kuma: preload Docker Hub images in E2E to avoid slow pulls

kuma/tests/conftest.py

@@ -2,10 +2,12 @@
 # All rights reserved
 # Licensed under a 3-clause BSD style license (see LICENSE)
 import os
+import platform
+from contextlib import contextmanager
 
 import pytest
 
-from datadog_checks.dev.kind import kind_run
+from datadog_checks.dev.kind import ClusterConfig, kind_run
 from datadog_checks.dev.subprocess import run_command
 
 KUMA_NAMESPACE = 'kuma-system'
@@ -14,6 +16,68 @@
 KUMA_METRICS_ENDPOINT = f'http://{KUMA_SERVICE}.{KUMA_NAMESPACE}.svc.cluster.local:5680/metrics'
 
 
+def _kuma_images():
+    # Images the Kuma Helm chart pulls from Docker Hub during a fresh install
+    # (the control-plane deployment and the CRD-install hook job). Preloading
+    # them into the kind nodes avoids slow or rate-limited pulls from docker.io
+    # on constrained connections; the chart defaults to `imagePullPolicy:
+    # IfNotPresent`, so the side-loaded copies are used. The `kubectl` hook
+    # image is pulled from registry.k8s.io (fast) and left to pull normally.
+    kuma_version = os.environ.get('KUMA_VERSION', '2.10.6')
+    return [
+        f'kumahq/kuma-cp:{kuma_version}',
+        f'kumahq/kumactl:{kuma_version}',
+    ]
+
+
+def _platform_spec():
+    # `docker save --platform` only accepts a single os/arch pair; default to
+    # the host platform so the saved archive contains exactly the blobs that
+    # were pulled (and no cross-platform references that break `kind load`).
+    return f'{platform.system().lower()}/{platform.machine()}'
+
+
+@contextmanager
+def _pull_images(images):
+    for image in images:
+        run_command(['docker', 'pull', '--platform', _platform_spec(), image], check=True)
+    yield
+
+
+class _PreloadImages:
+    """Load the Kuma images into the kind cluster before `setup_kuma` runs.
+
+    Uses `docker save --platform` + `kind load image-archive` rather than
+    `kind load docker-image`: on Docker daemons using the containerd image
+    store, `docker save` of a multi-arch image emits an OCI index referencing
+    blobs for every platform, and `kind load docker-image` (which pipes that to
+    `ctr images import --all-platforms`) then fails with "content digest ...
+    not found" for the non-host platforms. Saving a single-platform archive
+    sidesteps this and keeps the original image name so the chart's image
+    references resolve to the side-loaded copies.
+    """
+
+    def __init__(self, images):
+        self.images = images
+        self.cluster_name = None
+
+    def add_cluster_info(self, cluster_config: ClusterConfig):
+        self.cluster_name = cluster_config.cluster_name
+
+    def __call__(self):
+        if self.cluster_name is None:
+            raise RuntimeError('cluster_name must be set before preloading images')
+
+        spec = _platform_spec()
+        for image in self.images:
+            archive = f'/tmp/{image.replace("/", "_").replace(":", "_")}.tar'
+            run_command(['docker', 'save', '--platform', spec, image, '-o', archive], check=True)
+            run_command(
+                ['kind', 'load', 'image-archive', archive, '--name', self.cluster_name],
+                check=True,
+            )
+
+
 def setup_kuma():
     kuma_version = os.environ.get('KUMA_VERSION', '2.10.6')
     run_command(['kubectl', 'create', 'namespace', KUMA_NAMESPACE], check=True)
@@ -51,7 +115,11 @@ def setup_kuma():
 
 @pytest.fixture(scope='session')
 def dd_environment(dd_save_state):
-    with kind_run(conditions=[setup_kuma]) as kubeconfig:
+    images = _kuma_images()
+    with kind_run(
+        wrappers=[_pull_images(images)],
+        conditions=[_PreloadImages(images), setup_kuma],
+    ) as kubeconfig:
         instance = {'openmetrics_endpoint': KUMA_METRICS_ENDPOINT}
         metadata = {
             'agent_type': 'kubernetes',