Skip to content

Commit 35a61e6

Browse files
committed
test(tenant): pin ancestor egress to vlinsert and vminsert
The <tenant>-egress CiliumClusterwideNetworkPolicy opens child-tenant egress to vminsert/vlinsert/etcd/ingress in ancestor namespaces. These are allow rules on a clusterwide policy and act as a security boundary, so a silent template regression would quietly cut a child tenant's metric or log delivery. Add a helm-unittest suite for templates/networkpolicy.yaml covering the full contract: a depth-2 child tenant's egress must contain both the vlinsert and vminsert toEndpoints blocks scoped to the parent namespace, and the root tenant must emit no ancestor ingest block (guard pinned by notContains plus an exact egress-length assertion). Signed-off-by: IvanHunters <[email protected]>
1 parent 338b89c commit 35a61e6

1 file changed

Lines changed: 77 additions & 0 deletions

File tree

Lines changed: 77 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,77 @@
1+
suite: tenant egress to ancestor observability ingest
2+
# The <tenant>-egress CiliumClusterwideNetworkPolicy (networkpolicy.yaml)
3+
# opens child-tenant egress to a fixed set of ancestor-namespace services:
4+
# vminsert (VictoriaMetrics ingest), vlinsert (VictoriaLogs ingest), etcd,
5+
# and ingress controllers. A child tenant running monitoring-agents relies
6+
# on these allowances to ship metrics and container logs up to a parent
7+
# tenant's observability stack with no hand-edited CCNP.
8+
#
9+
# vlinsert is a security boundary: it is an allow rule on a clusterwide
10+
# policy, so a silent template regression would quietly cut a child
11+
# tenant's log delivery (observed as fluent-bit "no upstream connections
12+
# available", with the drop invisible downstream because the packet never
13+
# leaves the source node). Pin it here alongside vminsert so neither block
14+
# can disappear or drift its selector unnoticed.
15+
#
16+
# The ancestor namespace is derived from the release namespace by a pure
17+
# splitList/join walk (no cluster lookup), so helm-unittest renders it
18+
# offline. Depth-2 child tenant-htdev in namespace tenant-ktj resolves its
19+
# single ancestor to tenant-ktj.
20+
templates:
21+
- templates/networkpolicy.yaml
22+
set:
23+
_cluster:
24+
root-host: example.com
25+
tests:
26+
- it: child tenant egress allows vlinsert and vminsert in the parent namespace
27+
release:
28+
name: tenant-htdev
29+
namespace: tenant-ktj
30+
asserts:
31+
- documentSelector:
32+
path: metadata.name
33+
value: tenant-ktj-htdev-egress
34+
contains:
35+
path: spec.egress
36+
content:
37+
toEndpoints:
38+
- matchLabels:
39+
"k8s:app.kubernetes.io/name": "vlinsert"
40+
"k8s:io.kubernetes.pod.namespace": tenant-ktj
41+
- documentSelector:
42+
path: metadata.name
43+
value: tenant-ktj-htdev-egress
44+
contains:
45+
path: spec.egress
46+
content:
47+
toEndpoints:
48+
- matchLabels:
49+
"k8s:app.kubernetes.io/name": "vminsert"
50+
"k8s:io.kubernetes.pod.namespace": tenant-ktj
51+
52+
- it: root tenant egress emits no ancestor ingest blocks
53+
# The `ne (include "tenant.name" .) "tenant-root"` guard must suppress
54+
# every ancestor-egress block for the root tenant, which has no ancestor
55+
# to reach. Its egress list carries only the self-namespace allow, so a
56+
# broken guard (which would emit vlinsert scoped to tenant-root) is
57+
# caught both by the explicit notContains and by the exact length pin.
58+
release:
59+
name: tenant-root
60+
namespace: tenant-root
61+
asserts:
62+
- documentSelector:
63+
path: metadata.name
64+
value: tenant-root-egress
65+
notContains:
66+
path: spec.egress
67+
content:
68+
toEndpoints:
69+
- matchLabels:
70+
"k8s:app.kubernetes.io/name": "vlinsert"
71+
"k8s:io.kubernetes.pod.namespace": tenant-root
72+
- documentSelector:
73+
path: metadata.name
74+
value: tenant-root-egress
75+
lengthEqual:
76+
path: spec.egress
77+
count: 1

0 commit comments

Comments
 (0)