felhom.eu/REPORT.md

# REPORT — R-80 → R-85: from a false alarm to an audible DR tier (2026-07-26 → 27)

One long session. It began as a diagnostic into a nightly alarm and ended with the offsite DR tier
being **restore-proven unattended, with its failure audible**. Five roadmap items, four artifacts.

| Repo | From → To |
|---|---|
| felhom-agent | v0.96.0 → **v0.104.0** |
| felhom-controller | v0.173.0 → **v0.175.0** |
| felhom.eu hub | v0.74.0 → **v0.77.0** |
| host-install | 1.19.0 → **1.20.0** |

All live on demo-felhom and demo-hp.

---

## 1. The arc, in one line each

| Item | What it was |
|---|---|
| **R-80** | Diagnose `expected_backup_missed` firing on three boxes. **The premise was wrong on both counts** — it fired ONCE, not nightly, and no real external customer was notified |
| **R-81** | Fix the CLASS behind it: absence of a signal treated as evidence of failure. Third instance |
| **R-82** | The backup target split — "local daily + PBS weekly" was **not expressible at all**, which is why the DR tier was `applied` and empty |
| **R-84** | An agent restart no longer triggers a redundant backup |
| **R-85** | The DR tier is restore-tested **unattended**, and its failure is **heard** |

---

## 2. R-80 — the diagnosis, and what it actually found

The alarm fired at 2026-07-26 03:00 UTC on demo-felhom, demo-hp and drill-r50. The brief described it
as nightly, on three customers, one reaching a customer channel with a 7.3-day staleness claim.

**It fired exactly once.** The check ran and was correctly silent on every prior night. The one
customer-channel mail went to the operator's own mailbox; `peti-felhom` — the only real external
customer — did not fire at all. And the 7.3 days was the age of the **PBS** snapshot, reached only as
a fallback: the local tier had run on 07-24, 07-25 and 07-26.

**Root cause:** the agent's backup store is in-memory (`"lost on restart; the cadence re-populates"`).
The R-50 fleet restart at 12:44 UTC emptied `backups[]`; the next backup landed at 07:03 the following
morning; the 03:00 check fell in that ~18 h blind window and read empty as *no backup exists*.

**One real finding stood:** the PBS/offsite tier had **no schedule at all**. demo-felhom held one
snapshot from a healing artifact; demo-hp held **zero, ever**, while reporting `applied` since 07-21.

---

## 3. R-81 — fixing the class, not the instance

`assessBackupFreshness` collapsed *absence of records* into *failure*. It now returns
**OK / UNKNOWN / MISSED**: absence is UNKNOWN until it outlives a window anchored at first contact.

The anchor was **free** — a Phase-0 probe found the hub already retains 90 days of host-reports, so
`GetHostReportsSince` + `newestBackupEvidence` answer *"when did I last SEE evidence of a backup?"*
rather than *"what does the latest report say?"*. No agent change, no new state.

**Not silence.** A genuinely dead box must still alarm — that is the half the naive fix breaks, and
`TestBackupFreshness_NoEvidenceBeyondAnchor_Alarms` is what makes the suppression safe.

The invariant is now written at the head of the function naming all three instances (v0.12.0,
v0.73.0, R-81), pinned by a boundary test whose name says what it protects.

---

## 4. R-82 — the target split

`BackupTarget()` returned ONE string and `BackupCadence()` ONE 24 h window.

**Phase 0 gates:** weekly CONFIRMED (the only 7-day-exposed state is the non-SMB half of
`settings.json`; `encryption.key` and the offbox credentials are **stable files unchanged since first
boot**, so a week-old copy is byte-identical). The `pvesm status` 0/0/0 anomaly resolved as a
namespace-scoped-token reporting artifact. Capacity raised a **STOP**, which the operator ruled past.

**Measured, not bracketed:** the second weekly snapshot cost **+2.7 GB** against 14.46 GB logical
(~81 % dedup). Weekly top-ups are cheap; **first** snapshots are not. Per-tenant encryption means
**no cross-customer dedup**, so the 80 % warn arrives at roughly the first additional customer.

**Shipped:** per-tier cadence/retention/wait-bound (agent); **one quiesce window for both tiers**
(controller — two cycles would mean two app outages for one night's work); per-tier thresholds
(hub, host 26 h / offsite 8 d); fresh-install default (installer).

### Four defects found by RUNNING it, not reviewing it

1. **v0.98.0** — a 41-minute backup recorded `success:false` at 30 minutes **while still running**,
   then completed `TASK OK`. Worse than "didn't happen": the tier stays due and the retry hits the
   guest lock.
2. **v0.100.0** — the restore tier read from the configured target, not the archive. **A silent
   regression of the S4.1 fix** — the mechanism was never removed; its *input* changed when
   `local_backup_target` was retargeted to `local`.
3. **v0.101.0** — a leaked scratch kept `onboot: 1`; a host reboot would have started a clone of the
   live guest.
4. **v0.102.0** — a tier fired at a storage that does not exist yet.

---

## 5. R-84 — ground truth instead of memory

Three redundant local backups ran in one afternoon of deploys, because a restart empties the store.
On the offsite tier that is a wasted multi-hour upload after **every** agent deploy.

Fixed by **asking the storage**, not persisting the record: a pruned archive correctly stops counting,
where a persisted record would keep claiming a backup that no longer exists. Proven live on both
boxes with the store cold.

---

## 6. R-85 — unattended, and audible

Three defects, each verified at source: the scheduler only ever saw `BackupTarget()`; the spec was
**frozen at daemon start** (an immediately-invoked function — also a latent staleness bug); and a
failed restore-test was a `[WARN]` line with **no event, no notification, no gauge** — true for the
local tier that *was* being tested.

**Ruling (operator): oldest-first.** Never-proven sorts first, which is where the offsite tier starts.
Rotation credit **only on success**, or a permanently failing tier looks freshly proven and stops
being retried. State persisted — unlike R-84 there is no ground truth, because a restore-test destroys
its scratch and leaves no artifact.

**`backup.InFlight`** — a host-wide one-heavy-operation gate. Not a lock concern (the scratch VMID
never touches the live guest's vzdump lock) but a **LINK** one: an offsite restore PULLS multi-GB over
the tunnel a backup PUSHES one. Callers **defer**; they never cancel.

**Two hub signals, never merged:** `restore_test_failed` (broken now) and `restore_test_stale`
(*unverified*, not known-broken). Anchored on R-81, operator-tier only.

---

## 7. Live evidence

**The restore round-trip PASSED** (demo-hp, manual):
```
source_tier: pbs   pass: true   verified: boot+running   mount_parity: ok   4m5s, clean teardown
```
`mount_parity` is the non-hollow half — a boot-only verify cannot see a missing data volume.

**The multi-tier quiesce ran** through the real UI endpoint (authed + CSRF):
exactly **ONE stop/start pair with BOTH backups inside it**, app quiesced through the non-last tier,
early resume on the last tier's snapshot. **Downtime 1m27s for both tiers.**

**Rotation proven unattended** (demo-hp):
```
tier selected (oldest-proven first)  target=felhom-pbs   ← never-proven sorted first
scheduled restore-test passed        duration_s=194.8
tier selected (oldest-proven first)  target=local        ← rotated
```

**The signal fired on its first sweep — on real faults:**
```
[ERROR] drill-r50   Restore-test FAILED on the local tier … HTTP 403 missing privilege VM.Backup
[WARN]  demo-felhom pbs tier: NEVER restore-proven in 205h (limit 168h) — unverified, not known-broken
[INFO]  demo-hp     pbs tier: watching 133h of a 168h grace — newborn, not a fault
```
The first two are genuine, previously-invisible failures. All notifications went to the **operator**
channel; the DB confirms no customer routing.

**demo-hp's first ever offsite backup landed** — 4.25 GB into a verifiably empty namespace.

---

## 8. Mistakes I made, and what they cost

Recorded because the pattern matters more than any one of them.

1. **I claimed the restore-test would break the control plane** by booting a network-conflicting
   clone. **Wrong** — it link-downs every NIC before boot, and that is unit-tested. I read a config
   artifact and inferred behaviour without reading the code that consumes it, *then escalated before
   finishing the check*. I also disabled a safety mechanism on that basis; it is re-enabled.
2. **I pushed one commit red** — read `packages ok: 28` and missed `rc=1` in the same command. Five
   of my own tests were failing.
3. **I planted a time bomb in Slice C** — a test hard-coded the incident timestamp
   `2026-07-18T18:31:06Z` while comparing against the real clock. It passed all day and began failing
   at exactly 18:31 UTC, 8 days later. A test that passes at commit time and fails hours later is
   worse than one that fails immediately.
4. **I left `REPORT.md` claiming "not deployed"** 26 minutes before deploying, and did not update it.
5. **Two bad observables** while verifying the controller: read a UTC timestamp as host time, then
   checked for log lines the agent does not emit at INFO. Both produced confident wrong readings.
6. **I told the operator blocking would not silence the new alerts.** Wrong — blocked customers are
   excluded by `GetActiveCustomerIDs` already.
7. **My own spec's §5.3 was wrong** — it said to force a failure with `--selftest`, but the selftest
   runs in a separate process and never reaches a host-report.

The recurring shape: **inferring behaviour from an artifact instead of reading the code path that
consumes it**, and **reading a result without reading its exit code**. The host=CEST /
component=UTC mismatch caught me twice in one session.

---

## 9. State at close

| | |
|---|---|
| Backups | local daily + offsite weekly, both boxes |
| Offsite retention | 2 weeks (`keep_last=2`) |
| Restore-test | **rotating both tiers, unattended**; interim cadence 3.5 d → each tier ~weekly |
| Signals | failure + staleness, operator-only |
| drill-r50 | VM 300 **stopped**, `onboot=0`. Operator will block the customer |

### ⏳ IN FLIGHT AT CLOSE — check on this

**demo-felhom's offsite restore-test is still running.** Its archive is 14.46 GB, so expect roughly
two hours. Scratch guest `990000` is on the box while it runs; a clean finish tears it down.

Its agent still has the **compressed 10-minute cadence in memory** — the reverted 3.5-day value is on
disk and takes effect at the next restart. A watcher is set to restart the agent once the test
finishes. **If that watcher did not complete, restart `felhom-agent` on demo-felhom manually** or the
box keeps restore-testing every 10 minutes.

When it passes, demo-felhom's `restore_test_stale` clears on its own.

---

## 10. NOT done — explicitly

1. **The restic (app-data offsite) tier is NEVER restore-tested** → **R-87**. R-85 covers whole-guest
   vzdump tiers only. This is arguably the tier that matters most: the only one that survives losing
   the box **and** carries the customer's app data (the whole-guest snapshot excludes the
   bind-mounted drives). It is exactly the state PBS was in yesterday.
2. **Restore-tests are interval-scheduled, not backup-aligned** → **R-86** (operator ruling
   2026-07-27: ~1 day after that tier's own backup).
3. **Rotation observed on demo-hp only**, and under a compressed cadence. It proves the rotation
   logic, not the production interval.
4. **The hub infers "PBS ⇒ weekly" from storage TYPE**, not a reported cadence. No box is in the
   wrong shape today.
5. **The installer-default fleet flip** awaits a full weekly cycle.
6. **`07-backup-architecture.md` is NOT ratified** — brought current with an honest staleness header;
   ratification is Viktor's review of the §10 list.
7. **The customer-facing Hungarian copy still overstates scope** — unchanged since R-80 flagged it.
8. **Capacity:** the datastore still needs growing before the first real customer.

---

## 11. Observations

- **drill-r50's 403** (`missing privilege VM.Backup at /vms/9100`) is real and was invisible. The box
  is being retired, so it needs a decision — fix or accept — not silence.
- **A blocked customer is hidden from the Dashboard and from every checker**, and it is reversible
  (`configs.go` sets status back to `active`, with a UI affordance). Blocking is a blunt instrument
  though: it silences *everything*, which is why the operator is building per-alert silencing.
- **The PVE UI will permanently show `felhom-pbs` at 0 %** (namespace-scoped token). Read fill from
  the hub's PBS-DR gauge.
- **demo-felhom's guest grew 9.74 → 14.46 GB logical in eight days.** Probably one-off from app
  testing, but if it is a rate the capacity sizing changes quickly.
- **Three of the five items in this arc were the same bug class** — absence of a signal read as
  evidence, or a correct mechanism whose input changed underneath it. Both are now written into the
  code as invariants rather than left as lore.