feat(platform): guard the disk, clock, backups and monitoring after this week's incidents #10

Merged
donghyeon.kang merged 16 commits from claude/w1-infra-hardening into main 2026-11-11 22:43:49 +00:00
16 Commits
Author SHA1 Message Date
DongHyeonka 0d703ff249 docs(postgres): record why the backups go to the HDD
The owner asked on 2026-11-12 that production database backups stay
off the SSD. State the evidence that the AIStor volume is the HDD
(aistor-data-local-pv -> /srv/k3s/aistor -> /dev/sdb3, WD20EZAZ,
ROTA=1), why the NTFS archive on sdb2 was not used (no POSIX
permissions, no S3 API, a daily dump gives no point-in-time restore),
and that the restore drill uses a short-lived local-path volume on the
SSD because the HDD has no dynamic StorageClass.
2026-11-12 07:41:43 +09:00
DongHyeonka c04d79481c docs(host): point the time-sync step at the k3s restart bundle
scripts/bootstrap/apply-k3s-restart-bundle.sh now installs the two
time-sync drop-ins and enables systemd-time-wait-sync together with the
kubelet drop-ins, so the owner restarts k3s once. Say so in step 2 and
keep the manual commands for installing only these two files.
2026-11-12 07:36:11 +09:00
DongHyeonka d92dc5d163 feat(observability): read node-exporter textfile metrics
The host timers from the k3s maturity work (last Gitea dump and k3s
datastore backup, live drift count, API server and kubelet certificate
expiry) write their results as *.prom files. node-exporter does not
read any textfile directory, so Prometheus never sees them.

Pass --collector.textfile.directory pointing at
/var/lib/node_exporter/textfile_collector under the read-only host
root mount the chart already has. No volume is added. The host
installer creates the directory; until then the collector only
reports an error metric.

Verification: helm template of prometheus-node-exporter 4.56.1 with
these values renders the argument next to --path.rootfs=/host/root.
2026-11-12 07:29:48 +09:00
DongHyeonka 4551a2ece6 feat(platform-postgres): run the shared database as Guaranteed QoS
The one PostgreSQL instance holds the Gitea, Keycloak and tech-log
databases. It is Burstable at priority 0 with oom_score_adj 985, so
under memory pressure the kernel kills it before almost anything else,
and an eviction treats it like Grafana (k3s maturity review
2026-11-11, 7.2 items 2 and 3).

Set priorityClassName platform-critical (services/priority-classes)
and raise the requests to the existing limits, CPU 2 and memory 4Gi,
so the pod is Guaranteed and its oom_score_adj becomes -997. The
backup README now gives the restart, not pg_hba, as the reason to add
spec.backup with kubectl patch.

Risk: applying this changes the pod spec, so CloudNativePG restarts
the single instance and the three databases are down for that time.
It belongs in the k3s restart window. Node requests grow by 1.75 CPU
and 3.5Gi; peak use over the last 6 hours was 0.025 CPU and 195MB.

Verification: kubectl diff of the rendered build root against the live
cluster shows only priorityClassName, the two requests and the
DatabaseRole defaults that were already there; no pg_hba line differs.
render-phase1.sh passes.
2026-11-12 07:29:48 +09:00
DongHyeonka d47d57aa14 fix(platform-postgres): keep the tech-log pg_hba rules the cluster uses
The live platform-postgres Cluster has two pg_hba rules for the techlog
role that the repository does not declare. They were added outside the
repository for hyeonworks.com. Applying the current render would remove
them and the production backend could no longer reach its database.

Declare the same two rules, in the same position after the Keycloak
pair, and pin them with render invariants like the Gitea and Keycloak
pairs. The techlog DatabaseRole, Database, Secret and NetworkPolicies
stay outside this build root; where they belong is an owner decision
recorded in the 2026-11-11 maturity review.

Risk: none for the live cluster. The rendered Cluster now matches the
live object. A fresh install without the tech-log tenant gets two
rules for a role that does not exist, which PostgreSQL ignores.

Verification (desktop, Helm 3.19.4, Kustomize 5.8.1):
- render-phase1.sh with the old cluster.yaml and the new invariants
  fails: "the tech-log role must authenticate only to the tech-log
  database"; with the new cluster.yaml it passes.
- kubectl diff of the new render against the live cluster shows no
  Cluster change; only the pre-existing DatabaseRole default fields.
2026-11-12 07:28:31 +09:00
DongHyeonka d7c845ca5d refactor(postgres): keep the backup alerts with the backup
The observability core renderer rejects any alert or record whose
name contains "backup" (no_unscoped_monitors_rules): backup
monitoring is outside the observability core by design. f19ce54 put
PlatformPostgresBackupMissing and PlatformPostgresWALArchiveBacklog in
core-rules.yaml, so the pinned renderer failed with "out-of-scope
monitor or rule is present".

Move both alerts unchanged into
services/platform-postgres-backup/prometheus-rule.yaml in
platform-data, which Prometheus reads because the namespace has
observability.hyeonworks.com/enabled=true and the rule has
observability.hyeonworks.com/instance=home. They now point their
runbook_url at the backup README and are applied with the
ScheduledBackup. Renaming them to dodge the filter was rejected: it
would keep backup rules inside a scope that excludes them on purpose.

Verification: test-platform-disk-alerts.sh loads both rule files and
passes; promtool checks 13 rules. render-observability-core.sh and
test-render-observability-core.sh pass, and kubectl apply
--dry-run=server accepts the backup kustomization with the rule.
scan-platform-sensitive-source.sh stops at its observability-core
renderer with RC=23 here and on an origin/main snapshot alike, so it
says nothing about this change.

Fixes: f19ce54740 ("feat(postgres): back up platform-postgres to AIStor every day")
2026-11-12 07:28:31 +09:00
DongHyeonka 2b1310972c fix(observability): pin the heartbeat route in the core renderer
render-observability-core.sh keeps its own exact copy of the
AlertmanagerConfig, and test-render-observability-core.sh feeds it the
same copy as a fixture. 619b905 added the platform-heartbeat route and
receiver to the source and to test-observability-alerting.sh but not
to these two, so the pinned renderer rejected the source with
"platform AlertmanagerConfig is not the exact global Slack contract".

Add the route and receiver to both copies and name the heartbeat in
the reject message.

Verification: render-observability-core.sh (Helm v3.19.4, kustomize
v5.8.1) prints OBSERVABILITY CORE STATIC RENDER PASS, also with
--verified-output-dir; test-render-observability-core.sh passes.

Fixes: 619b90552f ("feat(observability): alert when Prometheus itself stops working")
2026-11-12 07:28:31 +09:00
DongHyeonka bed20b456b fix(observability): stop throttling node-exporter
node-exporter averages about 2m CPU, but each scrape runs its
collectors in parallel. With a 250m limit (25ms per 100ms CFS period)
26% of its CFS periods were throttled over 30 minutes on 2026-11-11
(container_cpu_cfs_throttled_periods_total / cfs_periods_total), which
delays scrapes and can make them time out.

Raise the CPU limit to one core. The request stays 100m, so
scheduling does not change, and a runaway collector is still capped.
Removing the limit was rejected because every other workload in this
repository sets one.

Risk: none beyond one core of headroom on a 12-core node. The live
DaemonSet changes only when the observability core is applied again.

Verification: the throttling ratio and usage came from live
Prometheus queries. helm template of prometheus-node-exporter 4.56.1
with these values renders limits.cpu "1". The effect after rollout is
not measured yet.
2026-11-12 07:28:03 +09:00
DongHyeonka 619b90552f feat(observability): alert when Prometheus itself stops working
Nothing tells the owner when monitoring itself fails. On 2026-11-11
Prometheus stored samples 9 hours ahead, then rejected every new
sample as out of bounds: rules kept running on no data, and the only
sign was a false KubeAPIDown. If Prometheus, Alertmanager or the node
went down, no rule could fire at all.

Add two rules and one route. PlatformMetricsIngestionStopped fires
when no up sample from node-exporter is stored for 5 minutes;
absent() still fires while the TSDB rejects samples. PlatformHeartbeat
always fires, and a new platform-heartbeat route sends it every 5
minutes to an external dead man's switch through a webhook whose URL
comes from the Secret alertmanager-heartbeat. When the pings stop,
that service alerts, which covers Prometheus, Alertmanager, the node
and the network being down. The chart Watchdog stays disabled; the
platform rule keeps the route inside this repository.

Rejected: an in-cluster CronJob that queries Prometheus and posts to
Alertmanager. It still depends on Alertmanager and the node, and
needs a new workload in the exact observability NetworkPolicy matrix.

Risk: the AlertmanagerConfig now references alertmanager-heartbeat;
applying it before that Secret exists keeps the operator from
rendering the new Alertmanager config. The alerting README orders the
steps: create the external check, create the Secret, then apply.
The heartbeat needs an account on an external service, which only
the owner can create.

Verification: promtool tests for both rules failed against the old
rules and pass now. test-observability-alerting.sh pins the new route
and receiver in its exact contract and passes the static contract,
the global Alertmanager render with Helm v3.19.4 and the server
dry run (--server-dry-run). kubectl apply --dry-run=server accepts
the updated PrometheusRule.
2026-11-12 07:28:03 +09:00
DongHyeonka f19ce54740 feat(postgres): back up platform-postgres to AIStor every day
platform-postgres holds the Gitea, Keycloak and tech-log databases and
had no backup: no spec.backup, no ScheduledBackup, and its only copy
is a local PV on the root SSD. A lost SSD or a bad migration had no
way back.

Add services/platform-postgres-backup: a merge patch that gives the
live Cluster barman-cloud WAL archiving and base backups to
s3://platform-postgres-backups on the AIStor ObjectStore (HDD sdb3)
with 30-day retention and gzip, a daily 03:00 KST ScheduledBackup that
also runs once at creation, the object-storage ingress for the
PostgreSQL pods, the bucket policy, and a restore drill Cluster that
only reads the bucket. scripts/bootstrap/create-postgres-backup-
credentials.sh creates the bucket (100GiB quota), policy, user and
the platform-data Secret through a restricted AIStor client Pod,
proves the user can write its bucket and cannot list loki, and prints
no key. Add PlatformPostgresBackupMissing (no base backup in 26h) and
PlatformPostgresWALArchiveBacklog (more than 10 WAL files waiting for
15m; a stuck archive fills the SSD volume).

The native barmanObjectStore is deprecated since CNPG 1.26 and the
operator warns it is removed in 1.31.0. The Barman Cloud Plugin needs
cert-manager, which this cluster does not have; installing two
operators before the first backup is slower and riskier. The bucket
layout is the same, so the plugin can take over before CNPG 1.31.
Volume snapshots do not work on the static local PV.

Risk: the live Cluster has pg_hba lines for techlog that the Phase 1
source lacks, so re-applying services/platform-postgres would cut the
tech-log backend off. The README applies only spec.backup with
kubectl patch. Backups stay on the same desktop; an off-machine copy
is not part of this change. Nothing was applied to the cluster.

Verification: kubectl apply --dry-run=server accepts the
ScheduledBackup, the NetworkPolicy and the drill Cluster; kubectl
patch --dry-run=server keeps both techlog pg_hba lines and adds
spec.backup. The row-count comparison from the drill steps ran
read-only on production (keycloak 100, techlog 44, gitea 116 tables).
mc command syntax was checked with the pinned AIStor client offline.
The two new alert tests failed against the old rules and pass now.
bash -n passes for the script; shellcheck is not installed on the
laptop or the desktop.
2026-11-12 07:28:03 +09:00
DongHyeonka 1349ee2f25 docs(ci): list daemon.json in the runner runtime folder 2026-11-12 07:28:03 +09:00
DongHyeonka fb5535aeb6 docs(host): add the owner steps for the desktop host settings
The host files need root, so agent sessions do not install them. The
owner needs one ordered list with checks and rollback for each step,
including when a Docker restart is safe.

Add infrastructure/host/README.md: why the disk and clock settings
exist, the BuildKit key meanings in Docker 29, and four steps. Pin the
tech-log images against kubelet image GC until they move to the Gitea
registry, install the k3s time-sync drop-ins, set the RTC to UTC on
the Windows side, and install the host daemon.json during a window
with no CI job, no agent gate holding the heavy lock, and workmachine
not in use.

Risk: the image pin is a manual label on the current tags; a new
deployment tag needs the label again until the registry move lands.

Verification: the commands were checked on the desktop where they do
not need root (systemctl show, timedatectl, docker buildx inspect,
docker exec gitea-runner docker ps); the sudo steps were not run.
2026-11-12 07:28:03 +09:00
DongHyeonka 03ed6719b0 feat(host): start k3s only after the clock is synchronized
The desktop dual-boots Windows, which writes local time (KST) to the
RTC, while Ubuntu reads the RTC as UTC. On 2026-11-11 the boot started
9 hours ahead; k3s issued its certificates at that time, timesyncd
then set the clock back, the certificates were not yet valid, and
every Ingress returned 502 until k3s was restarted by hand.

Add a k3s.service drop-in with Wants= and After=time-sync.target, and
a drop-in that caps systemd-time-wait-sync.service at 10 minutes.
Enabling systemd-time-wait-sync makes time-sync.target wait for the
first NTP sync; its packaged TimeoutStartSec=infinity would keep k3s
down forever when NTP is unreachable, so after 10 minutes it fails,
the target is reached anyway and k3s starts as it does today.

Risk: the order only applies at boot, so nothing changes until the
next reboot. If NTP is down at boot, k3s starts 10 minutes later. The
Windows side still writes local time until RealTimeIsUniversal is set
there; the host guide covers that step.

Verification: systemd-analyze verify passes for k3s.service and
systemd-time-wait-sync.service with the drop-ins, and an unknown key
added to each drop-in is reported, so both files are read.
2026-11-12 07:28:03 +09:00
DongHyeonka a5f0bff26e feat(ci): cap the runner Docker and prune it weekly
The Docker daemon inside the gitea-runner container keeps every job
image and all build cache in the gitea-runner_docker volume on the
root SSD that production k3s also uses. It grew to 16G in the two days
after the runner was installed and held 12.43G of images and 8.44G of
build cache on 2026-11-11. The README only said to look once a week
and prune by hand.

Mount daemon.json into the runner so the inner BuildKit GC keeps 2GB
to 10GB of build cache, and shrinks to 2GB while the disk has less
than 35GB free. Add gitea-runner-prune.service and .timer, a host
systemd timer for Sunday 04:30 KST that removes unused images created
more than 7 days ago and build cache older than 3 days. The service
runs as donghyeon in the docker group, so it needs root only to
install. A user timer was rejected because lingering is off for
donghyeon and user units stop with the last login session.

Risk: image prune filters on image creation time, so the pinned job
image is deleted weekly and the next job pulls about 543MB again.
Recreating the runner to pick up daemon.json cancels a running job,
so it is done only while the runner is idle.

Verification: systemd-analyze verify passes for both units, and the
calendar expression's next run is Sun 2026-11-15 04:30 KST. docker
compose config passes; dockerd --validate inside the runner (Docker
29.9.3) accepts daemon.json.
2026-11-12 07:28:03 +09:00
DongHyeonka 3bca2cf841 feat(host): cap the host Docker build cache
The host Docker daemon had no daemon.json. The default BuildKit policy
on this 228GiB disk keeps 21.42GiB of build cache untouched and allows
169.5GiB (docker buildx inspect default), so agent gate builds piled
up about 23G in two days before the 2026-11-09 disk incident.

Add infrastructure/host/docker/daemon.json with builder.gc
defaultReservedSpace 5GB, defaultMaxUsedSpace 15GB and
defaultMinFreeSpace 35GB. In Docker 29 the old defaultKeepStorage key
means defaultReservedSpace, a floor and not a cap, so the cap is
defaultMaxUsedSpace. 35GiB free is about 85% use, where the kubelet
starts deleting images; below it the cache shrinks to the floor first.
Moving the Docker data root to the HDD was rejected by the owner
because builds get slower.

Risk: applying needs sudo and a Docker restart, which restarts the
runner, workmachine and cloudflared and fails running CI jobs and
gates. The host apply guide added in this series says when that is
safe. dockerd --validate does not check keys inside builder.gc, so
the guide checks the live policy with docker buildx inspect.

Verification: dockerd --validate passes for the file on Docker
29.6.1; it also passes a typo inside builder.gc and rejects an unknown
top-level key, which is why the live check is in the guide.
2026-11-12 07:28:03 +09:00
DongHyeonka 7fe38428e3 feat(observability): alert on the root disk before kubelet acts
The kubelet starts deleting unused images at 85% root use and evicts
pods at 90%. The root disk warning fired at 85% and the critical at
90%, so both arrived when the kubelet was already acting. On
2026-11-09 the warning fired for 33 hours, then an eviction stopped
the tech-log pods, the image GC deleted their imagePullPolicy: Never
images, and production went down.

Move the warning to 80% for 10m and the critical to 85% for 5m, and
add PlatformRootFilesystemFillingFast (critical, for 15m): the 6h
linear trend of free bytes falls below 10% of the disk, the eviction
point, within 6 hours. On 2026-11-07 the disk lost 17G in 9 hours;
this forecast would have fired hours before the 80% warning. The
info alert at 75% stays.

Add scripts/validate/test-platform-disk-alerts.sh, promtool unit
tests for the usage bands and the forecast with the pinned Prometheus
image. The existing rules test needs live metric inventories, so this
one runs offline.

Risk: the live rules change only when the observability access apply
runs; this commit does not touch the cluster. The forecast can fire
on a large build that the BuildKit GC removes minutes later; the 15m
hold and the 6h window damp that.

Verification: the new test failed against the old rules (no alert at
81%, 86% or for a 1-per-minute fill) and passes after the change;
promtool check rules finds 9 rules. test-observability-alerting.sh
passes its static contract, then stops at the helm render because
helm v3.19.4 is not installed on the desktop.
2026-11-12 07:28:03 +09:00