vps cluster docs

Storage

This page describes the live storage layout after the 2026-10-27 vcap v1 retirement. Verify live state before you rely on any number here.

Live data

Dataset	Namespace	Backing	Size	Backup
Postgres `gitorgasm`	`gitorgasm`	static PV on hostPath `/srv/data/gitorgasm-postgres`, `Retain`	not measured	class A logical dump
Postgres `vcap-v1` (retired, retained)	`vcap-v1`	static PV `vcap-v1-postgres-data-vps`, hostPath `/srv/data/vcap-v1-postgres`, `Retain`	67 MB	class C raw
Final dump `vcap-v1` (retired, retained)	none	host directory `/srv/data/vcap-v1-final-dump`, file `vcap-v1-pg_dumpall-20261027.sql.gz`	not measured	class C raw
Application data `vcap-v1` (retired, retained)	`vcap-v1`	local-path PV `pvc-d4b91391-4691-43c4-9590-6bbe1ddf0ad9`, `Retain` since 2026-10-27	15 MB	class C raw
Postgres and MinIO `vcap-staging`	`vcap-staging`	static PV `vcap-backend-staging-data-vps`, hostPath `/srv/data/vcap-staging`, `Retain`	75 MB	class A raw
Postgres and MinIO `vcap-dev`	`vcap-dev`	static PV `vcap-backend-dev-data-vps`, hostPath `/srv/data/vcap-dev`, `Retain`	70 MB	class C raw
Postgres `meghmitra`	`meghmitra`	static PV `meghmitra-postgres-pvc-vps`, hostPath `/srv/data/meghmitra-postgres`, `Retain`	206 MB	class C raw
Postgres and MinIO `bitvault`	`bitvault`	local-path PVs `pvc-4f59ae33…` and `pvc-acbed98f…`, `Retain`	48 MB + 148 KB	class C raw
Mongo `rankstack`	`rankstack`	local-path PV `pvc-fb2fad5e…`, `Retain`	509 MB	class C raw
Mock service data `kirro`	`kirro`	local-path PV `pvc-623ce531…`, `Delete`	76 MB	class C raw
Voice logs `kirro`	`kirro`	local-path PV `pvc-905a8d2c…`, `Delete`	456 KB	class C raw
SQLite and model data	`adhigrahan-radar`	static PV `adhigrahan-radar-data-pvc-vps`, hostPath `/srv/data/adhigrahan-radar-data`, `Retain`	23 MB	class C raw plus consistent `vivaad.db` copy
Corpus	`meghmitra`	local-path PV `pvc-a5b3a672…`, `Retain`	4 KB	none
Grafana	`monitoring`	PV `pv-grafana`, hostPath `/var/lib/k8s-monitoring/grafana`, `Retain`	52 MB	class C raw plus consistent `grafana.db` copy

The vcap-staging, vcap-dev, vcap-v1, bitvault, rankstack, and kirro workloads are gone. Their namespaces, PVCs, PVs, Secrets, and data remain. Their Postgres and Mongo trees are static. Do not start a server on a retained tree. Copy it first. See the restore runbooks.

The meghmitra Postgres workload is also gone. The data remains. The adhigrahan-radar workload replaced the old meghmitra application and is live.

Namespace vcap-v1 also keeps these objects:

Secrets auth-env, backend-env, pull-secret, and vcap-v1-postgres-edge-tls.
ConfigMaps vcap-backend-v1-auth-config, vcap-backend-v1-backend-config, and vcap-frontend-v1-web-config.
PVCs vcap-backend-v1-data and vcap-backend-v1-postgres-data.

The vcap-v1 Postgres 16 data is in the pgdata/ directory. Postgres stopped cleanly before the retirement. No postmaster.pid file exists. The final dump has mode 600. It holds 58 tables and 104 live rows. Its gzip check passed. The local-path host directory for the application data is /var/lib/rancher/k3s/storage/pvc-d4b91391-4691-43c4-9590-6bbe1ddf0ad9_vcap-v1_vcap-backend-v1-data.

The namespace is still declared in k8s/platform/storage/namespaces-tenants.yaml.

Disposable and infrastructure data

Dataset	Backing	Size	Backup
Prometheus TSDB	PV `pv-prometheus`, `/var/lib/k8s-monitoring/prometheus`, `Retain`	3.8 GB	none
Loki chunks	PV `pv-loki`, `/var/lib/k8s-monitoring/loki`, `Retain`	3.4 GB	none
etcd	`/var/lib/rancher/k3s/server/db`	546 MB	restic `--tag etcd`, 5 local + 14 offsite
k3s token and TLS	`/var/lib/rancher/k3s/server/{token,tls}`	216 KB	same restic snapshot as etcd

Prometheus and Loki are disposable. Their loss does not affect an application dataset.

Space

Measured 2026-10-23:

Root disk: 75 GB ext4, 46 GB used, 27 GB free, 64%.
/srv/data: 20 GB ext4, 439 MB used, 19 GB free, 3%.
Class A source data: /srv/data/vcap-staging (75 MB, measured 2026-10-23) and the gitorgasm Postgres. The class-A snapshot also contains the gitorgasm logical dump and globals.
The retained vcap-v1 trees are class C. Measured 2026-10-23, the Postgres tree was 67 MB and the application data was 15 MB.
The largest single dataset is rankstack at 509 MB, followed by meghmitra-postgres at 206 MB.

Storage configuration

local-path is the default StorageClass. It provisions under /var/lib/rancher/k3s/storage. The StorageClass default reclaim policy is Delete. Individual PVs can differ: the bitvault, rankstack, and vcap-v1 application-data PVs are Retain, while the kirro PVs are Delete.
hcloud-volumes exists but has no PVs. No Hetzner CSI volume backs a PVC. The driver is not on a critical path.
Static PVs use hostPath, not .spec.local. The path is in .spec.hostPath.path. All static PVs use Retain and pin to node vps.
The monitoring PVs use the hostpath-retain class and point under /var/lib/k8s-monitoring.

Backup coverage

Class A runs every 30 minutes. Class C runs daily. See backups.md for the exact targets and retention.

gitorgasm Postgres is a logical target. It runs pg_dump -Fc and pg_dumpall --globals-only inside the live pod. A raw copy of a running Postgres tree can be inconsistent.
vcap-staging is a class-A raw target. Its Postgres and MinIO data are static.
Class C covers vcap-dev, adhigrahan-radar, meghmitra-postgres, the vcap-v1 Postgres tree, the vcap-v1 application data, the vcap-v1 final dump directory, both bitvault PVCs, rankstack, both kirro PVCs, and Grafana.
Class C is best effort. The owner accepts loss of class-C data.

Not backed up:

meghmitra-corpus-pvc: 4 KB and effectively empty. The corpus is re-downloadable.
meghmitra-data-pvc: it is empty and has no backup target. See below.
Prometheus and Loki: disposable monitoring data.

Known gaps

`meghmitra-data-pvc` is empty

PV meghmitra-data-pvc-vps is Bound and claims hostPath /srv/data/meghmitra-data. The directory exists. The Ansible data_volume role created it on 2026-10-23. It is empty. The PVC has no recoverable content. Do not delete the PVC or the PV. The binding records the configured intent, and a future restoration may write data to the path. No backup target exists for it because there is no data to read.

`smart-home-system-api` is out of this inventory

The deployment is scaled to zero by KEDA. It has a PVC in both its own namespace and default. This predates the retirement. Leave it alone unless a separate task takes it.

Historical notes

On 2026-09-28, vcap-staging, vcap-dev, and meghmitra Postgres moved to static PVs on vps-data. The old Hetzner CSI volumes 106574315 and 106574316 are gone. The 14-day rollback window closed on 2026-10-12. Do not plan around those volumes.
The old rankstack measurement of 369 MB allocated with 0 documents and 5 collections described the running database on 2026-09-28. It is historical. The current tree is 509 MB of static WiredTiger files.
The old meghmitra database size of 215 MB described the running database. The current raw tree is 206 MB.
The class-A composition changed on 2026-10-23: vcap-staging and meghmitra logical dumps were replaced by raw file targets. vcap-v1 remained the only logical Postgres target until 2026-10-27.
On 2026-10-27 vcap v1 was retired. The vcap-v1 Postgres tree, application data, and final dump moved from class A to class C as raw file targets. The application-data PV changed from Delete to Retain. gitorgasm Postgres is the only logical Postgres target.

backups.md — backup targets, retention, and restore drills
runbooks/restore-postgres.md
runbooks/restore-mongo.md
runbooks/restore-minio.md
disaster-recovery.md

This site is open source. Improve this page.