SPIKE ep0 connections: the leak is felhom-agent's, not the poll rate
gates / gates (push) Successful in 14s
gates / gates (push) Successful in 14s
R-341's first dated check, taken at +46.2 h: fd 17 -> 405 over 166,251 s
= 201.6/day. Pre-registered range was 370-450; observed 388. UNCHANGED,
as predicted. CLOSE-WAIT is 0 -- absent entirely, not merely flat.
Q1: exactly two peers, 194 each, no third party.
Q2: outcome (a). ep0 388 = 194 + 194 on the boxes, twice, and the four
new sockets carry the same source ports on both sides. 0 closed in 31 min.
The finding: all 388 are held by felhom-agent. pvestatd and
proxmox-backup-client made 162,404 requests and leaked zero. Mechanism is
a per-cycle http.Transport with a zero-value IdleConnTimeout that nothing
ever closes (internal/pbs/client.go:56, main.go:1486). R-336's premise
does not survive this -- cutting the poll rate would have fixed nothing.
Q3 NOT measured: Phase C held at STOP 1, prediction pre-registered first.
Part 0 captures the due-checks gate's first conviction on a real overdue
date (rc=1, names R-341, sole failure among 10 gates). Row cleared at
Part 4, after the result was recorded in R-341, not to make a push work.
New: R-344 (the transport leak), R-345 (hub/Makefile pushes :latest),
R-346 (ActiveEnterTimestamp reads 5h56m early -- NRestarts is still 0).
This commit is contained in:
1 parent
848368ec38
commit
19672e685e
22 files changed
+1330
-8
No files matched your search
@@ -636,16 +636,19 @@ class (an image `VOLUME` at an unmounted path) is still live — `immich-server`
|
||||
| **R-333** | **Two disk-health questions the deploy raised and did NOT act on.** **(a) The 55/60 °C bands are SPINNING-DISK bands applied to NVMe.** They were adopted unchanged from the operator's Prometheus config so the two systems cannot disagree — a deliberate, stated decision — but **measured on demo-hp 2026-08-14 the healthy Toshiba KXG50PNV1T02 NVMe idles at 53 °C, two degrees below Figyelmeztetés and seven below Hiba**, and NVMe routinely exceeds 60 °C under load with no fault whatever. As it stands a healthy customer NVMe under sustained write can be reported as **Hiba** — the single worst outcome this feature can produce. **(b) The agent runs bare `smartctl -a -j` with no `-n standby`** (`felhom-agent/internal/storage/hostops.go:368`), so every poll WAKES a spun-down drive; going 6h → hourly multiplies that by six. demo-hp is all-flash so the cadence measurement could not reveal it, and it was recorded rather than acted on per the task's own instruction. Mitigating datum from the fixture: the failing drive logged only **3375 load cycles in 60505 hours** (~one per 18h), i.e. that duty cycle barely spins down at all | **READY (S each) — NEW 2026-08-14** | — | (a) split the temperature bands by device class, or drop them for NVMe and rely on `critical_warning`; (b) add `-n standby` to the agent's smartctl invocation (an agent change, so fold it into R-330's session) | Viktor decides (a); CC does (b) |
|
||||
| **R-334** | **CLOSED 2026-08-18 — golden 0.216.0 baked, published and VOUCHED; CI green by run id.** ~~WAIVER + open item: controller v0.215.0 is released and deployed, and NO golden carries it.~~ Convicted by `golden_currency_gate.py` on the 2026-08-14 push: newest released controller **0.215.0**, newest golden bake **0.214.0** (`documentation/tests/golden-0.214.0-2026-08-12`). **A machine installed right now receives 0.214.0** — i.e. a brand-new box would ship WITHOUT the R-328 severity fix and would keep emailing nobody about a failing disk. The running fleet is unaffected (demo-hp guest 9201 is on 0.215.0 and healthy); this is purely the day-0 install path. **Not baked in this session deliberately:** the task scoped deployment to demo-hp only, and the second half of the fix — vouching the bake in the hub's day-0 artifact manifest — is **operator-password-gated, so CC cannot complete it**; a baked-but-unvouched golden is worse than none. **The push was made with `git push --no-verify` and it is stated here and in the session report**, per `.claude/rules/gates.md` — the gate has no waiver parser, so recording a waiver does not clear it. **STILL OPEN and now one version WIDER, 2026-08-18:** the newest released controller is **0.216.0** (v0.215.0's own follow-up fix, R-335) and the newest bake is still **0.214.0**, so a new install now misses *two* releases. Re-convicted on this date's documentation-only push, which was likewise made with `--no-verify`; the gate reads `felhom-controller/CHANGELOG.md` and `documentation/tests/golden-*`, **neither of which that session touched** — the conviction is inherited, not caused | **CLOSED 2026-08-18.** Baked from `RUNBOOK-manual-build.md` §4.0+§4.1 in the DooPlex drill VM and published: **`GOLDEN_VERSION=0.216.0`**, **`GOLDEN_SHA256=ac004dc90d8cefccc5448377892f9cff3a4c3e1e27d0e11129120e38ac31c34b`**, 656,970,239 bytes at `…/generic/felhom-golden/0.216.0/golden.tar.zst`. **The published bytes were verified, not just the script's print** — the artifact was downloaded back out of Gitea and hashed, and it matches. **Vouched by the operator, all THREE fields together**, confirmed by reading the hub's own store rather than the save: `artifact_golden_version=0.216.0`, `artifact_agent_version=0.129.0`, `artifact_min_agent=0.129.0` (2026-08-18 11:00:59–11:01:00), and the hub's recorded sha256 matches the downloaded artifact. The R-216 shape was checked on the machine: `MinAgent` 0.129.0 is **equal to**, not above, the newest **published** agent. **`golden_currency_gate.py` rc=0 and `repo_gates.py --fast` rc=0 — all nine gates — and CI is GREEN BY RUN ID: run **353**, `head_sha 7d81681d6`, conclusion `success`** (the two prior runs 351/352 on this same afternoon were red on exactly this row, which is the contrast). That push needed **no `--no-verify`** — the first of the day that did not. Evidence: `documentation/tests/golden-0.216.0-2026-08-18/`, report `REPORT-golden-0.216.0.md`. **Closed with the run id quoted deliberately**: this row was re-confirmed once and widened once, and closing it on a local green a third time would have left the same ambiguity | — | Bake a golden on **0.216.0** per `runbooks/RUNBOOK-manual-build.md` §4.1, then vouch it — a THREE-field change (`golden_version` + `agent_version` + `min_agent`). Until then every NEW install lacks the severity fix | CC bakes; **Viktor vouches** |
|
||||
| **R-335** | **One physical disk was walked TWICE per run, and the second walk sustained it against itself.** Found on live hardware ~2h after the v0.215.0 deploy, **by noticing the release's own positive observable disagreed with its own persisted artefact**: the hourly check logged *"3 disk(s) evaluated"* while `disk-health-state.json` held **two** records. Cause: demo-hp's `c11-scratch` and `felhom-backup` are the same physical NVMe (`/dev/nvme0n1`) and resolve to the same `diskKey`. **Not cosmetic** — `RunDiskHealthCheck` writes a disk's new record before the next entry reads it, so the SECOND copy consumed the FIRST copy's write as its prior: the disk **sustained against itself and reached Hiba on a FIRST sighting**, defeating truth-table row 6 — the exact rule separating a one-hour benign excursion from a false critical — and would have emitted **two identical events** for one drive. **Latent, not active, on demo-hp** (all three entries healthy, zero counters), but any aliased disk developing a single pending sector would have gone straight to Hiba. **This is the shape standing rule 3 warns about: an absent alarm was not evidence — the two artefacts had to be read AGAINST each other** | **CLOSED — controller v0.216.0, 2026-08-14.** Each `diskKey` is evaluated once per run; both entries stay marked `seen` so neither looks like a disappeared disk, and the card still renders both storage rows (the dedup is about state and alerts, not display). Pinned by `TestDiskCheck_SameDiskTwiceIsEvaluatedOnce`; companion red-proof run and reverted — deleting the guard makes the first sighting emit `Kind:2` (Hiba-from-sectors) at 8 sectors | — | — | CC |
|
||||
| **R-336** | **The offsite DR endpoint is polled about once per second, and that is what turned a slow leak into an outage.** ep0's PBS proxy served **~85,000 requests/day** — a flat **3,538/hour**, every hour, from two boxes: `74,445 GET /api2/json/admin/datastore` (`libwww-perl`, i.e. PVE's `pvestatd`) and `73,171 GET /admin/datastore/felhom-offsite/status` (`proxmox-backup-client`). Two pollers asking substantially the same question at the same rate. On 2026-08-18 this walked a connection leak in the proxy to its 1024-fd soft limit in **14 days**, wedging the offsite tier for 9½ hours (`audits/INCIDENT-ep0-pbs-fd-exhaustion-2026-08-18.md`). The `LimitNOFILE=65536` drop-in applied that morning raises the ceiling **but does not fix the leak** — it converts a fortnightly outage into a multi-year one, which is mitigation, not a fix. A DR endpoint that is written to weekly does not need to be asked about every second | **READY (M) — NEW 2026-08-18** | — | **CORRECTED 2026-08-18 (evening) — the easy lever named here does not exist.** This cell used to read *"PVE storage status is the prime suspect, and its interval is tunable"*. **The first half is right and the second half is false.** `pvestatd` stats EVERY configured storage on each 10-second cycle, and Proxmox staff have stated the interval is not designed to be configurable — so there is no knob to turn down. The only lever PVE actually offers is disabling the storage entry (`pvesm set <id> --disable 1`) around the backup window, and that is **substantially more than a tuning knob**: it collides with `felhom-agent/internal/pbsdr/manager.go`'s health model, where an inactive-but-existing entry drives the consume-the-one-time-secret recovery path. So the fix is a design question (does the hub still need a 15-minute fill reading at all, given R-339 now reports reachability separately?), not a config edit. **Doc-only correction — no agent code was changed.** The remaining step is unchanged: cut the poll rate by whatever means survives that question, then confirm the fd count between restarts stops climbing — the positive observable, per standing rule 3. **Baseline measured 2026-08-18, and the FIRST measurement published was WRONG.** The initial "~85/day, matching the ~73/day implied by the failure" came from a single 17-minute window whose delta was **one descriptor** — a sample of one cannot carry a daily rate, and the agreement that made it feel solid was coincidence. **Re-measured over two independent windows the same morning: 183/day (31 min) and 200/day (5.6 h)** — ~2.6x the published figure, putting the runway to the 65536 ceiling at **~357 days, not the ~2 years first claimed**. **And the named mechanism is the minority one:** across that window `CLOSE-WAIT` held flat at 1 while `ESTAB` grew 45→49 — *all* the growth was established connections, and at the wedge the split was 1011 ESTAB / 543 CLOSE-WAIT. **The fix must target connections the proxy never reaps, not just `CLOSE-WAIT` sockets.** The PBS 4.2.5-1 upgrade (2026-08-18) did NOT change the slope and was never expected to — see R-341 | CC |
|
||||
| **R-336** | **The offsite DR endpoint is polled about once per second, and that is what turned a slow leak into an outage.** ep0's PBS proxy served **~85,000 requests/day** — a flat **3,538/hour**, every hour, from two boxes: `74,445 GET /api2/json/admin/datastore` (`libwww-perl`, i.e. PVE's `pvestatd`) and `73,171 GET /admin/datastore/felhom-offsite/status` (`proxmox-backup-client`). Two pollers asking substantially the same question at the same rate. On 2026-08-18 this walked a connection leak in the proxy to its 1024-fd soft limit in **14 days**, wedging the offsite tier for 9½ hours (`audits/INCIDENT-ep0-pbs-fd-exhaustion-2026-08-18.md`). The `LimitNOFILE=65536` drop-in applied that morning raises the ceiling **but does not fix the leak** — it converts a fortnightly outage into a multi-year one, which is mitigation, not a fix. A DR endpoint that is written to weekly does not need to be asked about every second | **READY (M) — NEW 2026-08-18** | — | **CORRECTED 2026-08-18 (evening) — the easy lever named here does not exist.** This cell used to read *"PVE storage status is the prime suspect, and its interval is tunable"*. **The first half is right and the second half is false.** `pvestatd` stats EVERY configured storage on each 10-second cycle, and Proxmox staff have stated the interval is not designed to be configurable — so there is no knob to turn down. The only lever PVE actually offers is disabling the storage entry (`pvesm set <id> --disable 1`) around the backup window, and that is **substantially more than a tuning knob**: it collides with `felhom-agent/internal/pbsdr/manager.go`'s health model, where an inactive-but-existing entry drives the consume-the-one-time-secret recovery path. So the fix is a design question (does the hub still need a 15-minute fill reading at all, given R-339 now reports reachability separately?), not a config edit. **Doc-only correction — no agent code was changed.** The remaining step is unchanged: cut the poll rate by whatever means survives that question, then confirm the fd count between restarts stops climbing — the positive observable, per standing rule 3. **Baseline measured 2026-08-18, and the FIRST measurement published was WRONG.** The initial "~85/day, matching the ~73/day implied by the failure" came from a single 17-minute window whose delta was **one descriptor** — a sample of one cannot carry a daily rate, and the agreement that made it feel solid was coincidence. **Re-measured over two independent windows the same morning: 183/day (31 min) and 200/day (5.6 h)** — ~2.6x the published figure, putting the runway to the 65536 ceiling at **~357 days, not the ~2 years first claimed**. **And the named mechanism is the minority one:** across that window `CLOSE-WAIT` held flat at 1 while `ESTAB` grew 45→49 — *all* the growth was established connections, and at the wedge the split was 1011 ESTAB / 543 CLOSE-WAIT. **The fix must target connections the proxy never reaps, not just `CLOSE-WAIT` sockets.** The PBS 4.2.5-1 upgrade (2026-08-18) did NOT change the slope and was never expected to — see R-341 **SPIKE 2026-08-20 — THE PREMISE OF THIS ROW DOES NOT SURVIVE MEASUREMENT, and that is a change in what the row IS, not new evidence on it.** `audits/SPIKE-ep0-established-connections-2026-08-20.md`. **The leak is OURS, and the poll rate is not what feeds it.** Every one of the 388 leaked descriptors is an ESTABLISHED connection held open by **`felhom-agent`** on the boxes — 194 on each, `ss -tnp` naming a single PID per box, and **zero** held by `pvestatd` or `proxmox-backup-client`. Confirmed independently from ep0's access log over the same 46.18 h window: `libwww-perl` (pvestatd) **81,192 requests -> 0 descriptors**, `proxmox-backup-client` **80,061 requests -> 0 descriptors**, `Go-http-client/1.1` (the agent) **387 `/snapshots` calls -> 388 sockets — one per call, within one**. So **162,404 requests, 99.5% of the traffic, produce 0% of the leak.** **Mechanism, named from source:** `felhom-agent/internal/pbs/client.go:56-60` builds `&http.Transport{TLSClientConfig: tlsCfg}` — a composite literal, so `IdleConnTimeout` is the zero value = **no limit** (`http.DefaultTransport` sets 90 s; a literal does not inherit it) — and `cmd/felhom-agent/main.go:1486` (`pbsTargetsFromPVE`) builds **a fresh client every cycle**, as its own doc comment states. Each cycle therefore strands one idle keep-alive connection in a transport nothing ever closes; `CloseIdleConnections`/`IdleConnTimeout`/`MaxIdleConns` appear **nowhere** in the agent repo. Cadences reconcile without fitting: 900 s hub poll (184.7 cycles) + 6 h `DefaultVerifyCadence` (7.7 cycles) = 192.4 predicted vs **194 observed per box**. **CONSEQUENCE — RE-RANK.** The remaining step recorded above ("cut the poll rate, then confirm the fd count stops climbing") **would have produced a null result and read as a failed fix.** Cutting the Proxmox poll rate removes ~99.5% of ep0's reqLine truncated
|
||||
| **R-337** | **`/backup/status` lagged a completed backup by minutes on one box and not the other — and it RESOLVED ITSELF, which is why this is WATCHING and not a defect.** During the R-336 recovery on 2026-08-18, `demo-hp`'s snapshot landed on ep0 at **03:58:43Z** (complete manifest; the host's own task index says `OK`) — yet `GET /backup/status` was **still serving the superseded 03:27:00Z failure at ~04:03Z**, four-plus minutes later. `demo-felhom` showed its new result within ~40 s of completion. **The lag cleared on its own:** demo-hp's 04:07:35Z host report carries `felhom-pbs success=true, 4.29 GB`, and the hub is green for both boxes. **The first draft of this row claimed the success was "still reported as failed" — that was written before the next report arrived and it was wrong; the corrected claim is a several-minute skew between the two boxes, not a stuck value.** It is recorded because a status field that can trail its own artifact by minutes will, during an incident, be read as a second failure — this session nearly did — and because the asymmetry between the two boxes is unexplained | **WATCHING — NEW 2026-08-18** | another observation, ideally during an incident rather than constructed | **Do not open a fix on this as written.** First establish the intended refresh path for `/backup/status` after an out-of-schedule run; only if the skew is not simply collection cadence is there anything to pin. If it is cadence, close this row and say so | CC |
|
||||
| **R-338** | **`demo-hp` is not on the R-50 island at all, and `operations/nodes.md` states that it is.** The page records both fleet boxes as island-migrated 2026-07-25. True of `felhom-pve`; **false of `demo-hp`**, whose `agent.json` has `listen_addr: 192.168.0.87:8443` — the customer LAN address — and **no `island_bridge`/`island_guest_addr` keys at all**, whose guest 9201 has `net0` only (no `eth1`), and whose `vmbr9` exists with **zero members**. The controller's `controller.yaml` points at the LAN address, so the box works; this is inventory drift, not breakage. **Two costs.** A session trusting the page addresses the wrong endpoint — that happened on 2026-08-18 and the resulting timeout was briefly read as a fault. And the agent's local API is **bound to the customer LAN on this box** rather than to a point-to-point island, which is the exposure R-50 was built to remove — so a documented security property is claimed for a box that does not have it | **READY (S) — NEW 2026-08-18** | — | Decide which is true: migrate `demo-hp` to the island, or correct `nodes.md`. Leaving both is the one option that keeps the doc lying | Viktor decides; CC executes |
|
||||
| **R-341** | **Does the fd slope change after the PBS 4.2.5 upgrade? — two dated checks, and the answer is expected to be NO.** ep0 was upgraded 4.2.2-1 → 4.2.5-1 on 2026-08-18 09:51Z on the operator's ruling, **for rehearsal value, not as a fix**: the full changelog range was read (128 lines, all three entries) and swept for connection-handling vocabulary, and it contains **no mechanism** by which descriptor reaping would change — the single keyword hit was `S3 … honor the node's proxy settings`, HTTP-proxy config for S3, not the PBS proxy daemon. **The 32-minute post-upgrade window is indistinguishable from the before window** (+5 fd/1919 s = 225/day vs +4 fd/1885 s = 183/day; the two differ by ONE descriptor and Poisson uncertainty on such counts is ±2, so both are consistent with one unchanged rate — the higher after-figure is noise, not a regression). **Thirty minutes cannot settle it in either direction and this row exists so nobody pretends it did.** New `t0` = **fd 17 at 2026-08-18 09:51:22Z, proxy PID 551655**; before-rate to beat = **183–200/day**. **Interpretation fixed in advance** (`evidence-ep0-pbs-upgrade-2026-08-18/stop1-ruling.txt`, written before any numbers existed): unchanged = EXPECTED, not a failed upgrade; changed = a SURPRISE needing explanation, not a confirmation | **WATCHING — NEW 2026-08-18** | elapsed time only | **Two dated checks, both CC:** **+24 h — 2026-08-19 ~10:00Z** and **+7 d — 2026-08-25 ~10:00Z**. Command (the incident's own positive observable): `ssh root@<ep0> 'PID=$(systemctl show proxmox-backup-proxy -p MainPID --value); ls /proc/$PID/fd \| wc -l; ss -lnt "( sport = :8007 )"; ss -tn state all "( sport = :8007 )" \| awk "NR>1{print \$1}" \| sort \| uniq -c'`. **Record the ESTAB/CLOSE-WAIT split, not just the total** — the split is what says which leak it is. If PID ≠ 551655 the window is void: something restarted the proxy and the count began again | CC on both dates |
|
||||
| **R-341** | **Does the fd slope change after the PBS 4.2.5 upgrade? — two dated checks, and the answer is expected to be NO.** ep0 was upgraded 4.2.2-1 → 4.2.5-1 on 2026-08-18 09:51Z on the operator's ruling, **for rehearsal value, not as a fix**: the full changelog range was read (128 lines, all three entries) and swept for connection-handling vocabulary, and it contains **no mechanism** by which descriptor reaping would change — the single keyword hit was `S3 … honor the node's proxy settings`, HTTP-proxy config for S3, not the PBS proxy daemon. **The 32-minute post-upgrade window is indistinguishable from the before window** (+5 fd/1919 s = 225/day vs +4 fd/1885 s = 183/day; the two differ by ONE descriptor and Poisson uncertainty on such counts is ±2, so both are consistent with one unchanged rate — the higher after-figure is noise, not a regression). **Thirty minutes cannot settle it in either direction and this row exists so nobody pretends it did.** New `t0` = **fd 17 at 2026-08-18 09:51:22Z, proxy PID 551655**; before-rate to beat = **183–200/day**. **Interpretation fixed in advance** (`evidence-ep0-pbs-upgrade-2026-08-18/stop1-ruling.txt`, written before any numbers existed): unchanged = EXPECTED, not a failed upgrade; changed = a SURPRISE needing explanation, not a confirmation | **WATCHING — FIRST CHECK TAKEN 2026-08-20: the rate is UNCHANGED, exactly as pre-registered. Second check still due 2026-08-25** | elapsed time only | **Two dated checks, both CC:** **+24 h — 2026-08-19 ~10:00Z** and **+7 d — 2026-08-25 ~10:00Z**. Command (the incident's own positive observable): `ssh root@<ep0> 'PID=$(systemctl show proxmox-backup-proxy -p MainPID --value); ls /proc/$PID/fd \| wc -l; ss -lnt "( sport = :8007 )"; ss -tn state all "( sport = :8007 )" \| awk "NR>1{print \$1}" \| sort \| uniq -c'`. **Record the ESTAB/CLOSE-WAIT split, not just the total** — the split is what says which leak it is. If PID ≠ 551655 the window is void: something restarted the proxy and the count began again **FIRST CHECK — TAKEN 2026-08-20 08:02:13Z, and it was taken at +46.2 h, NOT +24 h.** No session ran between 18 and 20 August, so the 2026-08-19 date passed as pure elapsed time; the delay is stated here rather than backfilled. **The longer window is a BETTER measurement, not a degraded one** — the leaked count is **388** against the 4 and 5 descriptors the original answer rested on, roughly **97x**, so the uncertainty falls from about +/-50% to about +/-5%. **Precondition PASSED:** PID still **551655**, `ps -o lstart` 2026-08-18 09:51:04, `NRestarts=0` on both units. **Result:** fd **17 -> 405** over **166,251 s** = **201.6 fd/day**, Poisson +/-10.2/day (1s), 2s band **181.2-222.1**. Pre-registered range was **370-450** (confirm-band 330-490); **observed 388**, near the centre. **VERDICT: unchanged — the EXPECTED result, and not a failed upgrade.** **Composition:** ESTAB 0 -> 388, **CLOSE-WAIT 0 — absent from the histogram entirely**, so 100% of the growth is established connections and `CLOSE-WAIT` is not merely the minority half. Runway from fd 405 at 201.6/day to the 65536 ceiling: **~323 days (~2027-07-09)**. Full working: `audits/SPIKE-ep0-established-connections-2026-08-20.md` + `evidence-ep0-established-connections-2026-08-20/step1-slope-computation.txt`. **MEASUREMENT TRAP for the 2026-08-25 check, found on this run — see R-346:** the anchor must be `ps -o lstart= -p $MainPID`, NOT `systemctl show -p ActiveEnterTimestamp`, which reads 03:54:54Z for this generation (the upgrade re-exec'd the proxy; systemd never saw a stop, `NRestarts` is still 0) and would put the rate ~15% low. **PERTURBATION NOTE for the 2026-08-25 check:** this spike's Phase C quietens `pvestatd` on demo-hp for one overnight window inside that 7-day interval. Quantified so nobody reads the shortfall as a change in the leak: even if the leak were fully proportional to the request rate (which Parts 1-2 predict it is NOT), a 10 h half-rate window inside 168 h shifts the 7-day slope by **~3%**, far inside the +/-10% Poisson band on a ~1,400-descriptor count. **The +7 d reading remains usable.** | CC on both dates |
|
||||
|
||||
| **R-342** | **The ep0 snapshot covers less than it looks like it covers, and the next person will assume otherwise.** Quoting `audits/evidence-ep0-pbs-upgrade-2026-08-18/stop2-snapshot.txt` verbatim: *"covers — the 38 GB system disk /dev/sda (root), i.e. the PBS packages, unit files, /etc/systemd drop-ins, nftables and wg config. DOES NOT — /mnt/pbs-datastore. That is /dev/sdb, a separate 100 GB VOLUME, and Hetzner server snapshots do not include attached volumes. The backup data is therefore NOT protected by this snapshot."* Snapshot **421440873** (`felhom-hetzner-20260818`, 15.06 GB, Available) was taken as the rollback for the 4.2.2→4.2.5 PBS upgrade. **Rolling it back restores software state, not the datastore.** That was *acceptable for that change* — a package install writes no datastore content — and the file says so. **The problem is what happens next:** this fact lives in an evidence file nobody will open again, and a snapshot named as "the rollback" reads as protecting everything on the box. ep0 holds the only off-premises copy of a real customer's data | **READY (S) — NEW 2026-08-18** | — | **Decide the safeguard for any future ep0 procedure that could touch `/mnt/pbs-datastore` — it does not exist and has not been designed.** Candidates: a Hetzner **Volume** snapshot (a different object from the server snapshot), a PBS-level sync to a second location, or an explicit written acceptance that the datastore is unprotected for the duration. **Nothing may be added to a runbook implying a safeguard exists until one does** | **Viktor decides; CC executes** — a risk-to-customer-data question |
|
||||
| **R-343** | **The managed controller floor was raised 0.214.0 → 0.216.0 — and it was NOT the no-op it was expected to be: it moved a live box nine seconds later.** Raised by the operator 2026-08-18 **12:36:58Z**, in a **separate save after** the artifact vouch. **Read back from the store, not the form** (`hub_settings.min_controller_version`, `GetGlobalMinControllerVersion` — `hub/internal/store/store.go:1751`): `min_controller_version = 0.216.0`, updated 12:36:58. **WHY IT HAD BEEN BEHIND — this was NOT drift, and describing it as "two releases behind" without this context reads as a defect it was not.** `publish-train-rules.md` rule 1 is *manifest before floor*, and rule 2 requires the floor field to be filled **LAST, in a separate save**, because the DB row overrides the env floor and **acts immediately on the next report cycle**. That rule was earned: on the 2026-07-11 publish train the floor was saved together with the manifest, acted at once, and pushed controller 0.113.0 onto Peti's box **~9 minutes ahead of agent 0.81.0** — the exact forbidden skew, benign only because that box had no NAS shares. R-120's row records the same deliberate choice (*"Floor untouched per publish-train rule 2"*). So the floor sitting at 0.214.0 was **policy being followed**, not neglect. **What it was functionally while it sat there:** not a live problem — every reporting box was at or above it — but a **safety net set two versions low**. The floor is what drags a box forward if it ever falls behind (restored from an old backup, reinstalled, long offline), and at 0.214.0 it would have pulled such a box only to two versions back, missing R-328's severity fix and R-335's follow-up. **THE MEASURED BLAST RADIUS — five reads, and the third and fifth are the findings.** **(1) Floor:** `0.216.0` @ 12:36:58Z, from the store. **(2) Per-customer overrides:** **zero** — all five `customer_configs` rows carry an empty `min_controller_version`, so nothing hides behind a lower override and the global applies to everyone. **(3) Every box's controller version:** `demo-felhom` **0.216.0**, `demo-hp` **0.216.0** — both AT the floor **now**; `drill-r50` **0.213.0** (status `blocked`, last report 2026-08-12, powered off/reverted) and `peti-felhom` **0.115.0** (host row DELETED) are **below** it but are **not reporting boxes**. **(4) Directives/holds:** no `managed floor HELD` line exists; the hub logged `[INFO] Global controller-version floor set to "0.216.0"`. **(5) THE FINDING — a controller DID auto-update after the raise.** `demo-felhom` had been on **0.214.0** since 2026-08-12 16:44 and the hub recorded `controller_updated — Controller frissítve: 0.214.0 → 0.216.0` at **12:37:07Z**, then `controller_started (0.216.0)` at 12:37:12Z — **nine seconds after the raise**, exactly the immediate action rule 2 documents. `demo-hp` was already on 0.216.0 (hand-deployed 2026-08-14 08:31) and did not move. **No error, warning or critical event followed** — the update completed and the controller came back up. **So the change was real, not inert: "every reporting box is at or above the floor" is true BECAUSE of the raise, not independently of it.** **Why it is safe by construction**, cited rather than asserted: `ResolveManagedFloor` (`hub/internal/store/store.go:2068`) sets `Held` and clears the floor entirely when `Floor > GoldenVersion` (the R-216 shape) — floor 0.216.0 **equals** golden 0.216.0, so that guard does not trip — and holds per-box when the box's agent is below the manifest's `MinAgent`, or unknown, or unparseable; both boxes report agent 0.129.0 against `MinAgent` 0.129.0, so the floor was served rather than held. That second guard remains armed for any box reporting with an old agent. **No ISO rebuild is required:** the golden is fetched at first boot from the hub's manifest, which is 0.216.0 — at the floor, not below it — and rule 5's `assert_golden_ge_floor` is a **build-time** gate for *future* builds (`scripts/iso/build-felhom-iso.sh:77`, called at **:267**) which **fails open with a warning** when its inputs are absent (`:78-82`: `if [[ -z "$golden" \| \| -z "$floor" ]]` → `log_warn "… UNENFORCED …"` → `return 0`), both confirmed in the script. **`peti-felhom` was NOT contacted and needs no contact** — from the PETI row: its host row was deleted 2026-07-15 and *"a report from a deleted host 401s and is not persisted"*, so it cannot receive a floor directive at all and the raise cannot reach it | **OPEN — NEW 2026-08-18.** Deliberately NOT closed: the task's closing condition was *all five reads clean, no directive served*, and read 5 shows a live box updated. It went cleanly and is the floor working as designed — but a change recorded as a no-op when it moved a customer box is exactly the kind of record that misleads later | — | **Confirm the 0.216.0 update on `demo-felhom` is healthy in normal operation** (it reported and restarted clean, but it has not yet run a full bacLine truncated
|
||||
| **R-344** | **`felhom-agent` leaks one TCP connection to PBS per poll cycle, forever, on both sides — and it is the whole of the ep0 descriptor leak.** Found by the 2026-08-20 connections spike (`audits/SPIKE-ep0-established-connections-2026-08-20.md`), which was looking for a Proxmox poll-rate problem and found ours instead. **Two defects compounding.** **(a)** `internal/pbs/client.go:56-60` builds `&http.Transport{TLSClientConfig: tlsCfg}` as a composite literal, so **`IdleConnTimeout` is the zero value = no limit** — `http.DefaultTransport` sets 90 s and a literal does not inherit it. **(b)** `cmd/felhom-agent/main.go:1486` (`pbsTargetsFromPVE`) builds **a fresh `pbs.Client` every cycle**, as its own doc comment states, so each cycle strands one idle keep-alive connection in a transport that is then unreachable — and an unreachable `http.Transport` does **not** close its connections; the `persistConn` read-loop goroutine keeps the socket alive. `CloseIdleConnections` / `IdleConnTimeout` / `MaxIdleConns` appear **nowhere in the repo** (grep: no matches). **Measured live, both sides, twice:** ep0 held 388 ESTAB (194 from each box) at 08:02:39Z and 392 at 08:33:42Z; the boxes held 194+194 and 196+196 at the same instants, and the four new sockets carried **the same four source ports** on both sides. **Zero sockets closed in 31 minutes**, and all carried keepalive timers with `retrans=0` — mutually held live idle connections, not half-open ones. One socket per agent `/snapshots` call (387 calls vs 388 sockets). Rate **201.6/day** across the fleet; ep0 reaches its 65536 ceiling in **~323 days**. **It is bilateral and the box side is the under-watched half:** each agent holds **196 of its 208** descriptors in these sockets. Its limit is 524287, so the boxes are in no danger *today* — which is why this stayed invisible, not why it is harmless. | **READY (S) — NEW 2026-08-20** | none — it is a self-contained change in `felhom-agent` | **No fix is proposed here: the spike-first gate forbids it and the spec is a separate task.** What the spec must settle, and none of it is decided: whether the per-cycle client construction is the thing to remove or the transport is the thing to share; what `IdleConnTimeout` should be given the 900 s poll and the 6 h verify cadence; whether `internal/hub/client.go:53` and `internal/proxmox/client.go:69` — **the same composite-literal pattern, built once at start-up so not leaking by this route today** — should be changed in the same pass or left alone with a test pinning why. **The proof obligation is the fd count, not the diff:** per standing rule 3 the positive observable is ep0's ESTAB count going FLAT between proxy restarts, measured over a window long enough to matter — a green test suite proves nothing here, and a 30-minute window proves nothing here either (that error is already recorded twice in R-336 and R-341). | CC |
|
||||
| **R-345** | **`hub/Makefile` tags and pushes `:latest`, which the project's own rules forbid in two places.** Lines 21-22 of `docker-push`: `docker tag $(IMAGE):$(VERSION) $(IMAGE):latest` then `docker push $(IMAGE):latest`. `.claude/rules/hub.md:35` says *"Pin explicit versions, never `:latest`"* and `.claude/rules/manifests.md:15` repeats it. Verified present on `848368ec3`. Small, and the deployed manifests do pin a version, so nothing is currently broken by it — but a documented command that performs the prohibited action is a trap for whoever next reads the Makefile as the reference for how to publish, and a floating `:latest` on the registry is exactly the thing an emergency `kubectl set image` reaches for. Noticed while running the 2026-08-20 connections spike; unfiled until now. | **READY (XS) — NEW 2026-08-20** | — | Delete the two lines, or keep them behind an explicit opt-in target that says in a comment why it exists. Check whether a stale `:latest` tag already sits on `gitea.dooplex.hu/admin/felhom-hub` before deciding — an existing floating tag is the more dangerous half. | CC |
|
||||
| **R-346** | **`ActiveEnterTimestamp` answers a different question than the one a slope measurement asks, and on ep0 right now it is wrong by 5 h 56 m.** Found while taking R-341's first dated check. ep0's `proxmox-backup-proxy` has `MainPID=551655` started **2026-08-18 09:51:04Z** (`ps -o lstart=`), but `systemctl show -p ActiveEnterTimestamp` reads **03:54:54Z** and `NRestarts` reads **0** — because the 4.2.5-1 upgrade **re-exec'd** the daemon rather than restarting the unit, so systemd never observed a stop. Anyone anchoring "when did this proxy generation start" on `ActiveEnterTimestamp` would divide 388 descriptors by 52.1 h instead of 46.2 h and report **178/day instead of 201.6/day — ~12% low** — while every field consulted looks healthy and consistent. **This is the workspace rule's own case, in a new place:** ask of a timestamp *what exactly must have happened for this to be set?* Here the answer is "the unit entered active", which is not "this process started". `NRestarts=0` is the tell, and it reads like reassurance. | **READY (XS) — NEW 2026-08-20** | — | R-341's command already uses `ps -o lstart= -p $MainPID` and is correct; the risk is a future reader "improving" it to a systemd property. Add the reason as a comment beside that command in the R-341 row (done), and check whether any other slope or uptime check in the repo or in `scripts/felhom-tenantsync.sh` anchors on a systemd timestamp where it means a process start. | CC |
|
||||
|
||||
| **R-339** | **The hub was SILENT when it lost sight of the off-site stores — and a 9 h 37 m outage proved it.** Both box checkers (`OffsiteBoxChecker` over the Hetzner API, `PBSDRBoxChecker` over ep0's `usage` op) held their last snapshot and returned quietly on a failed fetch. That is **correct for a fill signal** — a missing reading must never be mistaken for 0%, which is why degraded data drives no band transition — but the consequence was that a completely dead off-site endpoint and a healthy one were **indistinguishable on the operator channel**. During the 2026-08-18 ep0 incident the hub said nothing for the entire outage; the only mails came from the boxes' own backup failures, and **only because the WEEKLY offsite run happened to fall inside the window**. Two days earlier, nothing would have fired at all | **SHIPPED — hub v0.106.0, 2026-08-18.** Reachability is now a second, independent signal: consecutive failed fetch windows counted per checker, `pbsdr_box_unreachable` / `offsite_box_unreachable` (severity `warning`) past a default **3 windows (≈30–45 min)**, with paired `*_recovered` all-clears wired into `recoveredPairedDownTypes` — necessary because both recoveries are severity `info` and `severityNotifies` drops `info`. Threshold tunable via `alerting.box_unreachable_windows`. **The fill logic is untouched**: no threshold, throttle, band or escalate-once behaviour changed. Evidence: `internal/monitor/box_reachability_test.go` (Scenarios A–F) + `internal/notify/dispatcher_box_reachability_test.go` (the cross-package wiring, asserting an actual operator mail), plus three companion red-proofs each seen failing with a message naming the right cause | — | **PROVEN-LIVE still owed.** No real or constructed outage has exercised the emit path end to end, and one cannot be manufactured without making ep0 or the Hetzner API unreachable — ep0 is Tier 2 protected, so that is forbidden. The honest route is a constructed outage against a scratch hub instance with the tenantsync client pointed at a blackholed address. **Do not close this row on the unit tests** | CC |
|
||||
| **R-340** | **The new reachability check does not touch the surface that actually failed.** R-339 reports when the hub cannot READ ep0 — but the read it performs is the `usage` op, which is `proxmox-backup-manager` plus `df` over SSH, and therefore rides the **local API daemon**. The 2026-08-18 incident explicitly CLEARED that daemon: `proxmox-backup.service` was healthy throughout, and it was the **HTTPS proxy on 8007** that was wedged with a full accept queue. **So R-339's check would have returned green for all 9 h 37 m of that outage.** It closes the case where ep0 is unreachable *as a host*; it does not close the case that actually happened. This is not a defect in R-339 — it is the honest boundary of what it watches, recorded so a future reader does not mistake a green box gauge for a working off-site tier | **READY (M) — NEW 2026-08-18** | a tenantsync endpoint-script version bump (the op is added on ep0, so it needs the same version-gated rollout `ErrUsageUnsupported` already models) | Add a **health op** to `scripts/felhom-tenantsync.sh` that probes `https://127.0.0.1:8007/` on ep0 and reports the proxy's fd count and listen-queue depth, then surface it as a third signal. **Overlaps the connections spike (R-336's remaining half): both want the same observations from ep0, so whichever runs SECOND must reuse the first's evidence rather than re-measuring a protected machine twice** | CC |
|
||||
| **R-340** | **The new reachability check does not touch the surface that actually failed.** R-339 reports when the hub cannot READ ep0 — but the read it performs is the `usage` op, which is `proxmox-backup-manager` plus `df` over SSH, and therefore rides the **local API daemon**. The 2026-08-18 incident explicitly CLEARED that daemon: `proxmox-backup.service` was healthy throughout, and it was the **HTTPS proxy on 8007** that was wedged with a full accept queue. **So R-339's check would have returned green for all 9 h 37 m of that outage.** It closes the case where ep0 is unreachable *as a host*; it does not close the case that actually happened. This is not a defect in R-339 — it is the honest boundary of what it watches, recorded so a future reader does not mistake a green box gauge for a working off-site tier | **READY (M) — NEW 2026-08-18** | a tenantsync endpoint-script version bump (the op is added on ep0, so it needs the same version-gated rollout `ErrUsageUnsupported` already models) | Add a **health op** to `scripts/felhom-tenantsync.sh` that probes `https://127.0.0.1:8007/` on ep0 and reports the proxy's fd count and listen-queue depth, then surface it as a third signal. **Overlaps the connections spike (R-336's remaining half): both want the same observations from ep0, so whichever runs SECOND must reuse the first's evidence rather than re-measuring a protected machine twice** **REUSE, per this row's own instruction — the connections spike ran FIRST (2026-08-20) and already produced most of what the health op wants; do not re-measure a protected machine a third time.** Available in `audits/evidence-ep0-established-connections-2026-08-20/`: the proxy **fd count** and its type breakdown (`lsof` + `/proc/<pid>/fd`), the **listen-queue depth** (`ss -lnt` — `Recv-Q 0`, `Send-Q 1024`), the **ESTAB/CLOSE-WAIT split**, the **per-peer** connection histogram, a **31-minute persistence diff** of full 4-tuples, and a **46.18 h** slope with Poisson bounds. What the health op would still add beyond these: a loopback `GET https://127.0.0.1:8007/` probe — the observation that distinguished "process problem" from "network problem" on 2026-08-18 and the one thing this spike did NOT take, because it is the surface R-339 cannot see. **And this spike sharpens what the op should report:** a rising **ESTAB** count is the live signal (CLOSE-WAIT was **0**, not merely flat), and per **R-344** the fd ceiling that matters may be the **agent's**, not only ep0's. | CC |
|
||||
|
||||
<!-- DUE-CHECKS-BEGIN — machine-readable. Parsed by scripts/due_checks_gate.py.
|
||||
One row per dated check. The R-number must have a row above. Dates are UTC.
|
||||
@@ -655,6 +658,5 @@ class (an image `VOLUME` at an unmounted path) is still live — `immich-server`
|
||||
the R-row. Duplicating them here would create the second source this design avoids. -->
|
||||
| item | due (UTC) | what to measure |
|
||||
|---|---|---|
|
||||
| R-341 | 2026-08-19 | ep0 proxy fd count + ESTAB/CLOSE-WAIT split; PID must still be 551655 |
|
||||
| R-341 | 2026-08-25 | same, +7 d |
|
||||
| R-341 | 2026-08-25 | same, +7 d. Anchor the elapsed time on `ps -o lstart= -p $MainPID`, NOT `ActiveEnterTimestamp` (R-346). A Phase-C perturbation sits inside this interval and is quantified in the R-341 row |
|
||||
<!-- DUE-CHECKS-END -->
|
||||
Reference in new issue
Block a user