no
protection. An apply would have turned all three off and discarded
the existing backups. Record them as they are.
- The firewalls had drifted both ways. Live, asxpio-ops also accepts
XMPP ports and tcp/5000, and asxpio-services accepts 80, 443, XMPP,
TURN, Valheim and Factorio ports; none of those services run. The
code in turn missed tcp/25566, where the second Minecraft server
listens.
Describe the firewalls as they should be: internal networks, plus the
public ports of services that run. Public SSH to asxpio-services goes
too; sshd and iptables already refuse it from outside the VPN.
This is not applied. The next plan will show the stale ports being
closed and nothing else.
Assisted-by: Claude:fable-5-2
The prune job now lives in /etc/cron.d on both hosts and root's crontab
entry is gone, so the tasks that removed it have nothing left to do.
Assisted-by: Claude:fable-5-2
grep -v exits 1 when it prints nothing, and with pipefail that failed
the task on both hosts, although the crontab had been rewritten
correctly. Ignore grep's exit status.
Assisted-by: Claude:fable-5-2
The READMEs still tracked the 2025 migration from bare metal and listed
Grafana, Immich, Valheim, Prosody and others that do not run. Replace
them with what is deployed, how a change reaches production and where
the secrets live.
Assisted-by: Claude:fable-5-2
Service changes were made by hand on the servers, which is how the
repository drifted. Apply site.yaml from a workflow instead: every
push dry-runs it against production, and a push to main applies it.
Jobs run in containers on asxpio-services. asxpio-ops is reached over
the private network; the services host is reached through asxpio-ops,
because a container connecting to its own host arrives from a Docker
bridge address that sshd and the firewall refuse. Host keys are pinned
in known_hosts.
Ansible comes from a pinned nixpkgs revision so every run uses the
same version.
Assisted-by: Claude:fable-5-2
The roles had drifted far from production: Forgejo was pinned at 12
while 15.0.11 runs, its config lacked the mailer, packages and the real
database path, and several services existed only on the hosts. Running
a role would have rolled production back. Take the servers as the truth
and make the repository match them.
site.yaml now holds the desired state of both hosts:
- base: unattended upgrades, fail2ban, Docker daemon config and the
weekly image prune, as found on the hosts. New: a 500M journal cap,
and the Docker systemd override is kept absent. That override once
opened dockerd on tcp/2375 without authentication.
- traefik: templates as deployed, version in a variable (v3.7.18).
The Coraza WAF files are removed; no router ever used them.
- forgejo: 15.0.11, app.ini generated from the live file with the
secrets taken from the vault.
- crawler_alert, sillytavern, minecraft, minecraft_fossils: added
from the hosts.
Forgejo's canonical name becomes git.asxp.io; git.czsk.it stays as an
alias on the same router.
docker_service restarts a stack when its config changes, no longer
forces ownership on existing data directories, and hides template
diffs because they carry secrets. Containers listed in
service_graceful_stop are saved and stopped with a long timeout before
compose recreates them: Docker's 10 s default is too short for a
Minecraft world. Both game servers get stop_grace_period for the same
reason.
Removed: the full-deployment and runner playbooks with their prompts,
the Forgejo backup and restore roles for the old bare-metal host, and
inventory entries for services that were never deployed. SSH hardening
stays a manual playbook, since a mistake there locks everyone out.
New secrets (SMTP, RCON, CurseForge key) went into the vault.
Assisted-by: Claude:fable-5-2
hcloud replaces a server when user_data, image or ssh_keys change, and
cloud-init only runs at first boot anyway. An edit to cloud-init, like
adding a user, would therefore plan the destruction of a production
machine. Ignore changes to those three attributes and set
prevent_destroy on both servers, so cloud-init describes a rebuild and
nothing else.
Add user zerrtika to cloud-init to match the servers; the user was
created on them by playbook 04.
Assisted-by: Claude:fable-5-2
The role had drifted from what runs on asxpio-services: capacity 4,
a bridge network per workflow so service containers start, the Docker
socket automounted into jobs, and a nix-latest label backed by a
persistent nix-store volume.
That volume caused a CI outage on 2026-10-18. nixos/nix:latest moved on,
and since the image's /bin/sh and /etc/nix/nix.conf link into /nix/store,
the old volume left them dangling. Pin the image through
forgejo_runner_nix_version and name the volume after it, so a bump is
deliberate and always gets a store seeded from its own image.
Assisted-by: Claude:opus-5-5
The ops server has 3.7 GB RAM and no swap. Forgejo, MinIO and the apps
fit with room to spare, but a spike (a big push, git gc, parallel
image pulls) would OOM-kill a service instead of paging out.
swappiness 10, so it only swaps under real pressure.
Assisted-by: Claude:opus-5
On the 38 GB ops disk the s3fs cache reserve (ensure_diskfree) never
fit, so all three mounts failed at every boot since 2026-09-22. Backups
now go to Hetzner S3 with rclone instead.
Add s3fs_enabled (default true); when false the role unmounts the
buckets and drops their fstab entries, keeping the credentials file
that the rclone wrapper reads.
Assisted-by: Claude:opus-5
