Data-driven user onboarding: add a JSON entry to Vault KV k8s_users,
apply vault + platform + woodpecker stacks, and everything is auto-generated.
Vault stack: namespace creation, per-user Vault policies with secret isolation
via identity entities/aliases, K8s deployer roles, CI policy update.
Platform stack: domains field in k8s_users type, TLS secrets per user namespace,
user domains merged into Cloudflare DNS, user-roles ConfigMap mounted in portal.
Woodpecker stack: admin list auto-generated from k8s_users, WOODPECKER_OPEN=true.
K8s-portal: dual-track onboarding (general/namespace-owner), namespace-owner
dashboard with Vault/kubectl commands, setup script adds Vault+Terraform+Terragrunt,
contributing page with CI pipeline template, versioned image tags in CI pipeline.
New: stacks/_template/ with copyable stack template for namespace-owners.
Kyverno ClusterPolicy reads dependency.kyverno.io/wait-for annotation
and injects busybox init containers that block until each dependency
is reachable (nc -z). Annotations added to 18 stacks (24 deployments).
Includes graceful-db-maintenance.sh script for planned DB maintenance
(scales dependents to 0, saves replica counts, restores on startup).
The API server doesn't re-resolve priority from PriorityClassName after
webhook mutation. Changed from remove+add to replace with explicit
priority=1200000 and preemptionPolicy=PreemptLowerPriority.
- Upgrade from 1.35.2 to 1.35.4 (fixes API key login userDecryptionOptions bug)
- Switch deployment strategy from RollingUpdate to Recreate (iSCSI PVC can't multi-attach)
Compaction of 5 years of TSDB blocks was OOM-killing at 3Gi (18 restarts
in 8h), causing sustained IO pressure on the PVE host spinning disk.
Increase liveness probe delay to 300s so WAL replay completes before
the probe kills the pod.
SQLite backup via Online Backup API + copy of RSA keys,
attachments, sends, and config. 30-day retention with rotation.
Pod affinity ensures co-scheduling with vaultwarden for RWO PVC access.
SQLite on NFS causes DB corruption due to unreliable POSIX fcntl locking.
iSCSI provides a block device with a local filesystem where locking works
correctly. Same approach used for Redis, MySQL, PostgreSQL, etc.
After node2 OOM incident, right-size memory across the cluster by setting
requests=limits based on max_over_time(container_memory_working_set_bytes[7d])
with 1.3x headroom. Eliminates ~37Gi overcommit gap.
Categories:
- Safe equalization (50 containers): set req=lim where max7d well within target
- Limit increases (8 containers): raise limits for services spiking above current
- No Prometheus data (12 containers): conservatively set lim=req
- Exception: nextcloud keeps req=256Mi/lim=8Gi due to Apache memory spikes
Also increased dbaas namespace quota from 12Gi to 16Gi to accommodate mysql
4Gi limits across 3 replicas.
CrowdSec Helm fix:
- Increase ResourceQuota requests.cpu from 1 to 4 — pods were at 302%
of quota, preventing scheduling during rolling upgrades
- Reduce Helm timeout from 3600s to 600s — 1 hour hang is excessive
- Add wait=true and wait_for_jobs=true for proper readiness checking
Prometheus startup guard:
- Add startup guard to 8 rate/increase-based alerts that false-fire
after Prometheus restarts (needs 2 scrapes for rate() to work):
PodCrashLooping, ContainerOOMKilled, CoreDNSErrors,
HighServiceErrorRate, HighService4xxRate, HighServiceLatency,
SSDHighWriteRate, HDDHighWriteRate
- Guard: and on() (time() - process_start_time_seconds) > 900
suppresses alerts for 15m after Prometheus startup
Root cause: sum(rate(node_nfs_requests_total[5m])) == 0 was too fragile:
- rate() returns nothing after Prometheus restarts (needs 2 scrapes)
- Individual nodes show zero NFS rate during scrape gaps or low activity
- The sum() could hit zero during quiet hours + scrape gaps
New expression uses:
- changes() instead of rate() — works with a single scrape
- Per-instance aggregation: count nodes with any NFS counter change
- Threshold < 2 nodes: single-node restarts won't trigger, real NFS
outage (all nodes affected) will
- Prometheus startup guard: skip first 15m after restart to avoid
false positives from empty TSDB
- Wider 15m changes() window to smooth out scrape gaps
CPU limits cause CFS throttling even when nodes have idle capacity.
Move to a request-only CPU model: keep CPU requests for scheduling
fairness but remove all CPU limits. Memory limits stay (incompressible).
Changes across 108 files:
- Kyverno LimitRange policy: remove cpu from default/max in all 6 tiers
- Kyverno ResourceQuota policy: remove limits.cpu from all 5 tiers
- Custom ResourceQuotas: remove limits.cpu from 8 namespace quotas
- Custom LimitRanges: remove cpu from default/max (nextcloud, onlyoffice)
- RBAC module: remove cpu_limits variable and quota reference
- Freedify factory: remove cpu_limit variable and limits reference
- 86 deployment files: remove cpu from all limits blocks
- 6 Helm values files: remove cpu under limits sections
The Redis K8s Service was load-balancing across both master and replica
nodes, causing READONLY errors when clients hit the replica. This broke
Nextcloud (DAV 500s, liveness probe timeouts, crash loops) and
potentially other services.
Replace the direct Service with HAProxy (2 replicas) that health-checks
each Redis node via INFO replication and only routes to role:master.
On Sentinel failover, HAProxy detects the new master within ~9 seconds.
DECISION: Disable Loki due to operational overhead vs benefit analysis
EVIDENCE FROM NODE2 INCIDENT:
- Loki was the root cause of major cluster outage (PVC storage exhaustion)
- Centralized logging was unavailable when needed most (Loki was down)
- All debugging was accomplished with simpler tools (kubectl logs, events, describe)
- Prometheus metrics proved more valuable than centralized logs
OPERATIONAL OVERHEAD ELIMINATED:
✅ 50GB iSCSI storage freed up (expensive)
✅ ~3.5GB memory freed up (Loki + Alloy agents across cluster)
✅ ~2+ CPU cores freed up for actual workloads
✅ Reduced complexity - fewer services to maintain and troubleshoot
✅ Eliminated single point of failure that can cascade cluster-wide
CONFIGURATION PRESERVED:
✅ All Terraform resources commented out (not deleted)
✅ loki.yaml preserved with 50GB configuration
✅ alloy.yaml preserved with log shipping configuration
✅ Alert rules and Grafana datasource preserved (commented)
✅ Easy re-enabling: just uncomment resources and apply
ALTERNATIVE DEBUGGING APPROACH:
✅ kubectl logs (always works, no storage dependency)
✅ kubectl get events (built-in Kubernetes events)
✅ Prometheus metrics (more reliable for monitoring)
✅ Enhanced health check scripts (direct status verification)
RE-ENABLING:
To restore Loki later, uncomment all /* ... */ blocks in loki.tf
and apply via Terraform. All configuration is preserved.
[ci skip] - Infrastructure changes applied locally first due to resource cleanup
ISSUE RESOLVED:
- Root cause: Loki's 15GB iSCSI PVC was completely full
- Symptom: 'no space left on device' errors during TSDB operations
- Impact: Loki service completely down, logging unavailable
- Side effects: Contributed to node2 containerd corruption incident
SOLUTION APPLIED:
- Expanded PVC storage: 15Gi → 50Gi via direct kubectl patch
- Triggered pod restart to complete filesystem resize
- Verified successful expansion and service recovery
CURRENT STATUS:
✅ PVC: 50Gi capacity (iscsi-truenas storage class)
✅ Loki StatefulSet: 1/1 ready
✅ Loki Pod: 2/2 containers running
✅ Service: Successfully processing log streams
✅ No storage errors in recent logs
TERRAFORM ALIGNED:
- Updated loki.yaml persistence.size to match actual PVC
- Infrastructure code now reflects deployed state
[ci skip] - Emergency fix applied locally first due to service outage
MySQL (3 OOM kills):
- Cap group_replication_message_cache_size to 128MB (default 1GB caused OOM)
- Reduce innodb_log_buffer_size from 64MB to 16MB
- Lower max_connections from 151 to 80 (peak usage ~40)
- Increase memory limit from 3Gi to 4Gi for headroom
Nextcloud (30+ apache2 OOM kills per incident):
- Reduce MaxRequestWorkers from 50 to 10 to prevent fork bomb
when SQLite locks cause request pileup
- Lower StartServers/MinSpare/MaxSpare proportionally
Uptime Kuma (Node.js memory leak):
- Increase memory limit from 256Mi to 512Mi
- Increase CPU limit from 200m to 500m
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The GPU operator needs ~19 CPU limits across 16 pods (NFD, device plugin,
driver, validators, exporters). The Kyverno auto-generated quota of 16 CPU
was insufficient, blocking NFD worker and GC pods from scheduling.
- Add custom-quota label to nvidia namespace to exempt from Kyverno generation
- Add explicit ResourceQuota with limits.cpu=32, limits.memory=48Gi
- Fix: nvidia namespace tier label was missing after CI re-apply, causing
Kyverno to use fallback LimitRange instead of tier-2-gpu specific one
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Pin actualbudget/actual-server from edge to 26.3.0 (all 3 instances) to
prevent recurring migration breakage from rolling nightly builds
- Add podAntiAffinity to MySQL InnoDB Cluster to spread replicas across nodes,
relieving memory pressure on k8s-node4
- Scale grampsweb to 0 replicas (unused, consuming 1.7Gi memory)
- Add GPU toleration Kyverno policy to Terraform using patchesJson6902 instead
of patchStrategicMerge to fix toleration array being overwritten (caused
caretta DaemonSet pod to be unable to schedule on k8s-master)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- MaxRequestWorkers 25→50: too few workers caused ALL workers to block
on SQLite locks, making liveness probes fail even faster (131 restarts
vs 50 before). 50 is a compromise — enough workers for probes.
- Excluded nextcloud from HighServiceErrorRate alert (chronic SQLite issue)
- MySQL migration attempted but hit: GR error 3100 (fixed with GIPK),
emoji in calendar/filecache (stripped), SQLite corruption (pre-existing
from crash-looping). Migration rolled back, Nextcloud restored to SQLite.
- JobFailed: only alert on jobs started within the last hour, so stale
failed CronJob runs don't keep firing after subsequent runs succeed
- Mail server alert: renamed to MailServerDown, now targets the specific
mailserver deployment instead of all deployments in the namespace
(was falsely triggering on roundcubemail going down)
- Updated inhibition rule to use new MailServerDown alert name
Adds Sealed Secrets (Bitnami) to the platform stack so cluster users can
encrypt secrets with a public key and commit SealedSecret YAMLs to git.
The in-cluster controller decrypts them into regular K8s Secrets.
- New module: sealed-secrets (namespace + Helm chart v2.18.3, cluster tier)
- k8s-portal setup script: adds kubeseal CLI install for Linux and Mac
Previously manualStartOnBoot=true and exitStateAction=ABORT_SERVER meant
any ungraceful shutdown required manual rebootClusterFromCompleteOutage().
New settings:
- group_replication_start_on_boot=ON: auto-start GR after crash
- autorejoin_tries=2016: retry rejoining for ~28 minutes
- exit_state_action=OFFLINE_MODE: stay alive on expulsion (don't abort)
- member_expel_timeout=30s: tolerate brief unresponsiveness
- unreachable_majority_timeout=60s: leave group cleanly if majority lost
LimitRange defaults had a 4-8x limit/request ratio causing the scheduler
to overpack nodes. When pods burst, nodes OOM-thrashed and became
unresponsive (k8s-node3 and k8s-node4 both went down today).
Changes:
- Increase default memory requests across all tiers (ratio now 2x):
- core/cluster: 64Mi → 256Mi request (512Mi limit)
- gpu: 256Mi → 1Gi request (2Gi limit)
- edge/aux/fallback: 64Mi → 128Mi request (256Mi limit)
- Add kubelet memory reservation and eviction thresholds:
- systemReserved: 512Mi, kubeReserved: 512Mi
- evictionHard: 500Mi (was 100Mi), evictionSoft: 1Gi (was unset)
- Applied to all nodes and future node template
- qBittorrent: use service port 80 (not container port 8080)
- Immich: add version=2 for new API endpoints (/api/server/*)
- Nextcloud: use external URL (internal rejects untrusted Host header)
- HA London: remove widget (token expired, needs manual regeneration)
- Headscale: remove widget (requires nodeId param, not overview)
Services expose port 80 via ClusterIP but widgets were using container
target ports (5000, 3001, 4533, 3000). Calibre was using external URL
through Authentik. All now use correct internal service URLs.
Wire homepage_credentials tokens through platform stack to enable
live widgets for Authentik, Shlink (URL shortener), and Home Assistant
London. Update SOPS with new credential entries.
