Merge pull request #569 from element-hq/gaelg/matrix-rtc-metrics

gaelgatelement · web-flow · commit bfc09288928e · 2025-06-30T15:55:42.000+02:00
matrix-rtc: fix services monitors
diff --git a/charts/matrix-stack/ci/fragments/matrix-rtc-pytest-extras.yaml b/charts/matrix-stack/ci/fragments/matrix-rtc-pytest-extras.yaml
@@ -4,6 +4,8 @@
 
 matrixRTC:
   replicas: 2
+  annotations:
+    has-no-service-monitor: "true"
   ingress:
     host: mrtc.{{ $.Values.serverName }}
     tlsSecret: "{{ $.Release.Name }}-matrix-rtc-tls"
diff --git a/charts/matrix-stack/ci/pytest-matrix-rtc-standalone-values.yaml b/charts/matrix-stack/ci/pytest-matrix-rtc-standalone-values.yaml
@@ -16,6 +16,8 @@ initSecrets:
 matrixAuthenticationService:
   enabled: false
 matrixRTC:
+  annotations:
+    has-no-service-monitor: "true"
   extraEnv:
     - name: LIVEKIT_INSECURE_SKIP_VERIFY_TLS
       value: YES_I_KNOW_WHAT_I_AM_DOING
diff --git a/charts/matrix-stack/ci/pytest-matrix-rtc-synapse-wellknown-values.yaml b/charts/matrix-stack/ci/pytest-matrix-rtc-synapse-wellknown-values.yaml
@@ -28,6 +28,8 @@ initSecrets:
 matrixAuthenticationService:
   enabled: false
 matrixRTC:
+  annotations:
+    has-no-service-monitor: "true"
   extraEnv:
     - name: LIVEKIT_INSECURE_SKIP_VERIFY_TLS
       value: YES_I_KNOW_WHAT_I_AM_DOING
diff --git a/charts/matrix-stack/templates/matrix-rtc/authorisation_service_monitor.yaml b/charts/matrix-stack/templates/matrix-rtc/authorisation_service_monitor.yaml
diff --git a/charts/matrix-stack/templates/matrix-rtc/sfu_service.yaml b/charts/matrix-stack/templates/matrix-rtc/sfu_service.yaml
@@ -1,5 +1,5 @@
 {{- /*
-Copyright 2024 New Vector Ltd
+Copyright 2024-2025 New Vector Ltd
 
 SPDX-License-Identifier: AGPL-3.0-only
 */ -}}
@@ -21,6 +21,9 @@ spec:
   - name: http
     port: 7880
     targetPort: http
+  - name: metrics
+    port: 6789
+    targetPort: metrics
   selector:
     app.kubernetes.io/instance: "{{ $.Release.Name }}-matrix-rtc-sfu"
 {{- end -}}
diff --git a/charts/matrix-stack/templates/matrix-rtc/sfu_service_monitor.yaml b/charts/matrix-stack/templates/matrix-rtc/sfu_service_monitor.yaml
@@ -1,5 +1,5 @@
 {{- /*
-Copyright 2024 New Vector Ltd
+Copyright 2024-2025 New Vector Ltd
 
 SPDX-License-Identifier: AGPL-3.0-only
 */ -}}
@@ -20,7 +20,7 @@ metadata:
 spec:
   endpoints:
   - interval: 30s
-    port: http
+    port: metrics
   selector:
     matchLabels:
       app.kubernetes.io/part-of: matrix-stack
diff --git a/newsfragments/569.fixed.md b/newsfragments/569.fixed.md
@@ -0,0 +1 @@
+matrix-rtc: Fix service monitors deployed.
diff --git a/poetry.lock b/poetry.lock
diff --git a/pyproject.toml b/pyproject.toml
@@ -37,6 +37,7 @@ types-pytz = "^2025.2.0.20250516"
 mypy = "^1.16.1"
 types-pyyaml = "^6.0.12.20250516"
 semver = "^3.0.4"
+prometheus-client = "^0.22.1"
 
 [build-system]
 requires = ["poetry-core"]
diff --git a/tests/integration/lib/helpers.py b/tests/integration/lib/helpers.py
@@ -6,8 +6,16 @@
 import time
 
 import pyhelm3
+from lightkube import AsyncClient
+from lightkube.models.core_v1 import (
+    Capabilities,
+    Container,
+    PodSpec,
+    SeccompProfile,
+    SecurityContext,
+)
 from lightkube.models.meta_v1 import ObjectMeta
-from lightkube.resources.core_v1 import ConfigMap, Endpoints, Namespace, Secret
+from lightkube.resources.core_v1 import ConfigMap, Endpoints, Namespace, Pod, Secret
 
 from ..artifacts import CertKey, generate_cert
 from .utils import merge
@@ -98,3 +106,63 @@ async def get_deployment_marker(kube_client, generated_data, marker: str):
         name=f"{generated_data.release_name}-markers",
     )
     return configmap.data.get(marker)
+
+
+async def run_pod_with_args(kube_client: AsyncClient, namespace, image_name, pod_name, args):
+    pod = Pod(
+        metadata=ObjectMeta(name=pod_name + "-" + str(int(time.time() * 1000)), namespace=namespace),
+        spec=PodSpec(
+            restartPolicy="Never",
+            containers=[
+                Container(
+                    name="cmd",
+                    image=image_name,
+                    args=args,
+                    securityContext=SecurityContext(
+                        seccompProfile=SeccompProfile(type="RuntimeDefault"),
+                        capabilities=Capabilities(drop=["ALL"]),
+                        readOnlyRootFilesystem=True,
+                        allowPrivilegeEscalation=False,
+                        runAsNonRoot=True,
+                        runAsUser=3000,
+                        runAsGroup=3000,
+                    ),
+                )
+            ],
+        ),
+    )
+    assert pod.metadata
+    assert pod.metadata.name
+    assert pod.metadata.namespace
+    try:
+        await kube_client.create(pod)
+        start_time = time.time()
+        now = time.time()
+        completed = False
+        while start_time + 30 > now and not completed:
+            found_pod = await kube_client.get(Pod, name=pod.metadata.name, namespace=pod.metadata.namespace)
+            if (
+                found_pod.status
+                and found_pod.status.containerStatuses
+                and found_pod.status.containerStatuses[0].lastState
+                and found_pod.status.containerStatuses[0].lastState.terminated
+                and found_pod.status.containerStatuses[0].lastState.terminated.reason == "Completed"
+            ):
+                completed = True
+            else:
+                now = time.time()
+                await asyncio.sleep(1)
+        else:
+            if start_time + 30 > now:
+                raise RuntimeError(
+                    f"Pod {pod.metadata.name} did not start in time "
+                    f"(failed after {time.time() - now} seconds), "
+                    f"pod status: {found_pod.status}"
+                )
+
+        log_lines = ""
+        async for log_line in kube_client.log(pod.metadata.name, namespace=pod.metadata.namespace, container="cmd"):
+            log_lines += log_line
+        return log_lines
+    finally:
+        await kube_client.delete(Pod, name=pod.metadata.name, namespace=namespace)
diff --git a/tests/integration/test_networking.py b/tests/integration/test_networking.py
@@ -9,9 +9,10 @@
 from lightkube import AsyncClient
 from lightkube import operators as op
 from lightkube.resources.core_v1 import Pod, Service
+from prometheus_client.parser import text_string_to_metric_families
 
 from .fixtures.data import ESSData
-from .lib.helpers import wait_for_endpoint_ready
+from .lib.helpers import run_pod_with_args, wait_for_endpoint_ready
 from .lib.utils import read_service_monitor_kind
 
 
@@ -139,9 +140,6 @@ async def test_pods_monitored(
                 service_port_names = [port.name for port in service.spec.ports if port.name]
                 if endpoint["port"] in service_port_names:
                     break
-            # This Service does not have the named port. Potentially there's another Service that covers it
-            else:
-                continue
 
             async for covered_pod in kube_client.list(
                 Pod, namespace=generated_data.ess_namespace, labels=service.spec.selector
@@ -162,3 +160,65 @@ async def test_pods_monitored(
     assert all_monitorable_pods == monitored_pods, (
         f"Some pods are not monitored : {', '.join(list(set(all_monitorable_pods) ^ set(monitored_pods)))}"
     )
+
+
+@pytest.mark.skipif(
+    os.environ.get("SKIP_SERVICE_MONITORS_CRDS", "false") == "true", reason="ServiceMonitors not deployed"
+)
+@pytest.mark.skipif(os.environ.get("MATRIX_TEST_FROM_REF", "") == "25.6.2", reason="This fails against 25.6.2.")
+@pytest.mark.asyncio_cooperative
+@pytest.mark.usefixtures("matrix_stack")
+async def test_service_monitors_point_to_metrics(
+    kube_client: AsyncClient,
+    generated_data: ESSData,
+):
+    async for service_monitor in kube_client.list(
+        await read_service_monitor_kind(kube_client),
+        namespace=generated_data.ess_namespace,
+        labels={"app.kubernetes.io/part-of": op.in_(["matrix-stack"])},
+    ):
+        found_metrics = False
+        async for service in kube_client.list(
+            Service, namespace=generated_data.ess_namespace, labels=service_monitor["spec"]["selector"]["matchLabels"]
+        ):
+            assert service.metadata, f"Encountered a service without metadata : {service}"
+            assert service.spec, f"Encountered a service without spec : {service}"
+            assert service.spec.ports, f"Ecountered a service without port : {service}"
+            assert service.spec.selector, f"Ecountered a service without selectors : {service}"
+
+            for endpoint in service_monitor["spec"]["endpoints"]:
+                service_port_names = [port.name for port in service.spec.ports if port.name]
+                if endpoint["port"] in service_port_names:
+                    assert await has_actual_metrics_on_endpoint(
+                        kube_client, generated_data, service, service_monitor["spec"]["endpoints"]
+                    )
+                    found_metrics = True
+        assert found_metrics, (
+            f"ServiceMonitor {service_monitor['metadata']['name']} does not point to any /metrics endpoint"
+        )
+
+
+async def has_actual_metrics_on_endpoint(
+    kube_client: AsyncClient, generated_data: ESSData, service: Service, endpoints
+):
+    assert service.metadata
+    assert service.spec
+    assert service.spec.ports
+    found_metrics = False
+    for endpoint in endpoints:
+        for port_spec in service.spec.ports:
+            if port_spec.name == endpoint["port"]:
+                metrics_data = await run_pod_with_args(
+                    kube_client,
+                    generated_data.ess_namespace,
+                    "curlimages/curl:latest",
+                    "curl",
+                    [
+                        "-s",
+                        f"http://{service.metadata.name}.{generated_data.ess_namespace}.svc.cluster.local:{port_spec.port}/metrics",
+                    ],
+                )
+                for metric_family in text_string_to_metric_families(metrics_data):
+                    assert metric_family.name, "Metric family has no name"
+                found_metrics = True
+    return found_metrics
diff --git a/tests/manifests/__init__.py b/tests/manifests/__init__.py
@@ -468,6 +468,7 @@ def make_synapse_worker_sub_component(worker_name: str, worker_type: str) -> Sub
         name="matrix-rtc",
         values_file_path=ValuesFilePath.read_write("matrixRTC"),
         has_topology_spread_constraints=False,
+        has_service_monitor=False,
         sub_components=(
             SubComponentDetails(
                 name="matrix-rtc-sfu",

Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`1`	`+matrix-rtc: Fix service monitors deployed.`