jakman85andClaude Opus 5
6a8953e288
Minimal prod delivery, authored directly on main (= prod's live revision d1a6628)
rather than by merging work/prod-standup-campaign. That branch carries the DNS apex
inversion, which would have set prod's BASE_DOMAIN to home.keithjackson.dev; prod
runs external-dns with policy: sync and txtOwnerId: prod, so it would have begun
rewriting apex records dev owns while the non-k3s prod still serves 443 there.
Operator sequencing is "dev migrates first". Nothing under clusters/ is touched
here at all, which is what makes this safe to land now.
Also deliberately EXCLUDED, because prod's live cluster-vars has no CNPG_BACKUP_PATH
key and StrictPostBuildSubstitutions would fail the whole tier: the R2 rework of
platform/velero/ and the CNPG destinationPath changes. Those land as
feat/r2-backup-prod together with the endpoint and bucket knobs.
WHAT THIS FIXES
1. Nothing was watching CNPG. All three prod databases have run since bootstrap with
ContinuousArchiving=False and firstRecoverabilityPoint=null, and no alert existed
anywhere. Because archive_mode=on and the archive command fails, Postgres refuses
to recycle WAL: keycloak-db holds every segment ever written (257 MiB on a 5 GiB
volume, ~6.8 MiB/h) — roughly 29 days to a disk-full PANIC on the cluster's SSO
database. Found by hand, 38h late. CNPGWALArchiveBacklog would have fired within
the hour, and evaluates to 3 firing series on prod today.
Rules key on cnpg_collector_* rather than cnpg_pg_stat_archiver_*, because
nextcloud-db's exporter cannot log in to its own database (permission denied for
database "nextcloud", both clusters) so the pg_stat_archiver family does not exist
for it — the obvious metric would have been blind to one database in three.
2. No PVC outside namespace=monitoring had any fill alert. New rules are partitioned
by namespace so they cannot double-fire with MonitoringVolumeFillingUp, pinned to
job="kubelet" (the apiserver copy of the same series reports namespace=default),
and guarded by inodes > 0 — block-mode PVCs such as unifi-rootdisk report
available=0 forever and would otherwise have fired critical permanently.
3. VPNTunnelDown could never clear on a cluster with no gateway node; it has been
firing critical on prod since 2026-09-05. Now guarded so prod self-arms the moment
a node carries the egress-gateway label, with dev's behaviour unchanged today.
4. The Keycloak realm import fails OPEN: an unresolved placeholder is stored
LITERALLY and the import still reports success, then sits at Done=True and never
re-runs. On 2026-09-06 all 12 prod OIDC clients held the literal
$(env:..._CLIENT_SECRET) string and every SSO login failed, while every signal was
green. The CronJob is the level-triggered correction; two alerts make its silence
observable. Repaired live already — this stops it recurring.
Also: the R-NET-04 conformance scan descended into nested agent worktrees under
.claude/, counting 4 egress policies where 1 is authored.
Co-Authored-By: Claude Opus 5 <[email protected]>
200 lines
12 KiB
YAML
200 lines
12 KiB
YAML
# =============================================================================
|
|
# L2 BEHAVIORAL — the default egress path is the WAN; the VPN is opt-in (R-NET-04) [RUNNABLE]
|
|
# Catalog: docs/14 L2 row ("real runtime behavior, probes against a cluster")
|
|
# Spec: docs/00 R-NET-04 ("The platform MUST support **multiple egress paths** selected by
|
|
# routing rules ... realized as two paths — the local `internet` path and the
|
|
# in-cluster `vpn` path").
|
|
# docs/03 §"Egress routing: two paths, one physical WAN": "the `internet` path is
|
|
# simply the default route out".
|
|
# docs/03 §"VPN egress model": "**Default is the WAN; the VPN is strictly opt-in.**
|
|
# Every pod — and the nodes themselves — egress the local/private WAN by default and
|
|
# **never** the VPN. Only services labeled `egress-class=vpn` are steered to the
|
|
# tunnel"; and the tunnel "is **split-tunnel**: ... only what Cilium points at `tun0`
|
|
# uses the VPN."
|
|
#
|
|
# WHAT THIS PROVES THAT tests/l2-behavioral/egress-class DOES NOT
|
|
# --------------------------------------------------------------
|
|
# That test (R-NET-05/07) compares two LABELLED pods — `internet` vs `vpn` — and shows their
|
|
# public IPs differ, then flips a label. Both of its pods declared a class, so it says nothing
|
|
# about the case R-NET-04 actually turns on: a pod that NO routing rule selects. This test adds
|
|
# that pod, and asserts the two claims that make a rule-selected path set complete:
|
|
#
|
|
# TOTALITY — an UNLABELLED pod egresses the WAN, i.e. the same public IP as the
|
|
# explicitly-`internet` pod. Absence of a matching rule IS a path, and it is
|
|
# the local one. If the steering rule ever widened (an empty selector, a
|
|
# matchExpression, a second rule), this pod's IP would move to the tunnel and
|
|
# this assertion is what would notice.
|
|
# COEXISTENCE — the WAN path and the VPN path are live AT THE SAME TIME, observed in one
|
|
# pass: three pods, two distinct public IPs. "Multiple egress paths" is a
|
|
# simultaneity claim, not a claim that one switch has two positions.
|
|
# PER-DESTINATION SELECTION — the vpn-class pod still reaches an in-cluster/RFC1918
|
|
# destination over the LOCAL path while its public traffic is tunnelled. The
|
|
# tunnel is destination-scoped (excludedCIDRs), not a default route, so ONE pod
|
|
# uses BOTH paths at once.
|
|
#
|
|
# ENVIRONMENT-AGNOSTIC (R-ENV-02): no address, hostname or range appears. Every assertion is a
|
|
# RELATION between observed values (equal / not-equal), so it holds against dev and prod alike
|
|
# and survives the WAN or the VPN provider rotating addresses.
|
|
#
|
|
# GIVEN one unlabelled, one `internet`-class and one `vpn`-class pod — the same workload three
|
|
# times, differing only in the class they declare (or do not),
|
|
# WHEN each asks an external echo for the public IP it egressed from, and the vpn-class pod
|
|
# also resolves a name through the in-cluster resolver,
|
|
# THEN the unlabelled pod's public IP EQUALS the internet-class pod's (the default route out
|
|
# is the WAN, and the VPN is strictly opt-in);
|
|
# AND the vpn-class pod's public IP DIFFERS from both (a second, simultaneously-live path);
|
|
# AND the vpn-class pod still resolves in-cluster (its private-destination traffic was not
|
|
# swallowed by the tunnel), so the same pod is using both paths at once.
|
|
#
|
|
# WHAT THIS DOES *NOT* PROVE (honest gaps — do not read this test as claiming them):
|
|
# - THAT NO OTHER RULE EXISTS. Three pods cannot enumerate a rule set. That exactly one
|
|
# steering rule is authored, that it selects only the vpn class, that no `ip rule` is
|
|
# unscoped and no nftables DROP unqualified is the L0 half, in
|
|
# tests/l0-static/egress-paths-selected-by-rules (same requirement label).
|
|
# - WHICH INTERFACE the packets left by. The observation is the public IP an external echo
|
|
# reports; it distinguishes the paths without naming them. Asserting "out tun0" would need
|
|
# host-level datapath inspection (Hubble/tcpdump on the gateway node), which no test in
|
|
# this suite does.
|
|
# - THE KILL-SWITCH. Tunnel-down behaviour is R-NET-06's, at L3 (vpn-killswitch): `vpn`
|
|
# fails closed. Its mirror — that `internet` is never blackholed — is asserted statically
|
|
# by the L0 half, not here.
|
|
# - NODE traffic. docs/03 also says "the nodes themselves" egress the WAN by default. That
|
|
# is a host-level fact this test cannot see from a pod.
|
|
# - EXTERNAL DEPENDENCY, stated plainly: the public-IP echo (api.ipify.org) is off-cluster,
|
|
# so a WAN outage or the echo being down fails this test. That is the same dependency
|
|
# tests/l2-behavioral/egress-class already carries; the alternative — trusting the
|
|
# cluster's own account of which path it used — would not be a behavioural test at all.
|
|
#
|
|
# CLEANUP: chainsaw deletes the per-test namespace (all three pods with it). Nothing outside
|
|
# that namespace is created or changed — no label on a platform object, no policy edit.
|
|
# =============================================================================
|
|
apiVersion: chainsaw.kyverno.io/v1alpha1
|
|
kind: Test
|
|
metadata:
|
|
name: egress-default-path
|
|
labels:
|
|
conformance.home.k3s/R-NET-04: "true"
|
|
annotations:
|
|
conformance.home.k3s/requirements: "R-NET-04"
|
|
conformance.home.k3s/layer: "L2"
|
|
conformance.home.k3s/safety-critical: "false"
|
|
conformance.home.k3s/executable: "true"
|
|
conformance.home.k3s/intent: >-
|
|
Given an unlabelled, an internet-class and a vpn-class pod, when each observes the public
|
|
IP it egressed from, then the unlabelled pod's IP equals the internet-class pod's (the
|
|
default route out is the WAN, the VPN is opt-in) while the vpn-class pod's differs — two
|
|
paths live at once — and the vpn-class pod still reaches in-cluster destinations over the
|
|
local path, so one pod uses both paths simultaneously.
|
|
spec:
|
|
timeouts:
|
|
apply: 60s
|
|
assert: 2m
|
|
exec: 3m
|
|
steps:
|
|
# -----------------------------------------------------------------------------------
|
|
- name: given-three-pods-differing-only-in-their-declared-egress-class
|
|
try:
|
|
- apply:
|
|
file: fixtures/three-egress-pods.yaml
|
|
# Bare pods (restartPolicy: Never) — Ready == Running here.
|
|
- assert:
|
|
resource:
|
|
apiVersion: v1
|
|
kind: Pod
|
|
metadata: { name: egress-unlabelled }
|
|
status: { phase: Running }
|
|
- assert:
|
|
resource:
|
|
apiVersion: v1
|
|
kind: Pod
|
|
metadata: { name: egress-internet }
|
|
status: { phase: Running }
|
|
- assert:
|
|
resource:
|
|
apiVersion: v1
|
|
kind: Pod
|
|
metadata: { name: egress-vpn }
|
|
status: { phase: Running }
|
|
# -----------------------------------------------------------------------------------
|
|
- name: then-the-unlabelled-pod-takes-the-wan-and-the-vpn-path-is-live-too
|
|
try:
|
|
- script:
|
|
content: |
|
|
set -eu
|
|
ns="$NAMESPACE"
|
|
get_ip() {
|
|
# Echo the pod's observed public IP, or empty on failure.
|
|
kubectl -n "$ns" exec "$1" -c curl -- \
|
|
curl -fsS --max-time 10 https://api.ipify.org 2>/dev/null || true
|
|
}
|
|
|
|
# Poll until all three answer AND the vpn class has diverged from the WAN:
|
|
# Cilium may still be programming the egress datapath just after the pods go
|
|
# Ready. Bounded by the loop, never a bare sleep-and-hope.
|
|
none="" inet="" vpn=""
|
|
for i in $(seq 1 12); do
|
|
[ -n "$none" ] || none="$(get_ip egress-unlabelled)"
|
|
[ -n "$inet" ] || inet="$(get_ip egress-internet)"
|
|
[ -n "$vpn" ] || vpn="$(get_ip egress-vpn)"
|
|
if [ -n "$none" ] && [ -n "$inet" ] && [ -n "$vpn" ]; then
|
|
[ "$vpn" != "$inet" ] && break
|
|
# Steering may not have converged; re-read the vpn pod only.
|
|
vpn=""
|
|
fi
|
|
sleep 5
|
|
done
|
|
echo "unlabelled_ip=$none internet_ip=$inet vpn_ip=$vpn"
|
|
|
|
[ -n "$none" ] || { echo "FAIL: the UNLABELLED pod never returned a public IP — a pod that declares no egress class must still reach the internet over the default WAN path (docs/03: 'the internet path is simply the default route out'); no path at all means the default was blackholed or steered into a dead tunnel" >&2; exit 1; }
|
|
[ -n "$inet" ] || { echo "FAIL: the internet-class pod never returned a public IP — the declared default path is down, so no comparison is possible" >&2; exit 1; }
|
|
[ -n "$vpn" ] || { echo "FAIL: the vpn-class pod never returned a public IP — the second egress path is not carrying traffic, so the platform is not offering MULTIPLE paths (R-NET-04)" >&2; exit 1; }
|
|
|
|
# TOTALITY: no rule selects the unlabelled pod, so it must land on the WAN —
|
|
# the same WAN the explicitly-internet pod uses.
|
|
if [ "$none" != "$inet" ]; then
|
|
echo "FAIL: the UNLABELLED pod egressed $none but the internet-class pod egressed $inet. They must be the same path: docs/03 — 'Default is the WAN; the VPN is strictly opt-in. Every pod ... egress the local/private WAN by default and never the VPN.' A different IP means some rule is selecting a pod that never opted in." >&2
|
|
exit 1
|
|
fi
|
|
# COEXISTENCE: and the vpn path is a genuinely different path, live right now.
|
|
if [ "$vpn" = "$inet" ]; then
|
|
echo "FAIL: the vpn-class pod egressed the same public IP as the WAN ($vpn) — there is only ONE live egress path, so 'multiple egress paths selected by routing rules' (R-NET-04) does not hold" >&2
|
|
exit 1
|
|
fi
|
|
echo "EGRESS_DEFAULT_PATH_IS_WAN_OK"
|
|
check:
|
|
(contains($stdout, 'EGRESS_DEFAULT_PATH_IS_WAN_OK')): true
|
|
# -----------------------------------------------------------------------------------
|
|
- name: then-the-vpn-class-pod-still-reaches-private-destinations-locally
|
|
try:
|
|
- script:
|
|
# PER-DESTINATION SELECTION. The steering rule tunnels 0.0.0.0/0 MINUS all
|
|
# RFC1918 space, so a vpn-class pod's private-destination traffic must stay on the
|
|
# local datapath. The cluster resolver is an in-cluster (RFC1918) address reached
|
|
# from the pod's own /etc/resolv.conf — no literal is written here — so a
|
|
# successful lookup from the vpn-class pod proves the tunnel did not swallow its
|
|
# local traffic, i.e. ONE pod is using BOTH paths at the same time.
|
|
content: |
|
|
set -eu
|
|
ns="$NAMESPACE"
|
|
resolver="$(kubectl -n "$ns" exec egress-vpn -c curl -- \
|
|
sh -c 'awk "/^nameserver/{print \$2; exit}" /etc/resolv.conf' 2>/dev/null || true)"
|
|
[ -n "$resolver" ] || { echo "FAIL: the vpn-class pod has no nameserver in /etc/resolv.conf, so its local-path reachability cannot be observed" >&2; exit 1; }
|
|
echo "vpn_pod_resolver=$resolver (in-cluster, RFC1918 — excluded from the tunnel)"
|
|
|
|
# A cluster-internal name, resolvable only by the in-cluster resolver: reaching
|
|
# it proves the private-destination path works from a vpn-class pod.
|
|
out=""
|
|
for i in $(seq 1 6); do
|
|
out="$(kubectl -n "$ns" exec egress-vpn -c curl -- \
|
|
nslookup kubernetes.default.svc.cluster.local 2>/dev/null || true)"
|
|
case "$out" in *"Address"*) break ;; esac
|
|
sleep 5
|
|
done
|
|
case "$out" in
|
|
*Address*) : ;;
|
|
*) echo "FAIL: the vpn-class pod could not resolve kubernetes.default.svc.cluster.local via its in-cluster resolver ($resolver). Its RFC1918 traffic is being steered into the tunnel (or blackholed) instead of taking the local path — the two paths are not coexisting per destination, and the tunnel is behaving as a default route (docs/03: split-tunnel, 'only what Cilium points at tun0 uses the VPN')" >&2; exit 1 ;;
|
|
esac
|
|
echo "EGRESS_PATHS_COEXIST_PER_DESTINATION_OK"
|
|
check:
|
|
(contains($stdout, 'EGRESS_PATHS_COEXIST_PER_DESTINATION_OK')): true
|