# 3-Pool Node Split Implementation Plan

> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.

**Goal:** Split the cluster into 3 isolated node pools (gitlab / platform / foxhunt) by creating a new `platform` pool, migrating infrastructure + monitoring there, and repurposing one `services` node as `foxhunt`.

**Architecture:** Create `platform` DEV1-L pool in Scaleway. Move stateless monitoring first (low risk), then stateful databases (requires PVC migration with backups), then rename services → foxhunt for app pods and drop the extra node.

**Tech Stack:** Scaleway Kapsule, Terragrunt/Terraform, kubectl, pg_dump/pg_restore, sed

---

### Task 1: Create `platform` Node Pool via Terraform (Phase 1 — additive only)

**Context:** The cluster is managed by Terragrunt at `infra/live/production/kapsule/`. We need to create the `platform` pool first (additive) WITHOUT renaming `services` → `foxhunt` yet. The rename would destroy the existing services pool before databases are migrated. We do this in two Terraform applies.

**Files already modified:**
- `infra/modules/kapsule/main.tf` — added `platform` pool resource, renamed `services` → `foxhunt`
- `infra/modules/kapsule/variables.tf` — added `platform_type`, `enable_platform_pool`, renamed `services_*` → `foxhunt_*`
- `infra/modules/kapsule/outputs.tf` — added `platform_pool_id`, renamed `services_pool_id` → `foxhunt_pool_id`
- `infra/live/production/kapsule/terragrunt.hcl` — updated inputs

**Step 1: Run terragrunt plan to preview changes**

```bash
cd infra/live/production/kapsule
terragrunt plan
```

Expected output should show:
- `scaleway_k8s_pool.platform[0]` — **create** (new platform pool)
- `scaleway_k8s_pool.services` — **destroy** (old name being removed)
- `scaleway_k8s_pool.foxhunt` — **create** (new name for app services)

**IMPORTANT:** Do NOT apply yet — the destroy of `services` would evict all pods before databases are migrated.

**Step 2: Use `-target` to apply ONLY the platform pool first**

```bash
cd infra/live/production/kapsule
terragrunt apply -target=scaleway_k8s_pool.platform
```

This creates the platform pool without touching the existing services pool.

**Step 3: Wait for platform node to become Ready**

```bash
kubectl get nodes -l k8s.scaleway.com/pool-name=platform -w
```

Expected: Node shows `STATUS=Ready` within 2-3 minutes.

**Step 4: Verify node resources**

```bash
kubectl get node -l k8s.scaleway.com/pool-name=platform \
  -o custom-columns="NAME:.metadata.name,CPU:.status.allocatable.cpu,MEM:.status.allocatable.memory"
```

Expected: CPU ~3800m, MEM ~6450608Ki (DEV1-L allocatable).

**Step 5: Commit Terraform changes**

```bash
git add infra/modules/kapsule/main.tf infra/modules/kapsule/variables.tf \
        infra/modules/kapsule/outputs.tf infra/live/production/kapsule/terragrunt.hcl
git commit -m "infra(tf): add platform pool, rename services → foxhunt in Kapsule module"
```

---

### Task 2: Move Stateless Monitoring to `platform` (Loki, Tempo, kube-state-metrics, Pushgateway)

**Context:** These pods are currently on the `gitlab` node (loki, tempo, kube-state-metrics) or `services` node (pushgateway). They have no persistent data that can't be recreated. Moving them is just a nodeSelector change.

**Files:**
- Modify: `infra/k8s/monitoring/loki.yaml:64`
- Modify: `infra/k8s/monitoring/tempo.yaml:60`
- Modify: `infra/k8s/monitoring/kube-state-metrics.yaml:118`
- Modify: `infra/k8s/monitoring/pushgateway.yaml:20`

**Step 1: Change nodeSelectors in manifests**

In each file, change `k8s.scaleway.com/pool-name` value to `platform`:

```bash
# loki.yaml line 64: gitlab → platform
sed -i 's/pool-name: gitlab/pool-name: platform/' infra/k8s/monitoring/loki.yaml

# tempo.yaml line 60: gitlab → platform
sed -i 's/pool-name: gitlab/pool-name: platform/' infra/k8s/monitoring/tempo.yaml

# kube-state-metrics.yaml line 118: gitlab → platform
sed -i 's/pool-name: gitlab/pool-name: platform/' infra/k8s/monitoring/kube-state-metrics.yaml

# pushgateway.yaml line 20: services → platform
sed -i 's/pool-name: services/pool-name: platform/' infra/k8s/monitoring/pushgateway.yaml
```

**Step 2: Apply and restart**

```bash
kubectl apply -f infra/k8s/monitoring/loki.yaml -n foxhunt
kubectl apply -f infra/k8s/monitoring/tempo.yaml -n foxhunt
kubectl apply -f infra/k8s/monitoring/kube-state-metrics.yaml -n foxhunt
kubectl apply -f infra/k8s/monitoring/pushgateway.yaml -n foxhunt
```

**Step 3: Verify pods rescheduled to platform node**

```bash
kubectl get pods -n foxhunt -l 'app.kubernetes.io/name in (loki,tempo,kube-state-metrics,pushgateway)' \
  -o custom-columns="POD:.metadata.name,NODE:.spec.nodeName,STATUS:.status.phase"
```

Expected: All 4 pods show a node name containing `platform` and STATUS=Running.

**Step 4: Verify Loki is ingesting**

```bash
GRAFANA_POD=$(kubectl get pod -n foxhunt -l app.kubernetes.io/name=grafana -o jsonpath='{.items[0].metadata.name}')
kubectl exec -n foxhunt "$GRAFANA_POD" -c grafana -- \
  curl -s "http://loki.foxhunt.svc.cluster.local:3100/loki/api/v1/labels" | head -5
```

Expected: JSON response with `"status": "success"` and label names.

**Step 5: Commit**

```bash
git add infra/k8s/monitoring/loki.yaml infra/k8s/monitoring/tempo.yaml \
        infra/k8s/monitoring/kube-state-metrics.yaml infra/k8s/monitoring/pushgateway.yaml
git commit -m "infra: move monitoring stack (loki, tempo, kube-state-metrics, pushgateway) to platform pool"
```

---

### Task 3: Move Grafana to `platform`

**Context:** Grafana is deployed via Helm with values in `grafana-values.yaml`. It has a 2Gi PVC for dashboards/state, but all dashboards are provisioned from ConfigMaps so PVC data loss is acceptable. The simplest approach is to change the nodeSelector, delete the old PVC, and let a new one get created on the platform node.

**Files:**
- Modify: `infra/k8s/gitlab/grafana-values.yaml:7`

**Step 1: Change nodeSelector in Helm values**

```bash
sed -i 's/pool-name: gitlab/pool-name: platform/' infra/k8s/gitlab/grafana-values.yaml
```

Verify the file now reads:
```yaml
nodeSelector:
  k8s.scaleway.com/pool-name: platform
```

**Step 2: Upgrade Grafana via Helm**

```bash
helm upgrade grafana grafana/grafana \
  -n foxhunt \
  -f infra/k8s/gitlab/grafana-values.yaml \
  --reuse-values
```

**Step 3: If pod is Pending due to PVC bound to old node**

The grafana PVC is ReadWriteOnce and may be bound to the old gitlab node. If the pod is stuck Pending:

```bash
# Delete the old PVC — dashboards are all provisioned, no data loss
kubectl delete pvc grafana -n foxhunt
# Restart the deployment to trigger new PVC creation
kubectl rollout restart deployment grafana -n foxhunt
```

**Step 4: Wait for Grafana to be ready**

```bash
kubectl rollout status deployment grafana -n foxhunt --timeout=120s
```

Expected: `deployment "grafana" successfully rolled out`

**Step 5: Verify Grafana is accessible and dashboards load**

```bash
GRAFANA_POD=$(kubectl get pod -n foxhunt -l app.kubernetes.io/name=grafana -o jsonpath='{.items[0].metadata.name}')
ADMIN_PASS=$(kubectl get secret grafana -n foxhunt -o jsonpath='{.data.admin-password}' | base64 -d)
kubectl exec -n foxhunt "$GRAFANA_POD" -c grafana -- \
  curl -s -u "admin:${ADMIN_PASS}" http://localhost:3000/api/search?type=dash-db | python3 -c "import sys,json; print(f'{len(json.load(sys.stdin))} dashboards loaded')"
```

Expected: `13 dashboards loaded`

**Step 6: Commit**

```bash
git add infra/k8s/gitlab/grafana-values.yaml
git commit -m "infra: move Grafana to platform pool"
```

---

### Task 4: Move Tailscale Proxies to `platform`

**Context:** Two tailscale deployments: `tailscale-gitlab-proxy` (in `gitlab/`) handles SSH to GitLab, and `tailscale-subnet-router` (in `tailscale/`) provides network access. Both are stateless and can simply be moved.

**Files:**
- Modify: `infra/k8s/gitlab/tailscale-proxy.yaml:54`
- Modify: `infra/k8s/tailscale/deployment.yaml:20`

**Step 1: Change nodeSelectors**

```bash
sed -i 's/pool-name: gitlab/pool-name: platform/' infra/k8s/gitlab/tailscale-proxy.yaml
sed -i 's/pool-name: services/pool-name: platform/' infra/k8s/tailscale/deployment.yaml
```

**Step 2: Apply both**

```bash
kubectl apply -f infra/k8s/gitlab/tailscale-proxy.yaml -n foxhunt
kubectl apply -f infra/k8s/tailscale/deployment.yaml -n tailscale
```

**Step 3: Verify pods moved**

```bash
kubectl get pod -n foxhunt -l app.kubernetes.io/name=tailscale-gitlab-proxy \
  -o custom-columns="POD:.metadata.name,NODE:.spec.nodeName"
kubectl get pod -n tailscale -l app.kubernetes.io/name=tailscale-subnet-router \
  -o custom-columns="POD:.metadata.name,NODE:.spec.nodeName"
```

Expected: Both show node containing `platform`.

**Step 4: Verify GitLab SSH still works through tailscale proxy**

```bash
ssh -T git@100.90.76.85 -p 2222 2>&1 | head -3
```

Expected: `Welcome to GitLab` or similar greeting (not connection refused).

**Step 5: Commit**

```bash
git add infra/k8s/gitlab/tailscale-proxy.yaml infra/k8s/tailscale/deployment.yaml
git commit -m "infra: move tailscale proxies to platform pool"
```

---

### Task 5: Migrate Redis to `platform` (Stateless Move)

**Context:** Redis is used for caching — no persistent data that can't be regenerated. Simply change nodeSelector and restart. No PVC migration needed.

**Files:**
- Modify: `infra/k8s/databases/redis.yaml:21`

**Step 1: Change nodeSelector**

```bash
sed -i 's/pool-name: services/pool-name: platform/' infra/k8s/databases/redis.yaml
```

**Step 2: Apply and wait**

```bash
kubectl apply -f infra/k8s/databases/redis.yaml -n foxhunt
kubectl rollout status deployment redis -n foxhunt --timeout=60s
```

Expected: `deployment "redis" successfully rolled out`

**Step 3: Verify Redis is responding**

```bash
kubectl exec -n foxhunt deployment/redis -- redis-cli ping
```

Expected: `PONG`

**Step 4: Commit**

```bash
git add infra/k8s/databases/redis.yaml
git commit -m "infra: move Redis to platform pool"
```

---

### Task 6: Migrate Postgres to `platform` (PVC Migration)

**Context:** Postgres has a 10Gi ReadWriteOnce PVC bound to the current services node. We need to back up the data, create a new PVC on the platform node, restore, and switch over. This causes a few minutes of downtime for postgres.

**Files:**
- Modify: `infra/k8s/databases/postgres.yaml:38`
- Modify: `infra/k8s/gitlab/postgres-init.yaml:15`

**Step 1: Backup postgres data**

```bash
# Get current postgres pod
PG_POD=$(kubectl get pod -n foxhunt -l app.kubernetes.io/name=postgres -o jsonpath='{.items[0].metadata.name}')

# Run pg_dumpall to capture everything
kubectl exec -n foxhunt "$PG_POD" -- pg_dumpall -U foxhunt > /tmp/foxhunt-pg-backup.sql
echo "Backup size: $(wc -c < /tmp/foxhunt-pg-backup.sql) bytes"
```

Expected: Non-zero file size (should be several MB with GitLab + foxhunt data).

**Step 2: Scale postgres to 0**

```bash
kubectl scale deployment postgres -n foxhunt --replicas=0
kubectl get pods -n foxhunt -l app.kubernetes.io/name=postgres
```

Expected: No pods running.

**Step 3: Delete old PVC and create a new one**

```bash
# Delete the PVC bound to the old node
kubectl delete pvc postgres-pvc -n foxhunt

# Apply the manifest again — it includes the PVC definition and will create a fresh one
# The new PVC will bind to the platform node when the pod schedules there
```

**Step 4: Change nodeSelector in postgres manifest**

```bash
sed -i 's/pool-name: services/pool-name: platform/' infra/k8s/databases/postgres.yaml
```

**Step 5: Apply manifest and scale up**

```bash
kubectl apply -f infra/k8s/databases/postgres.yaml -n foxhunt
# The deployment should automatically create 1 replica with the new nodeSelector
kubectl rollout status deployment postgres -n foxhunt --timeout=120s
```

Expected: Pod starts on the platform node with a fresh (empty) database.

**Step 6: Restore data**

```bash
PG_POD=$(kubectl get pod -n foxhunt -l app.kubernetes.io/name=postgres -o jsonpath='{.items[0].metadata.name}')

# Copy backup into pod and restore
kubectl cp /tmp/foxhunt-pg-backup.sql foxhunt/"$PG_POD":/tmp/backup.sql
kubectl exec -n foxhunt "$PG_POD" -- psql -U foxhunt -f /tmp/backup.sql postgres
```

**Step 7: Verify postgres is healthy**

```bash
kubectl exec -n foxhunt "$PG_POD" -- psql -U foxhunt -c "SELECT datname FROM pg_database;"
```

Expected: Lists databases including `foxhunt`, `gitlabhq_production`, etc.

**Step 8: Update postgres-init nodeSelector too**

```bash
sed -i 's/pool-name: services/pool-name: platform/' infra/k8s/gitlab/postgres-init.yaml
```

**Step 9: Commit**

```bash
git add infra/k8s/databases/postgres.yaml infra/k8s/gitlab/postgres-init.yaml
git commit -m "infra: migrate Postgres to platform pool with PVC migration"
```

---

### Task 7: Migrate QuestDB to `platform` (PVC Migration)

**Context:** QuestDB has a 10Gi ReadWriteOnce PVC. It stores ML model metrics (predictions, trade outcomes). Data is nice-to-have but not critical — QuestDB will recreate tables on startup. We can do a simple file copy if desired, or just start fresh.

**Files:**
- Modify: `infra/k8s/databases/questdb.yaml:38`

**Step 1: Scale QuestDB to 0**

```bash
kubectl scale deployment questdb -n foxhunt --replicas=0
```

**Step 2: Delete old PVC**

```bash
kubectl delete pvc questdb-pvc -n foxhunt
```

**Step 3: Change nodeSelector**

```bash
sed -i 's/pool-name: services/pool-name: platform/' infra/k8s/databases/questdb.yaml
```

**Step 4: Apply and scale up**

```bash
kubectl apply -f infra/k8s/databases/questdb.yaml -n foxhunt
kubectl rollout status deployment questdb -n foxhunt --timeout=120s
```

Expected: QuestDB starts fresh on the platform node.

**Step 5: Verify QuestDB is responding**

```bash
QUESTDB_POD=$(kubectl get pod -n foxhunt -l app.kubernetes.io/name=questdb -o jsonpath='{.items[0].metadata.name}')
kubectl exec -n foxhunt "$QUESTDB_POD" -- curl -s "http://localhost:9000/exec?query=SELECT+1" 2>/dev/null | head -3
```

Expected: JSON response with result.

**Step 6: Commit**

```bash
git add infra/k8s/databases/questdb.yaml
git commit -m "infra: migrate QuestDB to platform pool (fresh start)"
```

---

### Task 8: Migrate MinIO to `platform` (PVC Migration)

**Context:** MinIO has a 20Gi PVC containing compiled binaries and trained models. We need to preserve this data. Strategy: use a temp pod to mount both old and new PVCs and copy data via rsync.

**Files:**
- Modify: `infra/k8s/minio/minio.yaml:57,152`

**Step 1: Record what's in MinIO**

```bash
# Using minio mc client from the init job approach
MINIO_POD=$(kubectl get pod -n foxhunt -l app.kubernetes.io/name=minio -o jsonpath='{.items[0].metadata.name}')
kubectl exec -n foxhunt "$MINIO_POD" -- ls -la /data/
```

Note the contents for verification after migration.

**Step 2: Scale MinIO to 0**

```bash
kubectl scale deployment minio -n foxhunt --replicas=0
```

**Step 3: Create a migration pod to copy data**

```bash
cat <<'EOF' | kubectl apply -n foxhunt -f -
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: minio-data-new
  namespace: foxhunt
  labels:
    app.kubernetes.io/name: minio
    app.kubernetes.io/part-of: foxhunt
spec:
  accessModes:
    - ReadWriteOnce
  storageClassName: scw-bssd
  resources:
    requests:
      storage: 20Gi
---
apiVersion: v1
kind: Pod
metadata:
  name: minio-migrate
  namespace: foxhunt
spec:
  nodeSelector:
    k8s.scaleway.com/pool-name: platform
  restartPolicy: Never
  containers:
    - name: migrate
      image: debian:bookworm-slim
      command: ["sleep", "3600"]
      volumeMounts:
        - name: new-data
          mountPath: /new
  volumes:
    - name: new-data
      persistentVolumeClaim:
        claimName: minio-data-new
EOF
```

Wait for the migration pod to be Running:
```bash
kubectl wait --for=condition=Ready pod/minio-migrate -n foxhunt --timeout=120s
```

**Step 4: Copy data from old PVC to new PVC**

Since the old PVC is ReadWriteOnce and bound to a services node, we can't mount it directly on the platform node. Instead, we'll use a second pod on the services node to tar the data and pipe it:

```bash
# Create a reader pod on the old node
cat <<'EOF' | kubectl apply -n foxhunt -f -
apiVersion: v1
kind: Pod
metadata:
  name: minio-reader
  namespace: foxhunt
spec:
  nodeSelector:
    k8s.scaleway.com/pool-name: services
  restartPolicy: Never
  containers:
    - name: reader
      image: debian:bookworm-slim
      command: ["sleep", "3600"]
      volumeMounts:
        - name: old-data
          mountPath: /old
          readOnly: true
  volumes:
    - name: old-data
      persistentVolumeClaim:
        claimName: minio-data
EOF

kubectl wait --for=condition=Ready pod/minio-reader -n foxhunt --timeout=120s

# Pipe tar from reader to migrate pod
kubectl exec -n foxhunt minio-reader -- tar cf - -C /old . | \
  kubectl exec -i -n foxhunt minio-migrate -- tar xf - -C /new

# Verify data arrived
kubectl exec -n foxhunt minio-migrate -- ls -la /new/
```

Expected: Same directory listing as Step 1.

**Step 5: Clean up migration pods**

```bash
kubectl delete pod minio-reader minio-migrate -n foxhunt
```

**Step 6: Swap PVCs**

```bash
# Delete old PVC
kubectl delete pvc minio-data -n foxhunt

# Rename new PVC (can't rename in k8s — need to update the manifest to use the new name,
# or recreate with the original name)
# Simplest: update minio.yaml to reference minio-data-new
```

Actually, the simplest approach: edit the minio manifest to use `minio-data-new` as the PVC name:

```bash
sed -i 's/claimName: minio-data/claimName: minio-data-new/' infra/k8s/minio/minio.yaml
```

**Step 7: Change nodeSelectors**

```bash
sed -i 's/pool-name: services/pool-name: platform/' infra/k8s/minio/minio.yaml
```

**Step 8: Apply and start MinIO**

```bash
kubectl apply -f infra/k8s/minio/minio.yaml -n foxhunt
kubectl rollout status deployment minio -n foxhunt --timeout=120s
```

**Step 9: Verify MinIO has all data**

```bash
MINIO_POD=$(kubectl get pod -n foxhunt -l app.kubernetes.io/name=minio -o jsonpath='{.items[0].metadata.name}')
kubectl exec -n foxhunt "$MINIO_POD" -- ls -la /data/
```

Expected: Same contents as Step 1.

**Step 10: Commit**

```bash
git add infra/k8s/minio/minio.yaml
git commit -m "infra: migrate MinIO to platform pool with data migration"
```

---

### Task 9: Rename Services → Foxhunt Pool via Terraform (Phase 2 — destructive)

**Context:** All databases and monitoring are now on the `platform` node. The `services` pool only has foxhunt app pods + binary-cache PVCs (disposable). Now we run `terragrunt apply` to complete the rename: Terraform will destroy the old `services` pool (2 nodes) and create a new `foxhunt` pool (1 node). App pods will be evicted temporarily.

**Files:**
- Modify: `infra/k8s/services/api-gateway.yaml:39`
- Modify: `infra/k8s/services/trading-service.yaml:39`
- Modify: `infra/k8s/services/broker-gateway.yaml:39`
- Modify: `infra/k8s/services/ib-gateway.yaml:26`
- Modify: `infra/k8s/services/ml-training-service.yaml:82`
- Modify: `infra/k8s/services/trading-agent-service.yaml:39`
- Modify: `infra/k8s/services/backtesting-service.yaml:39`
- Modify: `infra/k8s/services/data-acquisition-service.yaml:39`
- Modify: `infra/k8s/services/web-gateway.yaml:39`
- Modify: `infra/k8s/training/idle-reaper.yaml:58`

**Step 1: Update all foxhunt service manifests to target `foxhunt` pool**

```bash
# All services currently say pool-name: services → change to pool-name: foxhunt
for f in infra/k8s/services/api-gateway.yaml \
         infra/k8s/services/trading-service.yaml \
         infra/k8s/services/broker-gateway.yaml \
         infra/k8s/services/ib-gateway.yaml \
         infra/k8s/services/ml-training-service.yaml \
         infra/k8s/services/trading-agent-service.yaml \
         infra/k8s/services/backtesting-service.yaml \
         infra/k8s/services/data-acquisition-service.yaml \
         infra/k8s/services/web-gateway.yaml \
         infra/k8s/training/idle-reaper.yaml; do
  sed -i 's/pool-name: services/pool-name: foxhunt/' "$f"
done
```

**Step 2: Verify changes are correct**

```bash
grep -r "pool-name: services" infra/k8s/services/ infra/k8s/training/idle-reaper.yaml
```

Expected: No output (all changed to `foxhunt`).

```bash
grep -r "pool-name: foxhunt" infra/k8s/services/ infra/k8s/training/idle-reaper.yaml | wc -l
```

Expected: 10 (9 services + 1 idle-reaper).

**Step 3: Delete all binary-cache PVCs (they'll be recreated on the new node)**

```bash
kubectl delete pvc -n foxhunt -l app.kubernetes.io/part-of=foxhunt --field-selector metadata.name!=postgres-pvc,metadata.name!=questdb-pvc,metadata.name!=minio-data-new 2>/dev/null || true

# More precisely — delete just the binary-cache PVCs
for SVC in api-gateway backtesting-service broker-gateway data-acquisition-service \
           ml-training-service trading-agent-service trading-service web-gateway; do
  kubectl delete pvc "binary-cache-${SVC}" -n foxhunt 2>/dev/null || true
done
```

**Step 4: Run Terraform Phase 2 — destroy services pool, create foxhunt pool**

```bash
cd infra/live/production/kapsule
terragrunt apply
```

This will:
- **Destroy** `scaleway_k8s_pool.services` (both services nodes get deleted)
- **Create** `scaleway_k8s_pool.foxhunt` (1× DEV1-L)

All pods on the old services nodes get evicted. Wait for the new foxhunt node to become Ready.

**Step 5: Wait for new foxhunt node**

```bash
kubectl get nodes -l k8s.scaleway.com/pool-name=foxhunt -w
```

Expected: 1 node with STATUS=Ready within 2-3 minutes.

**Step 6: Apply updated K8s manifests**

```bash
kubectl apply -f infra/k8s/services/ -n foxhunt
kubectl apply -f infra/k8s/training/idle-reaper.yaml -n foxhunt
```

**Step 7: Wait for all foxhunt services to be Running**

```bash
kubectl get pods -n foxhunt -l app.kubernetes.io/part-of=foxhunt -w
```

Expected: All 9 service pods + ib-gateway Running on the foxhunt node. Binary caches will be populated from S3 via initContainers.

**Step 8: Commit**

```bash
git add infra/k8s/services/ infra/k8s/training/idle-reaper.yaml
git commit -m "infra: rename services → foxhunt pool in all K8s manifests"
```

---

### Task 10: Move idle-reaper to `foxhunt` pool

**Context:** The idle-reaper CronJob currently targets `services`. Since it manages GPU node scaling, it logically belongs on the foxhunt/services node. Since we kept the pool name as `services`, no actual change is needed. However, if we want to be explicit:

**Files:**
- Verify: `infra/k8s/training/idle-reaper.yaml:58` — already says `services`, no change needed

**Step 1: Verify idle-reaper runs on the services (foxhunt) node**

```bash
kubectl get cronjob gpu-idle-reaper -n foxhunt -o jsonpath='{.spec.jobTemplate.spec.template.spec.nodeSelector}'
```

Expected: `{"k8s.scaleway.com/pool-name":"services"}` — already correct.

No commit needed.

---

### Task 11: Full Cluster Verification

**Context:** Verify the entire cluster is healthy after migration. Check every pool has the right pods, all services are communicating, dashboards work, CI works.

**Step 1: Verify node pool distribution**

```bash
echo "=== Pods by Node Pool ==="
for POOL in gitlab platform services; do
  echo ""
  echo "--- $POOL ---"
  for NODE in $(kubectl get nodes -l k8s.scaleway.com/pool-name="$POOL" -o jsonpath='{.items[*].metadata.name}'); do
    kubectl get pods -A --field-selector spec.nodeName="$NODE" --no-headers 2>/dev/null | \
      awk '{print $1"/"$2}' | sort
  done
done
```

Expected:
- `gitlab`: Only gitlab-* pods + daemonsets (node-exporter, promtail)
- `platform`: postgres, redis, minio, questdb, grafana, loki, tempo, kube-state-metrics, tailscale-*, pushgateway + daemonsets
- `services`: Only foxhunt app pods (api-gateway, trading-service, etc.) + ib-gateway + daemonsets

**Step 2: Verify Grafana dashboards**

```bash
GRAFANA_POD=$(kubectl get pod -n foxhunt -l app.kubernetes.io/name=grafana -o jsonpath='{.items[0].metadata.name}')
ADMIN_PASS=$(kubectl get secret grafana -n foxhunt -o jsonpath='{.data.admin-password}' | base64 -d)
kubectl exec -n foxhunt "$GRAFANA_POD" -c grafana -- \
  curl -s -u "admin:${ADMIN_PASS}" http://localhost:3000/api/search?type=dash-db | \
  python3 -c "import sys,json; ds=json.load(sys.stdin); print(f'{len(ds)} dashboards'); [print(f'  {d[\"title\"]}') for d in ds]"
```

Expected: 13 dashboards listed.

**Step 3: Verify Loki ingestion**

```bash
kubectl exec -n foxhunt "$GRAFANA_POD" -c grafana -- \
  curl -s "http://loki.foxhunt.svc.cluster.local:3100/loki/api/v1/labels" | python3 -c "import sys,json; d=json.load(sys.stdin); print(f'Loki status={d[\"status\"]}, {len(d.get(\"data\",[]))} labels')"
```

Expected: `Loki status=success, N labels`

**Step 4: Verify GitLab CI pipeline**

Push a trivial commit or trigger a pipeline manually:
```bash
git push  # The commit from earlier tasks triggers CI
```

Check pipeline status in GitLab UI or:
```bash
curl -s --header "PRIVATE-TOKEN: $(cat ~/.gitlab-token)" \
  "http://100.90.76.85:2222/api/v4/projects/1/pipelines?per_page=1" | python3 -c "import sys,json; p=json.load(sys.stdin)[0]; print(f'Pipeline #{p[\"id\"]}: {p[\"status\"]}')"
```

Expected: Pipeline running or passed.

**Step 5: Verify network connectivity between pools**

```bash
# Test foxhunt services can reach platform databases
kubectl exec -n foxhunt deployment/trading-service -- curl -s http://redis.foxhunt.svc.cluster.local:6379/ping 2>/dev/null || \
  kubectl exec -n foxhunt deployment/trading-service -- nc -zv redis.foxhunt.svc.cluster.local 6379 2>&1
```

Expected: Connection successful (TCP to Redis on platform node).

**Step 6: Check resource utilization on new layout**

```bash
kubectl top nodes
```

Expected: 3 nodes visible (gitlab, platform, services) with reasonable utilization.

**Step 7: Final commit**

```bash
git add -A
git status  # Verify only expected changes
git commit -m "infra: complete 3-pool node split (gitlab/platform/foxhunt)

- Created platform pool (DEV1-L) for databases + monitoring
- Migrated Postgres (pg_dump/restore), MinIO (tar copy), QuestDB (fresh)
- Moved Loki, Tempo, Grafana, kube-state-metrics, pushgateway to platform
- Scaled services pool from 2 to 1 node (now foxhunt app services only)
- Same EUR 61/mo total, proper blast radius isolation"
```

---

## Summary of Changes by File

| File | Change |
|------|--------|
| `infra/k8s/monitoring/loki.yaml` | `gitlab` → `platform` |
| `infra/k8s/monitoring/tempo.yaml` | `gitlab` → `platform` |
| `infra/k8s/monitoring/kube-state-metrics.yaml` | `gitlab` → `platform` |
| `infra/k8s/monitoring/pushgateway.yaml` | `services` → `platform` |
| `infra/k8s/gitlab/grafana-values.yaml` | `gitlab` → `platform` |
| `infra/k8s/gitlab/tailscale-proxy.yaml` | `gitlab` → `platform` |
| `infra/k8s/tailscale/deployment.yaml` | `services` → `platform` |
| `infra/k8s/databases/redis.yaml` | `services` → `platform` |
| `infra/k8s/databases/postgres.yaml` | `services` → `platform` |
| `infra/k8s/databases/questdb.yaml` | `services` → `platform` |
| `infra/k8s/minio/minio.yaml` | `services` → `platform`, PVC name update |
| `infra/k8s/gitlab/postgres-init.yaml` | `services` → `platform` |

**Unchanged (stay on `gitlab`):** `values.yaml`, `runner-values.yaml`, `runner-rl-values.yaml`, `pat-rotation.yaml`
**Changed `services` → `foxhunt`:** All 9 foxhunt service YAMLs, `idle-reaper.yaml`
**Terraform:** `infra/modules/kapsule/{main,variables,outputs}.tf`, `infra/live/production/kapsule/terragrunt.hcl`

## Rollback Procedure

**Before Phase 2 (services pool still exists):**
1. Revert `platform` nodeSelectors back to `gitlab` or `services`
2. `kubectl apply` each file
3. For stateful services (postgres, minio): restore from backups
4. `terragrunt destroy -target=scaleway_k8s_pool.platform` to remove platform pool
5. Revert Terraform changes in git

**After Phase 2 (services pool destroyed, foxhunt pool created):**
1. Revert Terraform to recreate `services` pool: `git revert` the TF commits, `terragrunt apply`
2. Revert all K8s manifests: `pool-name: foxhunt` → `pool-name: services`
3. Restore databases from backups
4. Destroy platform pool