feat(infra): forte-drop in-cluster postgres + nightly backup (prod) #17

Closed
jorgen.stensrud wants to merge 7 commits from feat/forte-drop-infra into main
Showing only changes of commit 94c7924e65 - Show all commits

View File

@@ -0,0 +1,143 @@
# forte-drop Postgres — backup & restore runbook
## What gets backed up
A CronJob (`forte-drop-pg-backup`, namespace `forte-drop`) runs nightly at **02:00 UTC**:
1. `pg_dump` of the `drops` database → gzip.
2. Upload to **UpCloud Managed Object Storage**: `s3://drops/_pgbackups/forte-drop-<TS>.sql.gz`
(the `_pgbackups/` prefix is collision-proof: app slugs match `/^[a-z0-9][a-z0-9-]{0,62}$/`
and can never start with `_`).
3. Retention: dumps older than **30 days** are pruned.
S3 creds come from the `forte-drop-secrets` Secret (`S3_ENDPOINT` / `S3_KEY` / `S3_SECRET`).
Postgres creds from `forte-drop-pg-creds` (`pgusername` / `pgpassword`).
> **Object storage is the durable tier.** App data + DB backups both live in UpCloud
> Managed Object Storage (replicated by UpCloud). The in-cluster Postgres PVC is the
> live working copy; the nightly dump is the recovery point. The PVC carries
> `Prune=false,Delete=false` so ArgoCD never deletes it.
## Prerequisites
```bash
export KUBECONFIG=~/Downloads/dev-fd-no-svg1_kubeconfig.yaml
# Confirm the namespace + DB pod are up:
kubectl -n forte-drop get pods -l app.kubernetes.io/name=postgresql
```
## List available backups
```bash
# Run an ephemeral mc pod with the app's S3 creds:
kubectl -n forte-drop run mc-list --rm -it --restart=Never \
--image=quay.io/minio/mc:RELEASE.2024-11-21T17-21-54Z \
--overrides='{"spec":{"containers":[{"name":"mc","image":"quay.io/minio/mc:RELEASE.2024-11-21T17-21-54Z","command":["sh","-c","mc alias set obj \"$S3_ENDPOINT\" \"$S3_KEY\" \"$S3_SECRET\" >/dev/null && mc ls obj/drops/_pgbackups/"],"envFrom":[{"secretRef":{"name":"forte-drop-secrets"}}]}]}}'
```
## Manually trigger a backup (before risky changes)
```bash
kubectl -n forte-drop create job --from=cronjob/forte-drop-pg-backup pg-backup-manual-$(date +%s)
# Watch:
kubectl -n forte-drop get jobs -l app.kubernetes.io/component=backup
kubectl -n forte-drop logs -l app.kubernetes.io/component=backup --tail=40
```
## Restore a dump
> **Destructive.** This overwrites the live `drops` database. Take a fresh manual
> backup first (above) and confirm with whoever owns the data before proceeding.
### 1. Pick the dump to restore
List backups (above), choose `forte-drop-<TS>.sql.gz`.
### 2. Run a restore pod that pulls the dump and pipes it into Postgres
```bash
DUMP="forte-drop-20260530T020000Z.sql.gz" # <-- set to the chosen file
kubectl -n forte-drop run pg-restore --rm -it --restart=Never \
--image=postgres:16-alpine \
--overrides='{
"spec": {
"containers": [{
"name": "restore",
"image": "postgres:16-alpine",
"command": ["sh","-c","set -euo pipefail; \
apk add --no-cache curl >/dev/null; \
# download via mc is simpler — use a 2-step instead (see note). \
echo placeholder"],
"envFrom": [
{"secretRef":{"name":"forte-drop-pg-creds"}},
{"secretRef":{"name":"forte-drop-secrets"}}
]
}]
}
}'
```
**Simpler 2-pod approach (recommended — avoids cramming mc + psql in one image):**
```bash
DUMP="forte-drop-20260530T020000Z.sql.gz"
# (a) Download the dump from object storage to a local file:
kubectl -n forte-drop run mc-get --rm -it --restart=Never \
--image=quay.io/minio/mc:RELEASE.2024-11-21T17-21-54Z \
--overrides='{"spec":{"containers":[{"name":"mc","image":"quay.io/minio/mc:RELEASE.2024-11-21T17-21-54Z","command":["sh","-c","mc alias set obj \"$S3_ENDPOINT\" \"$S3_KEY\" \"$S3_SECRET\" >/dev/null && mc cat obj/drops/_pgbackups/'"$DUMP"'"],"envFrom":[{"secretRef":{"name":"forte-drop-secrets"}}]}]}}' \
> /tmp/$DUMP
# (b) Pipe it into the live Postgres via the service:
gunzip -c /tmp/$DUMP | kubectl -n forte-drop run pg-restore --rm -i --restart=Never \
--image=postgres:16-alpine \
--overrides='{"spec":{"containers":[{"name":"psql","image":"postgres:16-alpine","stdin":true,"command":["sh","-c","PGPASSWORD=\"$pgpassword\" psql -h forte-drop-postgresql.forte-drop.svc -U \"$pgusername\" -d drops"],"env":[{"name":"pgusername","valueFrom":{"secretKeyRef":{"name":"forte-drop-pg-creds","key":"pgusername"}}},{"name":"pgpassword","valueFrom":{"secretKeyRef":{"name":"forte-drop-pg-creds","key":"pgpassword"}}}]}]}}'
```
> The app's schema is created idempotently on boot (`CREATE TABLE IF NOT EXISTS` +
> `ALTER TABLE ... ADD COLUMN IF NOT EXISTS` in `src/repo/pg.ts`), and `pg_dump`
> output includes the data. For a clean restore into a fresh DB this just works.
> To restore over an existing DB with conflicting rows, drop/recreate the `drops`
> database first (coordinate downtime — scale the web Deployment to 0 during the
> restore so the app isn't writing).
### 3. Verify
```bash
kubectl -n forte-drop run pg-check --rm -it --restart=Never \
--image=postgres:16-alpine \
--env="PGPASSWORD=$(kubectl -n forte-drop get secret forte-drop-pg-creds -o jsonpath='{.data.pgpassword}' | base64 -d)" \
--command -- psql -h forte-drop-postgresql.forte-drop.svc -U drops -d drops \
-c "SELECT count(*) AS drops FROM drops;" -c "SELECT count(*) AS view_hits FROM view_hits;"
```
### 4. Bring the app back
```bash
# If you scaled web to 0 for the restore:
kubectl -n forte-drop scale deploy/forte-drop --replicas=2
```
## Object data (uploaded drop files)
Drop files live in `s3://drops/<slug>/...` in the same managed bucket. They are
**not** part of the pg backup (the dump only holds metadata). Object storage is
UpCloud-managed/replicated, so no separate file backup is configured. If a
file-level backup is later required, mirror the bucket to a second bucket/region:
```bash
mc mirror --overwrite obj/drops/ backup-target/drops-mirror/
```
(Exclude `_pgbackups/` from the app-data mirror if you split them.)
## Disaster scenarios
| Scenario | Recovery |
|---|---|
| Postgres pod crash / reschedule | StatefulSet reattaches the PVC; ~12 min downtime; no data loss. |
| PVC lost / corrupted | Recreate StatefulSet, restore latest nightly dump (above). Data since last dump is lost. |
| Accidental `drops` table data loss | Restore latest dump; or `pg_restore` a single table from a dump. |
| Namespace deleted | PVC has `Prune=false,Delete=false`; recreate Applications, PVC re-binds, app recovers. Backups in object storage are independent. |
| Object storage bucket lost | UpCloud-managed (replicated). If the IAM key is rotated, update `forte-drop-secrets` (re-seal). |