4a3f6dee5998f119fc63fb8418bb89c8e25fc61d
12
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
  jakman85andClaude Opus 5.5
|
7c4eb439b2 |
feat(monitoring): alert when the keycloak passkey reconciler's Job fails
KeycloakPasskeyReconcilerFailing (warning, for 15m) on
kube_job_status_failed{namespace="keycloak", job_name=~"keycloak-passkey-reconciler-.*"},
following the sibling client-secret reconciler's rule. promtool unit tests pin it in
both directions: quiet when succeeding, firing after 15m on a failed Job, quiet on the
SIBLING's failure, and quiet on a single transient failure. Runbook added and projected
by the runbooks ConfigMap generator.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
||
|
jakman85andClaude Opus 5
|
8427b28b0e |
fix(obs): the deploy notification was inhibited before it could send
FluxRolloutInProgress fired correctly on the first real deploy and its mail
was suppressed before Alertmanager dispatched it. One label was wrong.
kube-prometheus-stack ships general.rules -> InfoInhibitor:
group by (namespace) (ALERTS{severity="info"} == 1)
unless on (namespace) group by (namespace)
(ALERTS{alertname!="InfoInhibitor", alertstate="firing",
severity=~"warning|critical"} == 1)
which fires for any namespace holding a firing info alert AND NOTHING ELSE
wrong in it, and the chart's third inhibit_rule (source alertname=InfoInhibitor
-> target severity="info", equal [namespace]) then drops exactly that alert.
For a deploy notification that is the worst possible interaction, because the
suppressed case is the ONLY case that matters: on a healthy rollout of a healthy
cluster flux-system holds no warning or critical alert, so the notification is
inhibited precisely when it is wanted, and would arrive only when something else
in flux-system was already broken. Measured on prod: InfoInhibitor went active
2026-09-30T04:58:48Z, the same moment the alert began firing for 12
Kustomizations, and amtool showed no FluxRolloutInProgress in Alertmanager
either active or --inhibited, with zero send failures.
Upstream's intent is explicit and fair: severity "info" means "only worth
showing next to a real problem". A deploy notice is not that, so it must not
claim that severity. FluxRolloutInProgress now carries `notification: deploy`
and no severity at all, and the route matches on that label — no inhibit_rule
in the chart targets anything but severity values, and omitting severity also
stops InfoInhibitor firing for flux-system rather than merely working around it.
FluxRolloutStuck keeps severity: warning and is unaffected.
Also verified on prod in the process, all previously UNVERIFIED: the 33 metric
series exist exactly as designed, the `revision` label IS populated (so the
lastAppliedRevision/artifact.revision paths are right), the podAnnotations roll
forced a new KSM pod, and the rule both fired for 12 lagging Kustomizations and
then CLEARED on convergence — the completion signal's whole mechanism, on real
data. Routing proven by `amtool config routes test notification=deploy` ->
email-deploy.
Still unverified: no mail observed in the inbox. The 226->227 email counter
increment during the window cannot be attributed (no receiver label, other
alerts active), and I am not claiming it either way. See the handoff entry.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
|
||
|
jakman85andClaude Opus 5
|
55b232e2ce |
feat(obs): notify on Flux deploy start and completion
The operator got no mail when Flux began applying a git change and none when it finished. Measured cause: notification-controller v1.4.0 is running and its CRDs serve v1beta3, but there are ZERO Provider and ZERO Alert objects in the cluster — nothing in git ever created any. The obvious fix does not work. Neither Flux event can be filtered into a deploy signal: ReconciliationSucceeded is a heartbeat (101 events in one hour across 16 Kustomizations, message byte-identical whether the apply changed anything), and Progressing is worse (five Kustomizations re-apply the same object on every tick — event counts 2505/2504/2501/2495/2490). gotk_* carries no revision label at all (gotk_reconcile_condition 120 series, gotk_suspend_status 40, none with one). The applied revision is the only signal that moves when and only when git moves, and it lives in object status alone. So export it via the kube-state-metrics custom-resource-state path that already serves the Velero Schedule metrics — kube_flux_kustomization_info/_suspended and kube_flux_gitrepository_info — and ask the question in PromQL: FluxRolloutInProgress (severity: info, no `for:`) fires the moment a non-suspended Kustomization's applied revision is not the revision the GitRepository fetched from main, and CLEARS when all have caught up. Fired mail = a deploy started; RESOLVED mail = it completed. The completion half is send_resolved, not a second rule. FluxRolloutStuck (severity: warning, for: 30m) is the same expression as the escalation for a rollout that started and never converged. Two guards, each with a test whose only job is to fail if the clause is removed: `and on () (count(gitrepository_info) > 0)` — without it a missing exporter makes `unless on (revision) <empty>` remove nothing and all 16 Kustomizations alert forever, the permanently-firing mode this repo shipped once already — and the suspended-Kustomization exclusion, without which a suspended tier holds the rollout open and the "applied" mail never arrives. A dedicated Alertmanager severity=info route + email-deploy receiver keeps deploy notices off the [WARNING] catch-all. podAnnotations carries a roll-counter because the chart sets no checksum over the custom-resource-state ConfigMap, so a config-only edit would leave KSM on the old collector set with the new alerts silently inert. Verified: promtool SUCCESS on 6 scenarios (both guards, plus the convergence test that proves the alert empties itself); lint/l0/render-golden/l1/decisions-log/ spec-immutability green in the pinned toolbox. NOT verified: no mail has been sent — the metrics do not exist until this merges. Full gap list, and the deliberately-cut notification-controller half, in the handoff entry. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> |
||
|
jakman85andClaude Opus 5
|
a3f4ddb971 |
merge: feat/egress-gateway-staleness-alert — finished so it can actually pass the gate
Brings the two alerts for the silent egress-gateway failure: VPNEgressSteeringStale
(critical) and VPNEgressUplinkLabelMissing (warning), plus the kube-state-metrics
allowlist entry that makes them non-inert.
WHY THE BRANCH SAT IN THE BACKLOG: it could not pass the merge gate. Three
whole-repo checks failed on it, and each one was a real gap rather than a nuisance:
* alert-runbooks-resolvable — both alerts link to /runbooks/vpn-egress-stale,
which did not exist. An alert firing now would hand the operator a 404.
* alert-rules-unit-tested — alerts.yaml is already a covered file, so new rules
in it must be tested or a "tested" file quietly grows untested rules.
* alert-metrics-have-scrape-targets — see the guard bug below.
WHAT THE ALERTS CATCH. egress-class=vpn pods leave through one gateway node's tun0;
Cilium resolves it to an ifindex and caches it. An OpenVPN reconnect destroys and
re-creates tun0 with a NEW ifindex, and the node label
homelab.keithjackson.dev/egress-uplink=<addr>-<ifindex> is what forces Cilium to
re-resolve. If the label does not follow the tunnel, the cache points at a destroyed
interface and every vpn-class pod is silently dropped — while tun0 stays up and
VPNTunnelDown says nothing. That ran ~80 minutes undetected on the dev gateway.
Verified this is live on prod, not inert: k3s-prod-3 carries egress-gateway=true and
egress-uplink=10.150.192.86-139, and vpn-egress-steering has been in place 14 days.
The label was exported on 0 series before this change, so the allowlist entry is
what switches the alerts on.
runbooks/vpn-egress-stale.md — written and wired into the configMapGenerator, with
the four required operator sections. It leads with proving egress from a pod the
policy selects, because the symptom is "outbound traffic silently disappears", and
it says plainly not to "fix" this by deleting the policy: that does not restore
egress, it sends vpn-class traffic out of the real ISP connection, which is the leak
the class exists to prevent.
platform/monitoring/tests/vpn-egress.test.yaml — four promtool fixtures. 1 and 2 are
the SAME tunnel reconnect with and without the label following it, which isolates the
mechanism the rule actually keys on: a label VALUE change is not a changing sample,
it retires one series and starts another. 3 is the inert case (gateway node, no uplink
label) where the critical rule CANNOT fire, which is why the warning exists.
Two fixes the branch needed that were defects in the checks themselves:
1. alert-metrics-have-scrape-targets asserted the KSM allowlist matched
`nodes=[...egress-gateway]` with a literal closing bracket — i.e. that the list
contained EXACTLY that one label. Adding a second label correctly (egress-uplink,
which these alerts need) therefore FAILED the check while the property it guards
was perfectly satisfied. Loosened to `[,\]]`: in the list, not the whole list. A
guard that breaks when you correctly extend the thing it guards teaches people to
edit the guard, which is how it stops guarding.
2. UNTESTED_ALERTS listed VPNTunnelDown, and the check flagged it as stale because
"asserted on" is a bare word match over the test file — my header COMMENT
mentioning VPNTunnelDown was enough to count as coverage. Rather than delete the
register entry and claim coverage that did not exist, the fixtures now assert
VPNTunnelDown QUIET in all four cases. That required making them faithful: they
had no node_network_receive_bytes_total{device="tun0"} at all, so VPNTunnelDown
would have been firing throughout unnoticed, because promtool only checks the
alertnames you name. A healthy tun0 passes traffic, and that silence IS the claim
these rules exist for.
NOTE FOR LATER, not fixed here: that word-match means any alert named in prose counts
as tested. Tightening it to `alertname:` positions could reveal other prose-only
"coverage", and the register is shrink-only, so the cascade wants doing deliberately
rather than at 01:00 unattended.
Conflicts resolved: kept HEAD's Velero custom-resource-state block and took only the
branch's one-line allowlist change (14 lines down to 1); inserted the branch's
2026-09-12 DECISIONS-LOG entry in date order rather than appending it out of sequence.
Gated: offline l0,l1 --include-scope whole-repo — 24 cases, 0 failures, 0 skips,
0 unwired. The new unit tests pass under the pinned promtool 3.14.0.
Co-Authored-By: Claude Opus 5 <[email protected]>
|
||
|
jakman85andClaude Opus 5
|
79c703bc79 |
fix(conformance): whole-repo checks audited nested agent worktrees as source
Six discovery sites walked ROOT without pruning dot-directories, so they descended into .claude/worktrees/agent-*/ — up to 18 full copies of the repo, created for isolated agent runs and gitignored (.gitignore:43). The visible half was false FAILs: four ServiceMonitors reported as selecting no Service, and nine PrometheusRule manifests reported as "under test but not in the repo", all of them worktree paths. The whole-repo L0 gate was red for reasons that had nothing to do with authored surface. The dangerous half is quieter. A check that asks "does anything in the repo declare X" can be satisfied by a worktree copy instead of real source, so the same bug that invents failures can also manufacture a pass. That is the reason to fix all six sites rather than only the ones that happened to be failing. egress-paths-selected-by-rules already had this fix and documents it (it once saw 5 CiliumEgressGatewayPolicy objects instead of the 1 authored); this applies the same dot-directory predicate everywhere else, which skips dot-DIRECTORIES while still matching dot-FILES such as .sops.yaml. platform/monitoring/tests/run.sh is the other side of the same failure: its `--list` reported each manifest once per worktree, and the L0 check compares that listing against the tree. Both find calls now prune .claude alongside .git. `--list` returns 21 real pairs where it previously returned them ~19x over. Verified in the pinned toolbox: L0 --include-scope whole-repo is 21 cases, 0 failures, 0 skips, 0 unwired (was 4 failures, then 1). L1 3/3. The 18 stale worktrees are all on already-merged branches and remain on disk; cleaning them up is a separate change and is no longer load-bearing for the gate. Co-Authored-By: Claude Opus 5 <[email protected]> |
||
|
jakman85andClaude Opus 5
|
79d9ab2c8c |
feat(monitoring): per-export NFS latency alerts — server-slow vs we-are-backed-up
Closes the gap `feat/io-pressure-diagnostic-context` deliberately left open. That
change enabled node-exporter's `--collector.mountstats` and recorded its own honest
limit: "an NFS export that is merely SLOW cannot be distinguished today", because
node_mountstats_* did not exist on prod until the HelmRelease rolled and the L0 guard
alert-metrics-have-scrape-targets rightly refuses a rule on an unscraped metric.
The rollout has landed. Verified on prod 2026-09-15: 1,842 node_mountstats_* series
(~0.4% of a 451,650-series head), six mounts, 426 per-operation series in each of the
five operation families and 6 in each transport family. Every metric referenced below
returns series under `promtool query instant`.
FIVE RULES in platform-health.node-io-cause, each naming its cause in `summary`:
NFSExportServerSlow response_time > 50 ms/op, our op rate > 20/s,
queue_time < 5 ms/op -> the server is slow BECAUSE OF US
NFSExportServerSlowNotOurLoad same, but our op rate 1-20/s -> the server is slow and
the load is NOT OURS. This is the operator's Proxmox
hypothesis made testable.
NFSClientQueueTimeHigh queue_time > 5 ms/op -> WE are saturating the transport
NFSTransportBacklogDeep backlog depth > 8 per send -> even the grown slot table is full
NFSExportMajorTimeouts any major timeout in 15m -> the export is HANGING (critical)
THE DISCRIMINATOR. mountstats splits a slow NFS op into the half the server owns
(response_time) and the half we own (queue_time). High response with low queue means the
SERVER; high queue means US. The two server rules then partition that state by our own op
rate at 20/s, so exactly one of them can fire and it names which reading applies. The
`unless queue_time > 5ms` clause is what licenses reading response_time as a statement
about 192.168.2.102 rather than about us, and it is pinned by fixture 5 (a client queueing
30 ms/op while the server appears to take 60 ms/op: both server rules must stay quiet).
MEASURED BASELINE, prod 2026-09-15, re-measured for this change:
server response time 4.40-4.81 ms/op (30m window), uniform across all six mounts
client queue time 0.013-0.067 ms/op; snapshot maximum 0.333 ms/op
our own op rate 0.019-0.163 ops/s — STATFS keepalive and nothing else
major timeouts 0 on every mount
backlog depth 0 on every mount
slot-table high-water 586 (prod-1), 162 (prod-3), 2 (prod-2)
THREE LIMITS ON THAT BASELINE, recorded in the rules and the runbook rather than papered
over:
(a) the history is MINUTES, not days — the series begin when the DaemonSet rolled
(count_over_time(...[24h]) returned 18, and time() - process_start_time_seconds read
558-561 s on all three nodes), so no rule uses a range longer than 15m;
(b) it is an IDLE baseline — STATFS keepalive only, no READ/WRITE traffic, so the 50 ms
threshold is argued from the link and the array (one outstanding 1 MB read at 50 ms is
20 MB/s) and must be re-cut after the first real media event. The runbook carries the
query;
(c) prod-1's /mnt/Storage/plex and /mnt/Storage/plex/downloads report byte-identical
per-operation counters (0.12443893333333335 on both, to every digit), as do prod-3's two
nextcloud mounts — mounts sharing one NFS superblock share the kernel's RPC statistics.
One slow export on prod-1 therefore raises two alerts. Documented, not aggregated away.
PROOF, not assertion. platform/monitoring/tests/nfs-export-latency.test.yaml: seven fixtures
(measured healthy; healthy under 60 ops/s of real load; server-slow-ours; server-slow-not-ours;
client-backed-up; backlog-deep; export-failing), and every fixture asserts all five alertnames
in the fires or the quiet direction. The measured baseline is pinned as a band, the
shared-superblock duplication is pinned, and both hazards the sibling recorded are asserted:
the PromQL precedence trap (`(ratio) and rate(ops) > 1 > 0.050` returns a series on a healthy
60 ops/s mount; the shipped form returns none) and the op-rate guards (a mount doing half an
operation a second at 1 s/op response, 1 s/op queue and 20 deep in the backlog — 20x, 200x and
2.5x the three thresholds — must stay silent).
MUTATION-CHECKED, 14 mutations, every one of them making the test file FAIL: each of the five
thresholds moved in both directions where the fixtures allow it, both `unless` clauses deleted,
and all four rate guards loosened. The two clause deletions fail with an exp/got alert diff,
not a parse error. No threshold here is pinned only by a promql_expr_test restating it.
Live on prod, all five shipped expressions return NO series while all four of their
sub-expressions return six each — so the silence is a verdict, not missing data.
runbooks/node-io.md: the triage table grows from 10 rows to 15 (per-export server response
time, client queue time, our own demand, major timeouts, backlog depth), with real measured
HEALTHY values and an explicit "not collected" in both incident columns, because the collector
postdates both events and that can never be backfilled. A new four-way verdict table reads rows
11-13 together and maps each combination to the alert that says it.
HONEST REMAINING GAP: these rules rule US out; they cannot rule Proxmox IN. The per-client view
of an export exists only on the server. What would settle it is one read on 192.168.2.102 while
NFSExportServerSlowNotOurLoad is firing — `nfsstat -s` and /proc/fs/nfsd/clients/*/info — which
is out of scope from this repo. Recorded in the runbook.
Co-Authored-By: Claude Opus 5 <[email protected]>
|
||
|
jakman85andClaude Opus 5
|
b9f3143f42 |
feat(monitoring): name the CAUSE of node I/O pressure, not just the stall
NodeIOPressureCritical has now fired twice on k3s-prod-1 for opposite causes and
said the same thing both times. On 2026-09-14 02:40-02:45Z it fired at PSI(full)
20.98% — twice the threshold, four times the 2026-09-07 level — while every local
device was idle (0.09-0.27 utilisation) and fast (1.4-4.8 ms per write), NFS
retransmissions were zero and no Longhorn volume was degraded. Establishing that
took eight ad-hoc PromQL queries after the event had already recovered.
Adds the group platform-health.node-io-cause: eight rules, one per cause class,
each naming its cause in its own summary.
NodeDiskDeviceSaturated device saturation >0.85 busy, 10m
NodeDiskWriteLatencyHigh device latency >20ms/write, 10m
NodeDiskWriteLatencyCritical device latency >50ms/write, 5m
NodeDiskReadLatencyHigh device latency >250ms/read, 15m
NodeNFSRetransmissionRateHigh NFS transport >1% retransmitted, 10m
NodeIOPressureNotLocalDisk not local disk PSI>10% with all of the above absent
NodeEtcdRequestsSlow etcd >0.5% of etcd requests over 1s
ContainerIOStallHigh which workload per-container PSI >15%
Every threshold is measured on prod over 24-48h, not guessed, and each rule
records its headroom. The PSI thresholds themselves are deliberately unchanged:
the 21% stall was real, it fired twice in 24h for two real events on the one node
that has a problem, and raising the bar would also have raised it above the 11.36%
at which prod-1 was already losing etcd on 2026-09-07.
Context reaches the operator at page time three ways: the summary states the cause
class; every rule carries a diagnostic_query annotation with the exact companion
PromQL; and node-pressure-dashboard.yaml gains a WHY row (seven panels) plotting
saturation, write/read service time, NFS retransmissions, our own NFS op rate,
the etcd slow-request fraction and per-container PSI side by side. runbooks/node-io.md
gains a 10-row triage table with the real healthy / 2026-09-14 / 2026-09-07 columns,
which turns those eight queries into a two-minute read.
etcd and NFS metric availability, verified on prod rather than assumed:
* etcd SERVER disk metrics do NOT exist. etcd_disk_wal_fsync_duration_seconds and
etcd_disk_backend_commit_duration_seconds are absent (kubeEtcd disabled for k3s);
a query for etcd_disk_*/etcd_server_*/etcd_mvcc_* returns nothing at all.
NodeEtcdRequestsSlow uses the apiserver's client-side view of its own local
member instead, and says in its own description that it is a proxy.
* node_nfs_* DOES exist — the nfs collector is on by default, host-wide via
/proc/net/rpc/nfs, confirmed on all three nodes. It gives op counts and
retransmissions and NO latency, so it can see a dropping export but not a slow
one. The mountstats collector is enabled here to close that gap; NO rule
references node_mountstats_* yet, because the series do not exist until the
HelmRelease rolls out and the collector's visibility through the node-exporter
mount namespace must be verified on prod first. The runbook records the check
and says to revert if it comes back empty.
Proof, per the repo's two shipped alerting failure modes:
* platform/monitoring/tests/node-io-cause.test.yaml — 7 fixtures, all from
measured profiles: healthy (all 8 quiet), 2026-09-14 (NotLocalDisk fires and
names sonarr), 2026-09-07 re-queried at offset 164h (write-latency and etcd
fire, NotLocalDisk correctly silent), rebuild storm, NFS transport, read
latency, and the `unless` semantics.
* 10 mutations of the shipped rules, all killed. The mutation run found a real
gap: the etcd threshold could be changed without failing any test, because it
was only pinned by a promql_expr_test that restated the threshold in the test
file. An alert_rule_test now pins it.
* All 11 referenced metrics verified to return series on prod with
`promtool query instant`, and all 8 expressions verified quiet on today's state.
Also extends the alert-metrics-have-scrape-targets owner table for container_*
(the kubelet ServiceMonitor's cAdvisor path, which this entry already claimed in
its title) and for etcd_request_duration_seconds_* by EXACT NAME, deliberately not
by an etcd_ prefix, so a future rule on the unscraped etcd_disk_* names still fails
the gate.
Co-Authored-By: Claude Opus 5 <[email protected]>
|
||
|
jakman85
|
52398d4426 |
merge: Longhorn alerting measured the wrong things, three ways
(1) LonghornVolumeCapacityHigh alarmed on ALLOCATION while calling itself a capacity alarm. Prometheus's TSDB volume read 0.973 while the kubelet saw the same filesystem at 0.208, because 37.73 GiB of it is a Longhorn SYSTEM snapshot — the kind taken for the full replica rebuild that prod-1's drive replacement forced — growing +152 MB/h while the filesystem stayed flat. The rule was therefore unclearable by design, i.e. permanently firing. Deleted as redundant, shown rather than asserted: 23 of 24 volumes fall under the kubelet-based filling rules, and the 24th is volumeMode Block, which the deleted rule was not covering either. (2) LonghornVolumeDegraded and LonghornVolumeFaulted could not fire even with the scrape working. longhorn_volume_robustness on 1.12.1 is ONE-HOT over a state label with values 0/1, not the 0-3 enum the rules and their comment assumed, so a faulted volume would have alerted nobody. This survived the scrape fix landing hours earlier. (3) LonghornNodeStorageHigh was wrong twice: its 0.85 threshold sat above the 75% at which storage-minimal-available-percentage=25 already removes a disk from the scheduling pool, and longhorn_node_storage_usage_bytes SUMS a node's disks, so one full disk hid behind a half-empty sibling. Replaced with per-disk LonghornDiskAllocationHigh (>0.70) and LonghornDiskUnschedulable, the latter reading Longhorn's own verdict instead of re-deriving a formula from three settings no metric exports. Baseline green with 11/11 mutations red. retentionSize 45GiB of 50Gi and discard/TRIM are both REPORTED with recommendations, not changed: TRIM would not have helped here, since remove-snapshots-during-filesystem-trim is false and the 37.7 GiB is a snapshot. # Conflicts: # DECISIONS-LOG.md |
||
|
jakman85andClaude Opus 5
|
4687521d65 |
fix(monitoring): delete the Longhorn rule that measured allocation and called it capacity
LonghornVolumeCapacityHigh went pending within minutes of the Longhorn scrape
landing today. Re-measured on prod: it read 0.9731 on the Prometheus TSDB volume
(actual_size 52243681280 / capacity 53687091200) while the kubelet reported the
SAME filesystem 20.8% used. actual_size is head + pinned snapshots, and one
user_created=false system snapshot from prod-1's replica rebuild holds 40510484480
bytes (37.7 GiB) of it; Longhorn returns no freed blocks without discard/TRIM, and
retentionSize 45GiB on a 50Gi volume targets 90% by design. The ratio is pinned
near 1.0 and cannot fall back under 0.85 — a permanently-firing alert, the same
shape as CNPGNoRecoverabilityPoint.
DELETED, not re-thresholded. Raising the threshold would tune away a true
statement about allocation. Redundancy proven, not asserted: of 24 Longhorn
volumes on prod, 23 appear in kubelet_volume_stats_inodes > 0 and so fall inside
MonitoringVolumeFillingUp (namespace="monitoring") or PersistentVolumeFillingUp /
PersistentVolumeAlmostFull. The 24th is unifi/unifi-rootdisk, volumeMode Block —
no filesystem for anyone to measure, and the deleted rule was not covering it
either.
AND A SECOND INERTNESS the scrape did not fix. longhorn_volume_robustness on
longhorn 1.12.1 is one-hot over a `state` label with values 0/1, not the 0-3 enum
the rules assumed:
count(longhorn_volume_robustness) => 96 (24 volumes x 4 states)
count by (state) (longhorn_volume_robustness == 1) => {state="healthy"} 24
count(longhorn_volume_robustness == 2) => (empty)
count(longhorn_volume_robustness == 3) => (empty)
So a FAULTED volume would still have alerted nobody. Re-keyed to the label; the
"44 series / 72 unknown(0, detached)" note in platform/longhorn/helmrelease.yaml
was the same misreading and is corrected in place.
The allocation question moved to where allocation is a budget. LonghornNodeStorageHigh
fired at 0.85, but storage-minimal-available-percentage is 25 on this cluster, so a
disk leaves the scheduling pool at 75% used — the old threshold could only fire
after scheduling had already been refused. Replaced by:
LonghornDiskAllocationHigh (longhorn_disk_usage_bytes / longhorn_disk_capacity_bytes) > 0.70, 30m
LonghornDiskUnschedulable longhorn_disk_status{condition="schedulable"} == 0, 15m
Per DISK, not per node: longhorn_node_storage_usage_bytes sums a node's disks, so
one full disk could hide behind a half-empty sibling. Measured prod usage
0.2254 / 0.2306 / 0.2369 on prod-1/2/3.
PROVEN NEITHER PERMANENTLY FIRING NOR INERT. platform/monitoring/tests/longhorn.test.yaml
adds 6 cases pinning every rule quiet on today's measured prod state and firing on
a constructed bad state, with every `for:` asserted, plus regression assertions that
the old `== 2`/`== 3` forms match nothing and that per-disk 0.72 is visible where the
node-summed 0.475 is not. Mutation-checked 11 ways; each turns the suite RED. Live on
prod: all four new expressions return empty, while the sum of their selector
cardinalities is 54. The L0 UNTESTED_ALERTS register shrinks by seven (7/56 -> 14/58
alert rules unit-tested), which also brings MonitoringVolumeFillingUp,
PersistentVolumeFillingUp and PersistentVolumeAlmostFull under test since the
redundancy claim rests on them.
REPORTED, NOT CHANGED (DECISIONS-LOG, owner decisions): retentionSize 45GiB on a
50Gi volume is too tight — recommend 36GiB, with reasoning; and discard/TRIM on the
longhorn StorageClass is feasible but not advisable now — remove-snapshots-during-
filesystem-trim reads false on prod, so trim would not have reclaimed the 37.7 GiB
anyway, and continuous discard on nodes where the etcd WAL shares the spindle points
at this cluster's known failure mode.
Co-Authored-By: Claude Opus 5 <[email protected]>
|
||
|
jakman85andClaude Opus 5
|
cf7d97e546 |
fix(obs): twelve app SLI rules read a signal that exists; kill the dead ServiceMonitors
Twelve `<app>-down` SLI rules were STRUCTURALLY INERT. Each read `up{job="<app>"} == 0`,
and the ServiceMonitor rendered beside it named `endpoints[].port: metrics` — a port NO
rendered app Service defined. Every monitor discovered ZERO targets, so no `up{job="<app>"}`
series existed and the rule could not fire. Verified on prod before this change:
`count by (job) (up)` listed 24 jobs and not one of the twelve appeared; `up{job="plex"}`
and `up{job="gitea"}` both returned nothing. `PrometheusTargetsMissing` could not see it
either — a target that was never discovered has no `up` series to be 0.
THE PORT NAME WAS NEVER THE MISSING PIECE. Ground truth per app: Plex, Pi-hole, Jackett,
qBittorrent, Sonarr, Radarr, Seerr, Semaphore and Nextcloud export NO Prometheus metrics
without a separate exporter sidecar; Gitea can but only with app.ini metrics enabled;
Keycloak does, on the operator's management port, already scraped by
platform/keycloak/servicemonitor.yaml; vm-demo is a KubeVirt VM with no Service at all, so
its monitor selected nothing whatsoever. `metrics: true` was asserting a fact none of them
had. There was also a THIRD layer of inertness nobody had hit yet: every app namespace ships
default-deny with `ingress: []` and the renderer emitted no monitoring allow, so even a
correct port would have been dropped by policy — while renderer-contract.md claimed the
baseline carried such an allow.
FIXED IN THE GENERATOR, not in twelve rendered files:
* `slis: [up]` now renders a `<app>-down` alert against kube-state-metrics readiness of the
service's own workload object —
`kube_deployment_status_replicas_available{namespace,deployment} == 0 or absent(...)`,
`kube_statefulset_status_replicas_ready` for a StatefulSet, and
`absent(kubevirt_vmi_info{namespace,name,phase="running"})` for a VM. The same idiom
KeycloakDown has used since it was fixed. kube-state-metrics IS scraped, so this needs no
exporter and no new target. The `or absent(...)` half is load-bearing: the signal
DISAPPEARING now fires the alert instead of silencing it.
* schema 1.6.0 adds `observability.metricsEndpoint` — WHERE metrics are served. A
ServiceMonitor is rendered ONLY from it, and it renders all three halves of a working
scrape together: the named Service port, the monitor endpoint, and the
monitoring-namespace CiliumNetworkPolicy allow. No catalog entry declares one, so twelve
dead ServiceMonitors are gone. `metrics: true` keeps meaning logs-to-Loki plus
kube-state-metrics liveness.
* schema 1.6.0 also adds `observability.livenessWorkload`, and catalog/keycloak.yaml uses
it. Keycloak's server is a StatefulSet run by the operator; the default Deployment shape
would have made its `absent()` half fire FOREVER — the other failure mode this repo has
shipped.
* renderer selftest gains cases 9/9a/9b/9c: no metricsEndpoint => no monitor; the up SLI
must not read `up{job=...}`; endpoint port must exist on the Service; the metrics port
must NOT reach the Gateway allow; an unknown livenessWorkload.kind is REFUSED.
PROOF, both directions, for all twelve (apps/wave-{0,1,2}/tests/app-slis.test.yaml, run by
platform/monitoring/tests/run.sh over the RENDERED manifests):
* quiet on the healthy state; firing on no-ready-replica; firing when the series is ABSENT;
the VM asserted both gone and in a non-running phase; per-app isolation asserted, and
keycloak asserted un-maskable by its neighbours (keycloak-operator Deployment,
keycloak-db StatefulSet ready in the same namespace).
* mutation-checked: six mutations, each turning its suite red — reverting plex to
`up{job=...} == 0`; dropping seerr's `absent()` half; pointing keycloak at a Deployment;
capitalising the VMI phase to `Running`; widening keycloak's matcher to the namespace;
inverting plex's `== 0` to `!= 0`. Control run green.
* live on prod with `promtool query instant`: all ten deployed app Deployments export
`kube_deployment_status_replicas_available`,
`kube_statefulset_status_replicas_ready{namespace="keycloak",statefulset="keycloak"}` = 2,
`kubevirt_vmi_info` present with a lowercase `phase="running"`; and evaluating the twelve
SHIPPED expressions returns no series for eleven (quiet, not permanently firing). The
twelfth is vm-demo, which is not in apps/wave-0/kustomization.yaml and so is not deployed
— its rule is not applied either.
DELETION WAS NOT THE ANSWER, and the evidence is recorded rather than assumed:
kube-prometheus-stack's `kubernetes-apps` group IS live on prod (17 rules, KubePodNotReady /
KubeDeploymentReplicasMismatch among them), so app pods are not unmonitored — but that is a
generic warning keyed to a namespace, not the per-service CRITICAL SLI R-OBS-01 declares.
platform/monitoring/alerts.yaml carries no generic app-readiness rule at all.
GUARD SO IT CANNOT RECUR SILENTLY: new whole-repo L0 invariant
conformance/tests/l0-static/servicemonitor-ports-resolve. It walks the MONITORS, not the
alerts, so it needs no `up{job=...}` rule to reach one, and resolves each endpoint by SELECTOR
across the whole tree rather than by same-file adjacency. Fixing the alerts DELETED the only
thing that made the broken monitors visible to the sibling check, which is exactly why this
axis needed its own home. It declares `conformance.home.k3s/scope: whole-repo` and adds no
ci/l0.sh stage. Proven to FAIL on the pre-fix tree: 15 failures — eleven port disagreements,
vm-demo's monitor matching no Service at all, and the three generator/contract assertions —
and to pass after. Its resolver is additionally proven on in-check fixtures (agreeing pair
accepted, disagreeing pair rejected, target-less endpoint rejected), because after the fix
there are zero in-repo monitors left for it to judge and a guard whose only evidence is
"nothing failed" cannot be told apart from one that stopped working. It also asserts
r_servicemonitor does not HARDCODE a port name — the mechanism by which one edit gave twelve
services the same wrong one.
Also fixed, found on the way:
* platform/monitoring/tests/run.sh discovered the SUITE convention only inside
platform/monitoring/tests/ while documenting it as general — a suite added anywhere else
was never run and still looked like coverage. Now discovered tree-wide; a misplaced
*.test.yaml fails loudly instead of vanishing.
* KNOWN_UNWIRED_JOBS in alert-metrics-have-scrape-targets is now EMPTY, and the twelve
apps/* entries in alert-rules-unit-tested's UNTESTED_FILES are gone — both because the
gaps were CLOSED, not waived. Both registers are shrink-only and a stale entry fails.
* renderer-contract.md claimed the default-deny baseline carried a standing monitoring
scrape allow. It never did; the claim is corrected and the allow is now real, per declared
metrics endpoint and scoped to that one port.
Co-Authored-By: Claude Opus 5 <[email protected]>
|
||
|
jakman85andClaude Opus 5
|
78d0f1b0f6 |
feat(ci): promtool in the pinned toolchain, one alert-rule harness, and the gaps declared
The CNPG alert unit tests (324 lines, mutation-checked) were run by NOTHING. promtool was
not in the baked toolchain, no gate stage evaluated any *.promtool-test.yaml, and the only
runner `curl | tar -xz`'d a promtool release from GitHub at TEST-RUN time if one was
missing. Hours later a second, independent harness landed for the Velero rules
(platform/monitoring/tests/) with its own fallback — a `podman run` of
quay.io/prometheus/prometheus. So the repo had two harnesses, two acquisition paths, and
coverage that looked present and was enforced nowhere: the exact failure mode that had just
produced a permanently-firing CNPG alert and a structurally inert Velero alert.
PROMTOOL IS NOW PINNED LIKE EVERY OTHER TOOL. PROMTOOL_VERSION=3.14.0 in
infra/devcontainer/versions.env and ci/versions.env (the two stay in lockstep, per the
repo's rule), baked by the Containerfile from the prometheus release tarball (promtool ships
inside it), recorded in /etc/dev-toolchain-versions, passed by dev.sh, and installed for the
l0 stage by .gitea/workflows/scripts/install-tools.sh so the CI runner is not permanently
SKIPPED. The pin is not free-floating: promtool IS the rule engine, so it tracks the
Prometheus prod actually runs — verified against the live cluster, not assumed:
kubectl --context k3s-prod get prometheus -n monitoring -o jsonpath='{.items[*].spec.version}'
-> v3.14.0-distroless (image quay.io/prometheus/prometheus:v3.14.0-distroless)
ONE HARNESS. platform/cnpg/run-alert-tests.sh is folded into
platform/monitoring/tests/run.sh, which now covers every PrometheusRule in the repo that has
tests, via two conventions it discovers automatically: `<manifest>.promtool-test.yaml`
beside its manifest (cnpg), and `<component>/tests/*.test.yaml` over that component's
manifests (monitoring). It writes the flattened rules under whatever name each test file's
own `rule_files:` declares, so no test file needs to know the harness exists, and it
`promtool check rules` first — a rules file that does not parse would make every unit test
pass vacuously by finding no rules at all. The cnpg script's reasoning is preserved: why a
CR must be flattened (`.spec.groups`, never re-typed), and why generated files go to a temp
dir. Its `${BASE_DOMAIN}` substitution is dropped in favour of extract_rules.py's verbatim
placeholders, with the trade-off recorded there — the hazard it guarded against is addressed
by the temp dir, and verbatim leaves one transformation fewer between the shipped rule and
the rule under test.
NO ENGINE IS ACQUIRED AT RUN TIME any more — not the GitHub release, not the container
image. Not because a pinned download is unusual here (the Containerfile downloads every tool
it bakes) but because those two decided the version, the network and the machine SILENTLY,
at test time, on whatever host ran the tests; and each quietly turned "the toolchain has no
promtool" into a green run, which is how the toolchain stayed without one. Missing promtool
is exit 3 (UNWIRED, never green); a PROMTOOL= that does not resolve is exit 2.
ONE WIRING: conformance/tests/l0-static/alert-rules-unit-tested, NOT a gate stage. The
`prom-rules` stage (ci/prom-rules.sh + its Makefile target + ci/gate.sh entry) is deleted:
it is the bespoke-per-check pattern
|
||
|
jakman85andClaude Opus 5
|
7057bbb711 |
fix(monitoring): alert when a Velero schedule has NEVER backed up, or is paused
VeleroNoRecentBackup could not fire for the failure mode that actually happened.
`(time() - velero_backup_last_successful_timestamp{schedule!=""}) > 24*3600` is the
right SHAPE — per-schedule, un-aggregated — but Velero only creates that series once
a backup for the schedule SUCCEEDS, and `time() - <absent series>` yields no series.
So it detects "backed up once, then stopped" and is structurally INERT on "never
backed up". Prod ran with `velero-daily` (the schedule carrying VOLUME DATA) paused
and never-successful: no volume backups at all, nothing alerted, and it was masked
because `velero-daily-resources-only` completed cleanly every 03:00, so every other
backup signal read green while only Kubernetes objects were protected.
A PromQL-only fix is impossible, verified rather than assumed: velero v1.18.1 exports
35 velero_* metric names and every one carrying a `schedule` label is created by a
backup ATTEMPT, not by the Schedule object. A throwaway `paused: true` Schedule on
prod produced ZERO velero_* series, and nothing exposes `spec.paused`. The fact is
not in the TSDB, so the Schedule CRs are exported via kube-state-metrics'
custom-resource-state collector — the path this repo already uses for
kube_node_labels — as kube_velero_schedule_{created,paused,protects_volume_data,info}.
Proven by running KSM v2.20.0 with this exact config against prod's API: the CRD
list/watch rule the chart adds is MANDATORY, not cosmetic — without it KSM emits the
CRS metrics as an empty set, silently.
Three rules, all keyed on PROPERTIES rather than schedule names:
* VeleroScheduleNeverBackedUp (critical) — a Schedule older than 36h with no
successful backup ever. The 36h grace comes from creationTimestamp so a newly
added schedule is not paged before its first run is even due.
* VeleroSchedulePaused (warning) — the direct root-cause signal; a paused backup
schedule must never be silently indefinite.
* VeleroNoVolumeDataBackup (critical) — THE ANTI-MASKING RULE. Fires unless some
schedule with spec.template.defaultVolumesToFsBackup=true succeeded in 36h, so a
healthy resources-only schedule can never make the platform look protected.
Reading the property that DEFINES volume protection (not a name list) matches
platform/velero/volume-policies.yaml and survives renames. `or vector(0)` is
load-bearing: `count(<empty>) == 0` returns no series and would be inert —
absence of the exporter is deliberately the alarm.
Adds the repo's first Prometheus rule-unit-test harness (platform/monitoring/tests/,
wired as the `prom-rules` gate stage). It flattens the COMMITTED PrometheusRule CRs —
never a copy — and pins every rule in the group both ways: quiet on today's healthy
prod state, firing on the reconstructed bad state, with `for:` durations asserted.
Two mutation checks confirm the tests are not vacuous: dropping `or vector(0)` and
breaking the `unless on (schedule)` join each turn the suite red. This is the tool
that makes both of this repo's recurring alert bugs — permanently-firing and
structurally inert — provable offline, which neither yamllint, Kyverno, nor a healthy
cluster can do.
Proof of the fix is by construction; the live metrics cannot show it any more, because
`velero-daily` was unpaused on 2026-09-12 and has since completed 2492/2492 items.
The throwaway probe Schedule and the probe KSM deployment were deleted and the
deletion confirmed; `velero-daily` remains unpaused.
Co-Authored-By: Claude Opus 5 <[email protected]>
|