Document gated Minecraft cutovers and rollback

2026-10-21 13:04:53 +02:00
parent da2ede7ec8
commit 146f8841ab
2 changed files with 57 additions and 0 deletions
@@ -0,0 +1,18 @@
+# ATM10 cutover and rollback
+
+Current safety gate: `server.enabled=false`. The byobu process remains the sole
+writer of `/home/feedc0de/ATM10`. Chart Git tag `v6.6` is published as
+`minecraft-atm10` chart `6.6.0-0`.
+
+1. Ask the operator to type `stop` in the ATM10 console only. Confirm the Java
+   process has exited and record a file count and size for its `world/` tree.
+2. Change only `server.enabled` to `true` in `values.yaml`, commit, and wait for
+   Argo to synchronize. Confirm the new pod is on arschrock, UID/GID 1000,
+   `Recreate`, one replica, the unchanged host directory, and port 25566.
+   Test the dynamic LoadBalancer address and both direct arschrock addresses.
+3. Rollback: set `server.enabled=false` and wait until the pod is fully gone.
+   Only then restart the existing Java 21 launch from the same host directory:
+   `java @user_jvm_args.txt @libraries/net/neoforged/neoforge/21.1.224/unix_args.txt nogui`.
+   Do not let both runtimes write the same files.
+
+Do not stop the Paper server as part of this first cutover.
@@ -0,0 +1,39 @@
+# Paper cutover and rollback
+
+Current safety gate: `server.enabled=false` and `dynmap.route.enabled=false`.
+The byobu process remains the sole writer of `/home/feedc0de/mcserver` and the
+old host PostgreSQL `minecraft` database. Paper chart Git tag `v26.3-166` is
+published as chart `26.3.166-0`, with that exact JAR baked into the matching
+image. The new PostgreSQL has a retained 8 GiB PVC and a preliminary dump;
+its Secret remains cluster-managed, not in this public repository.
+
+After ATM10 has been cut over and checked:
+
+1. Ask the operator to type `stop` in the Paper console. Confirm Java has
+   exited, port 25565 is free, and no old PostgreSQL plugin connections remain.
+   Record `world*/` file counts/sizes and take a final `pg_dump -Fc` of the
+   host database. Keep the host database and previous dump untouched.
+2. Check the kube context, new namespace, new PVC UID, and database pod before
+   restoring the final dump **only into `paper-postgresql`**. Since the
+   preliminary dump is present, restore with `--clean --if-exists` and
+   `--single-transaction`. Confirm all nine table counts match the final dump.
+3. Enable `server.enabled`, but leave `dynmap.route.enabled=false`. Commit and
+   let Argo launch one pod on arschrock. The init container first copies each
+   changed plugin config to `.pre-kubernetes`, then changes the three PostgreSQL
+   hosts and Dynmap's bind address. Verify game port 25565, both direct
+   arschrock addresses, dynamic LoadBalancer address, plugin database access,
+   and the internal `paper-dynmap` Service.
+4. In one GitOps change, enable the Paper chart's Dynmap route and remove the
+   old `default/mcmap` route and its middleware/legacy Service definitions.
+   Verify `https://brunner.ninja/mcmap` serves the new Dynmap before deleting
+   the manually managed legacy EndpointSlice. Move the old source manifest
+   under `traefik/legacy-services/_deleted/`. Dynmap needs no MikroTik forward.
+5. Add and verify Grafana panels for player count, CPU, memory, network, and
+   executing node after both services are healthy.
+
+Rollback requires disabling the Kubernetes game Deployment and waiting for its
+pod to terminate **before** restarting byobu. If plugins wrote to the new
+database, export/reconcile those writes into the old database first; starting
+against the stale old database loses data. Restore the `.pre-kubernetes`
+configuration files only once the pod is stopped. Keep the new PVC and both
+database dumps for recovery.