51313ee088
The k8s-master gate pod OOM-killed child kubectls 149x/7d (accelerating: 0/day → 15 → 134) while master sat in pending-reboot. Root cause: only the pending-reboot node's gate pod runs the kubectl-heavy hot path each cycle, and the immortal bash loop slowly leaks (kubectl forks + Check-4 process substitution) past the 64Mi cgroup limit. PID 1 bash survives each kill, so the pod never restarts — just silent oom_events. Fix: raise limit 64Mi→256Mi (headroom for ~30-50Mi kubectl forks) + add a MAX_ITER=72 self-exit (~6h) so kubelet restarts the pod fresh and the leak can never accumulate, regardless of how long a node stays pending-reboot. Docs: post-mortem + automated-upgrades.md gate note. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
347 lines
14 KiB
HCL
347 lines
14 KiB
HCL
# kured — Kubernetes Reboot Daemon
|
||
#
|
||
# Auto-reboots nodes when /var/run/reboot-required exists on the host (set by
|
||
# unattended-upgrades). The reboot process is gated by a custom sentinel file
|
||
# (kured-sentinel-gate DaemonSet below) and by Prometheus alerts so reboots
|
||
# only happen when:
|
||
# - all nodes Ready
|
||
# - all calico-node pods Running
|
||
# - no node has transitioned Ready in the last 24 hours (24h soak)
|
||
# - no Prometheus alert is firing (excluding self-referential ignore-list)
|
||
#
|
||
# History:
|
||
# - 2026-03 post-mortem (memory 390): 26h cluster outage triggered by kured
|
||
# rebooting nodes while containerd's overlayfs snapshotter was corrupted.
|
||
# Remediation included the sentinel gate and a tight reboot window
|
||
# (Mon-Fri 02:00-06:00 London).
|
||
# - 2026-04-18: adopted into Terraform (Wave 5a). Previously helm-installed
|
||
# manually + kubectl-applied sentinel gate.
|
||
# - 2026-05-10: re-enabled unattended-upgrades (cloud_init.yaml flipped from
|
||
# remove → install). Sentinel cool-down stretched 30m → 24h. Added Helm
|
||
# values prometheusUrl + alertFilterRegexp so any non-ignored firing alert
|
||
# halts the rollout. New "Upgrade Gates" alert group in monitoring stack
|
||
# (KubeAPIServerDown, KubeStateMetricsDown, PrometheusRuleEvaluationFailing,
|
||
# PVCStuckPending, RecentNodeReboot, MysqlStandaloneDown,
|
||
# ClusterPodReadyRatioDropped, NodeMemoryPressure, NodeDiskPressure,
|
||
# KubeQuotaAlmostFull) provides explicit cluster-health gating.
|
||
|
||
resource "kubernetes_namespace" "kured" {
|
||
metadata {
|
||
name = "kured"
|
||
labels = {
|
||
"istio-injection" = "disabled"
|
||
tier = local.tiers.cluster
|
||
"keel.sh/enrolled" = "true"
|
||
}
|
||
}
|
||
lifecycle {
|
||
# KYVERNO_LIFECYCLE_V1: goldilocks-vpa-auto-mode ClusterPolicy stamps this label on every namespace
|
||
ignore_changes = [metadata[0].labels["goldilocks.fairwinds.com/vpa-update-mode"]]
|
||
}
|
||
}
|
||
|
||
# -----------------------------------------------------------------------------
|
||
# kured Helm release
|
||
# -----------------------------------------------------------------------------
|
||
|
||
resource "helm_release" "kured" {
|
||
namespace = kubernetes_namespace.kured.metadata[0].name
|
||
create_namespace = false
|
||
name = "kured"
|
||
chart = "kured"
|
||
repository = "https://kubereboot.github.io/charts/"
|
||
version = "5.11.0"
|
||
|
||
values = [yamlencode({
|
||
configuration = {
|
||
period = "1h0m0s"
|
||
timeZone = "Europe/London"
|
||
startTime = "02:00"
|
||
endTime = "06:00"
|
||
# All 7 days — operator decision 2026-05-16. The Mon–Fri restriction
|
||
# was a 2026-03-16-era guardrail (overlapping with weekend on-call
|
||
# response). Today the rest of the safety net (halt-on-alert,
|
||
# sentinel-gate Check 4 = 24h soak, single-concurrency, the
|
||
# K8sUpgradeStalled alert) is strong enough to operate any day; the
|
||
# weekday-only schedule was just slowing the backlog down.
|
||
rebootDays = ["mo", "tu", "we", "th", "fr", "sa", "su"]
|
||
# IMPORTANT: must match where kured-sentinel-gate writes (below):
|
||
# `touch /host/var-run/gated-reboot-required` → host
|
||
# `/var/run/gated-reboot-required`. The kured chart derives the host
|
||
# path from `dirname(rebootSentinel)`, so this single setting controls
|
||
# BOTH the in-pod mountPath AND the host hostPath. Previously
|
||
# `/sentinel/gated-reboot-required` — that pointed the chart's hostPath
|
||
# at `/sentinel/` (empty, auto-created by hostPath:Directory) while the
|
||
# gate kept writing to `/var/run/`. kured never saw the open gate so
|
||
# nodes stopped rebooting on 2026-05-10 when unattended-upgrades was
|
||
# re-enabled. Fixed 2026-05-16.
|
||
rebootSentinel = "/var/run/gated-reboot-required"
|
||
notifyUrl = data.vault_kv_secret_v2.secrets.data["slack_kured_webhook"]
|
||
concurrency = 1
|
||
rebootDelay = "30s"
|
||
# Fail closed instead of looping forever. Default is 0 (unlimited) — if
|
||
# a future PDB or finalizer stalls drain, kured retries indefinitely and
|
||
# the node stays cordoned silently. 30m gives CNPG / shared-store
|
||
# Anubis / any other stateful workload plenty of time to settle, but
|
||
# caps the silent-failure window. After timeout kured logs the abort
|
||
# and waits for the next period; node stays Schedulable so the cluster
|
||
# doesn't lose capacity. Fixed 2026-05-16.
|
||
drainTimeout = "30m"
|
||
# Halt rolling reboots when ANY firing Prometheus alert is not in the
|
||
# ignore-list. The ignore-list excludes self-referential / always-firing
|
||
# alerts that would otherwise deadlock kured. alertFilterMatchOnly stays
|
||
# false (default) so the regex marks alerts to IGNORE — every other
|
||
# firing alert blocks. See "Upgrade Gates" group in monitoring stack.
|
||
prometheusUrl = "http://prometheus-server.monitoring.svc.cluster.local:80"
|
||
alertFilterRegexp = "^(Watchdog|RebootRequired|KuredNodeWasNotDrained|InfoInhibitor)$"
|
||
alertFiringOnly = true
|
||
alertFilterMatchOnly = false
|
||
}
|
||
reboot_days = "mon,tue,wed,thu,fri,sat,sun"
|
||
window_end = "06:00"
|
||
window_start = "22:00"
|
||
service = {
|
||
annotations = {
|
||
"prometheus.io/scrape" = "true"
|
||
"prometheus.io/port" = "8080"
|
||
"prometheus.io/path" = "/metrics"
|
||
}
|
||
}
|
||
})]
|
||
}
|
||
|
||
data "vault_kv_secret_v2" "secrets" {
|
||
mount = "secret"
|
||
name = "kured"
|
||
}
|
||
|
||
# -----------------------------------------------------------------------------
|
||
# kured-sentinel-gate
|
||
#
|
||
# Runs a DaemonSet that creates /var/run/gated-reboot-required ONLY when all
|
||
# safety preconditions are met (see script). kured's rebootSentinel points at
|
||
# this file, so reboots are effectively blocked unless every check passes.
|
||
# -----------------------------------------------------------------------------
|
||
|
||
resource "kubernetes_service_account" "kured_sentinel_gate" {
|
||
metadata {
|
||
name = "kured-sentinel-gate"
|
||
namespace = kubernetes_namespace.kured.metadata[0].name
|
||
}
|
||
# Token IS mounted — the script uses kubectl to read nodes + pods state for
|
||
# the safety checks. Without an authenticated token, kubectl falls back to
|
||
# localhost:8080 (no proxy in distroless-ish image), every check silently
|
||
# no-ops on parse-empty stdout, and the gate appears to PASS when it
|
||
# shouldn't. Mount the token. (Found 2026-05-10 during Test 3 validation.)
|
||
automount_service_account_token = true
|
||
}
|
||
|
||
resource "kubernetes_cluster_role" "kured_sentinel_gate" {
|
||
metadata {
|
||
name = "kured-sentinel-gate"
|
||
}
|
||
rule {
|
||
api_groups = [""]
|
||
resources = ["nodes"]
|
||
verbs = ["list"]
|
||
}
|
||
rule {
|
||
api_groups = [""]
|
||
resources = ["pods"]
|
||
verbs = ["list"]
|
||
}
|
||
}
|
||
|
||
resource "kubernetes_cluster_role_binding" "kured_sentinel_gate" {
|
||
metadata {
|
||
name = "kured-sentinel-gate"
|
||
}
|
||
role_ref {
|
||
api_group = "rbac.authorization.k8s.io"
|
||
kind = "ClusterRole"
|
||
name = kubernetes_cluster_role.kured_sentinel_gate.metadata[0].name
|
||
}
|
||
subject {
|
||
kind = "ServiceAccount"
|
||
name = kubernetes_service_account.kured_sentinel_gate.metadata[0].name
|
||
namespace = kubernetes_namespace.kured.metadata[0].name
|
||
}
|
||
}
|
||
|
||
resource "kubernetes_daemon_set_v1" "kured_sentinel_gate" {
|
||
metadata {
|
||
name = "kured-sentinel-gate"
|
||
namespace = kubernetes_namespace.kured.metadata[0].name
|
||
labels = {
|
||
app = "kured-sentinel-gate"
|
||
tier = local.tiers.cluster
|
||
}
|
||
}
|
||
spec {
|
||
selector {
|
||
match_labels = {
|
||
app = "kured-sentinel-gate"
|
||
}
|
||
}
|
||
template {
|
||
metadata {
|
||
labels = {
|
||
app = "kured-sentinel-gate"
|
||
}
|
||
}
|
||
spec {
|
||
service_account_name = kubernetes_service_account.kured_sentinel_gate.metadata[0].name
|
||
automount_service_account_token = true
|
||
enable_service_links = false
|
||
# bitnami/kubectl:latest runs as uid=1001 by default. The hostPath
|
||
# /var/run is root:root 0755 → final `touch
|
||
# /host/var-run/gated-reboot-required` fails with EACCES, so the gate
|
||
# never opens. Run as root inside the container (the hostPath mount
|
||
# already gives privileged-equivalent access; this just lets us write
|
||
# to /var/run). Found 2026-05-10 during Test 3 validation.
|
||
security_context {
|
||
run_as_user = 0
|
||
}
|
||
toleration {
|
||
effect = "NoSchedule"
|
||
key = "node-role.kubernetes.io/control-plane"
|
||
operator = "Equal"
|
||
}
|
||
toleration {
|
||
effect = "NoSchedule"
|
||
key = "node-role.kubernetes.io/master"
|
||
operator = "Equal"
|
||
}
|
||
container {
|
||
name = "gate"
|
||
image = "bitnami/kubectl:latest"
|
||
image_pull_policy = "Always"
|
||
command = [
|
||
"/bin/bash",
|
||
"-c",
|
||
<<-EOT
|
||
# Self-restart guard (added 2026-05-31): a node stuck in
|
||
# pending-reboot keeps THIS pod on the kubectl-heavy hot path
|
||
# every cycle. The long-lived bash loop slowly leaks (repeated
|
||
# kubectl forks + the Check-4 process substitution) until the
|
||
# cgroup OOM-kills child processes — PID 1 bash survives, so the
|
||
# pod never restarts, it just racks up silent oom_events
|
||
# (149 in 7d / accelerating on k8s-master, 2026-05-30..31).
|
||
# Exit 0 every MAX_ITER cycles (~6h at 300s) so kubelet restarts
|
||
# the pod fresh and memory can never accumulate.
|
||
ITER=0
|
||
MAX_ITER=72
|
||
while true; do
|
||
ITER=$((ITER + 1))
|
||
if [ "$ITER" -gt "$MAX_ITER" ]; then
|
||
echo " Iteration cap ($MAX_ITER) reached — exit 0 for a clean restart (leak guard)"
|
||
exit 0
|
||
fi
|
||
echo "[$(date)] Checking reboot gate conditions... (iter $ITER/$MAX_ITER)"
|
||
|
||
# Check 1: Does the host need a reboot?
|
||
if [ ! -f /host/var-run/reboot-required ]; then
|
||
echo " No reboot required on this host"
|
||
rm -f /host/var-run/gated-reboot-required
|
||
sleep 300
|
||
continue
|
||
fi
|
||
echo " Host has /var/run/reboot-required"
|
||
|
||
# Check 2: Are ALL nodes Ready?
|
||
NOT_READY=$(kubectl get nodes --no-headers | grep -v ' Ready' | wc -l | tr -d ' ')
|
||
if [ "$NOT_READY" -gt 0 ]; then
|
||
echo " BLOCKED: $NOT_READY node(s) not Ready"
|
||
rm -f /host/var-run/gated-reboot-required
|
||
sleep 300
|
||
continue
|
||
fi
|
||
echo " All nodes Ready"
|
||
|
||
# Check 3: Are ALL calico-node pods Running?
|
||
CALICO_NOT_RUNNING=$(kubectl get pods -n calico-system -l k8s-app=calico-node --no-headers 2>/dev/null | grep -v Running | wc -l | tr -d ' ')
|
||
if [ "$CALICO_NOT_RUNNING" -gt 0 ]; then
|
||
echo " BLOCKED: $CALICO_NOT_RUNNING calico-node pod(s) not Running"
|
||
rm -f /host/var-run/gated-reboot-required
|
||
sleep 300
|
||
continue
|
||
fi
|
||
echo " All calico-node pods Running"
|
||
|
||
# Check 4: No node rebooted in last 24 hours (soak window).
|
||
# Stretched from 30m to 24h on 2026-05-10 so the de-facto canary
|
||
# node has a full day of observation before the next node drains.
|
||
RECENT_REBOOT=0
|
||
while IFS= read -r transition_time; do
|
||
if [ -n "$transition_time" ]; then
|
||
transition_epoch=$(date -d "$transition_time" +%s 2>/dev/null || date -j -f "%Y-%m-%dT%H:%M:%SZ" "$transition_time" +%s 2>/dev/null)
|
||
now_epoch=$(date +%s)
|
||
diff=$(( now_epoch - transition_epoch ))
|
||
if [ "$diff" -lt 86400 ]; then
|
||
RECENT_REBOOT=1
|
||
break
|
||
fi
|
||
fi
|
||
done < <(kubectl get nodes -o jsonpath='{range .items[*]}{range .status.conditions[?(@.type=="Ready")]}{.lastTransitionTime}{"\n"}{end}{end}')
|
||
|
||
if [ "$RECENT_REBOOT" -eq 1 ]; then
|
||
echo " BLOCKED: A node transitioned Ready within the last 24 hours (soak window)"
|
||
rm -f /host/var-run/gated-reboot-required
|
||
sleep 300
|
||
continue
|
||
fi
|
||
echo " No recent node reboots (24h soak window clear)"
|
||
|
||
# All checks passed — create gated sentinel
|
||
echo " ALL CHECKS PASSED — creating /var/run/gated-reboot-required"
|
||
touch /host/var-run/gated-reboot-required
|
||
sleep 300
|
||
done
|
||
EOT
|
||
]
|
||
resources {
|
||
requests = {
|
||
cpu = "10m"
|
||
memory = "32Mi"
|
||
}
|
||
# 64Mi was too tight for the kubectl-heavy hot path (each kubectl
|
||
# fork is a ~30-50Mi Go binary). Raised to 256Mi 2026-05-31 after
|
||
# the k8s-master gate pod OOM-killed child kubectls 149x/7d while
|
||
# master sat in pending-reboot. The self-restart guard (loop above)
|
||
# is the primary leak fix; this just gives comfortable headroom.
|
||
limits = {
|
||
memory = "256Mi"
|
||
}
|
||
}
|
||
volume_mount {
|
||
name = "var-run"
|
||
mount_path = "/host/var-run"
|
||
}
|
||
}
|
||
volume {
|
||
name = "var-run"
|
||
host_path {
|
||
path = "/var/run"
|
||
type = "Directory"
|
||
}
|
||
}
|
||
}
|
||
}
|
||
}
|
||
lifecycle {
|
||
ignore_changes = [
|
||
spec[0].template[0].spec[0].dns_config, # KYVERNO_LIFECYCLE_V1
|
||
metadata[0].annotations["keel.sh/policy"],
|
||
metadata[0].annotations["keel.sh/trigger"],
|
||
metadata[0].annotations["keel.sh/pollSchedule"], # KYVERNO_LIFECYCLE_V2
|
||
metadata[0].annotations["keel.sh/match-tag"],
|
||
spec[0].template[0].spec[0].container[0].image, # KEEL_IGNORE_IMAGE — Keel manages tag updates
|
||
metadata[0].annotations["kubernetes.io/change-cause"],
|
||
metadata[0].annotations["deployment.kubernetes.io/revision"],
|
||
spec[0].template[0].metadata[0].annotations["keel.sh/update-time"], # KEEL_LIFECYCLE_V1
|
||
]
|
||
}
|
||
}
|
||
|
||
# CI retrigger 2026-05-16T13:42:57+00:00 — bulk enrollment apply (pipeline #689 killed)
|
||
# CI retrigger v2 2026-05-16T13:46:35+00:00
|