## Adopting Existing Resources — Use `import {}` Blocks, Not the CLI
When bringing a live cluster/Vault/Cloudflare resource under Terraform management, use an HCL `import {}` block (Terraform 1.5+). Do **NOT** use `terraform import` on the CLI for anything landing in this repo — the CLI path leaves no audit trail and makes multi-operator adoption fragile.
**Canonical workflow:**
1. Write the `resource` block that matches the live object.
2. In the same stack, add an `import {}` stanza naming the target and the provider-specific ID:
```hcl
import {
to = helm_release.kured
id = "kured/kured" # Helm ID format: <namespace>/<release-name>
3.`scripts/tg plan` — every change it proposes is real divergence between HCL and live state. Iterate on values until the plan is **0 changes**.
4.`scripts/tg apply` — the import runs alongside whatever zero-change apply you have. If your plan is 0 changes, this commits only the state-ownership transfer.
5. After the apply lands cleanly, **delete the `import {}` block** in a follow-up commit. The resource is now fully TF-owned and the stanza would be a no-op that clutters diffs.
**Why `import {}` and not `terraform import`:**
- Reviewable in PRs before any state mutation. The CLI path is an out-of-band action nobody sees.
- Plan-safe: the `import` plan step shows the exact object being adopted. Mistyped IDs or the wrong resource address are caught before apply, not after.
- Survives state backend changes (Tier 0 SOPS vs Tier 1 PG) transparently — both work identically from the operator's perspective because both use `scripts/tg`.
- Re-runnable: if the apply fails partway through, the `import {}` block is idempotent. The CLI path's state mutation is not.
**Finding the provider-specific ID:** each provider has its own convention.
- **CI/CD**: Woodpecker CI — PRs run plan, merges to master auto-apply all stacks
## Key Paths
-`stacks/<service>/main.tf` — service definition
-`stacks/platform/modules/<service>/` — core infra modules
-`modules/kubernetes/ingress_factory/` — standardized ingress with auth, rate limiting, anti-AI, and auto Cloudflare DNS (`dns_type = "proxied"` or `"non-proxied"`)
- **NFS server**: Proxmox host at 192.168.1.127 (sole NFS). HDD NFS at `/srv/nfs` (2TB ext4 LV `pve/nfs-data`), SSD NFS at `/srv/nfs-ssd` (100GB ext4 LV `ssd/nfs-ssd-data`). Exports use `async` mode (safe with UPS + databases on block storage). TrueNAS (VM 9000, 10.0.10.15) decommissioned 2026-04-13. Legacy `nfs-truenas` StorageClass name retained (48 PVs bind it; SC names are immutable on PVs) but now points to the Proxmox host, identical to `nfs-proxmox`.
- **SQLite on NFS is unreliable** (fsync issues) — always use proxmox-lvm or local disk for databases.
- **NFS mount options**: Always `soft,timeo=30,retrans=3` to prevent uninterruptible sleep (D state).
- **NFS export directory must exist** on the Proxmox host before Terraform can create the PV.
- **Backup (3-2-1)**: Copy 1 = live PVCs on sdc. Copy 2 = sda `/mnt/backup` (PVC file backups, auto SQLite backups, pfSense, PVE config, **VM images via `vzdump-vms`**). Copy 3 = Synology offsite (two-tier: sda→`pve-backup/`, NFS→`nfs/`+`nfs-ssd/` via inotify change tracking).
- **vzdump-vms** (Daily 01:00): live `vzdump --mode snapshot` of hand-managed VMs (NOT in TF) → `/mnt/backup/vzdump/`, keep 3/VMID. `VZDUMP_VMIDS` default `102` (devvm) — the only VM imaged today; before this (2026-06-09) no VM was ever imaged. NOT in the incremental offsite manifest; monthly full pass mirrors it. See `docs/architecture/backup-dr.md`.
- **nfs-change-tracker.service**: inotifywait on `/srv/nfs` + `/srv/nfs-ssd`, logs to `/mnt/backup/.nfs-changes.log`. Incremental syncs complete in seconds.
## Redis Service Naming (read before wiring a new consumer)
The Redis stack (`stacks/redis/`) exposes three distinct entry points. Pick the one that matches the client's connection pattern — the wrong one causes READONLY errors or silent connection drops.
| Endpoint | Port(s) | Use for | Backed by |
|----------|---------|---------|-----------|
| `redis-master.redis.svc.cluster.local` | 6379 (redis), 26379 (sentinel) | **Default for new services.** Write-safe — HAProxy health-checks nodes and routes only to the current master. Matches `var.redis_host`. | `kubernetes_service.redis_master` → HAProxy → Bitnami StatefulSet |
| `redis-node-{0,1,2}.redis-headless.redis.svc.cluster.local` | 26379 | **Long-lived connections (PUBSUB, BLPOP, MONITOR, Sidekiq).** Use a sentinel-aware client with master name `mymaster`. Example: `stacks/nextcloud/chart_values.yaml:32-54`. | Bitnami-created headless service → pod DNS |
| `redis.redis.svc.cluster.local` | 6379 | **Do NOT use.** Helm chart's default service — selector patched by `null_resource.patch_redis_service` to match `redis-haproxy`, so today it behaves like `redis-master`. This patch is load-bearing but temporary; consumers hard-coded on this name are tracked in a beads follow-up (T0). | Bitnami chart (patched) |
**HAProxy's `timeout client 30s` closes idle raw Redis connections** — any client that holds a connection open for pub/sub, blocking commands, or replication streams MUST use the sentinel path. Uptime Kuma's Redis monitor hit this limit and had to be re-pointed at the sentinel endpoint (see memory id=748).
**When onboarding a new service:** start from `redis-master.redis.svc.cluster.local:6379` via `var.redis_host`. Only reach for sentinel discovery if the client library supports it natively (ioredis, redis-py Sentinel, go-redis FailoverClient, Sidekiq `sentinels` array) AND the workload uses long-lived connections.
Kyverno's admission webhook mutates every pod with a `dns_config { option { name = "ndots"; value = "2" } }` block (fixes NxDomain search-domain floods — see `k8s-ndots-search-domain-nxdomain-flood` skill). Terraform does not manage that field, so without suppression every pod-owning resource shows perpetual `spec[0].template[0].spec[0].dns_config` drift.
**Rule**: every `kubernetes_deployment`, `kubernetes_stateful_set`, `kubernetes_daemon_set`, and `kubernetes_cron_job_v1` MUST include the following `lifecycle` block, tagged with the `# KYVERNO_LIFECYCLE_V1` marker so every site is greppable:
**Why not a shared module?** Terraform's `ignore_changes` meta-argument only accepts static attribute paths. It rejects module outputs, locals, variables, and any expression. A DRY module is therefore impossible — the canonical pattern IS the snippet + marker. When `kubernetes_manifest` resources get Kyverno `generate.kyverno.io/*` annotations mutated, a sibling convention `# KYVERNO_MANIFEST_V1` will be introduced (Phase B).
**Audit**: `rg "KYVERNO_LIFECYCLE_V1" stacks/ | wc -l` — should grow (never shrink). Add the marker to every new pod-owning resource. The `_template/main.tf.example` stub shows the canonical form.
### `# KYVERNO_LIFECYCLE_V2` — Keel auto-update annotations
When a namespace is labeled `keel.sh/enrolled=true`, the `inject-keel-annotations` ClusterPolicy (`stacks/kyverno/modules/kyverno/keel-annotations.tf`) injects these annotations on every Deployment / StatefulSet / DaemonSet:
```
keel.sh/policy: patch
keel.sh/trigger: poll
keel.sh/pollSchedule: "@every 1h"
```
**`keel.sh/match-tag` is NO LONGER injected — it is actively STRIPPED.** It was the pre-2026-05-26 default (`force + match-tag`), proven unreliable: under `force` it let Keel rewrite tag strings and cross-assign images between containers in multi-image pods. The `blog` deployment was a casualty — its `nginx` ⇄ `nginx-exporter` images got swapped and the site was down 2026-05-26 → 2026-06-01. The policy now sets the annotation to `null` (strips on admission); the 194 pre-existing workloads still carrying it were swept once via `kubectl annotate … keel.sh/match-tag-` on 2026-06-01. The `ignore_changes` line for it (below) is retained as a harmless no-op. See `docs/post-mortems/2026-06-01-keel-match-tag-image-swap.md`.
To suppress the resulting Terraform drift, **enrolled workloads** must carry the complete `ignore_changes` block below. This is the canonical form — it folds together every marker (see the legend after it):
**Marker legend** (the names are historical; grep each to audit coverage):
| Marker | Ignores | Why |
|---|---|---|
| `# KYVERNO_LIFECYCLE_V1` | `dns_config` | Kyverno injects pod DNS `ndots` config |
| `# KYVERNO_LIFECYCLE_V2` | `keel.sh/policy`, `/trigger`, `/pollSchedule` | Kyverno-injected Keel control annotations |
| `# KEEL_IGNORE_IMAGE` | `container[N].image` (one line **per container index**, incl. `init_container[N]`) | Keel rewrites the image tag on `policy=patch`; without this, `apply` reverts the bump (a **downgrade**) |
| `# KEEL_LIFECYCLE_V1` | `keel.sh/match-tag`, `keel.sh/update-time` (pod template), `kubernetes.io/change-cause`, `deployment.kubernetes.io/revision` | every Keel digest-update restamps these; without ignoring them `apply` strips them → forces a rollout → Keel re-stamps → fight loop |
**Multi-container caveat**: `container[0].image` only covers the first container. Add one `container[N].image` line for **every** container index, plus `init_container[N].image` for init containers — otherwise the un-ignored container's image still drifts/downgrades.
The `KEEL_LIFECYCLE_V1` + per-container `KEEL_IGNORE_IMAGE` lines were swept across all enrolled workloads on **2026-05-28** (previously only `llama-cpp` had them; the rest fought on every apply). New enrolled workloads must include the full block. Workloads in un-enrolled namespaces don't receive the annotations and don't need the block.
Per-workload opt-out: add the label `keel.sh/policy: never` on the Deployment metadata (not pod template); the policy's `exclude` clause respects it, no annotation gets injected, no `ignore_changes` needed.
**Audit**: `rg "KYVERNO_LIFECYCLE_V2" stacks/` — count should equal the number of enrolled workloads. `rg "KEEL_LIFECYCLE_V1" stacks/` should match it (every enrolled workload also carries the V1 lines).
- **Nodes**: k8s-master (10.0.20.100), node1 (GPU, Tesla T4), node2-4
- **GPU**: `node_selector = { "nvidia.com/gpu.present" : "true" }` + toleration `nvidia.com/gpu`. The label is auto-applied by NFD/gpu-feature-discovery on any node with an NVIDIA PCI device — nothing is hostname-pinned, so the GPU card can move between nodes without Terraform edits.
- **Pull-through cache**: 10.0.20.10 — docker.io (:5000), ghcr.io (:5010) only. Caches stale manifests for :latest tags — use versioned tags or pre-pull with `ctr --hosts-dir ''` to bypass.
- **pfSense**: 10.0.20.1 (gateway, firewall, DNS forwarding)
- **MySQL InnoDB Cluster**: 1 instance on proxmox-lvm (scaled from 3 — only Uptime Kuma + phpIPAM remain), PriorityClass `mysql-critical` + PDB, anti-affinity excludes any GPU node (`nvidia.com/gpu.present=true`) so MySQL moves off the GPU host automatically if the card is relocated
