github/kubernetes
1
0
Fork 0
mirror of https://github.com/k3s-io/kubernetes.git synced 2025-08-06 10:43:56 +00:00
Code Issues Projects Releases Wiki Activity

node: device-plugin: e2e: Additional test cases

Browse Source 
Additional test cases added:
Keeps device plugin assignments across pod and kubelet restarts (no device plugin re-registration)
Keeps device plugin assignments after the device plugin has re-registered (no kubelet or pod restart)

Signed-off-by: Swati Sehgal <swsehgal@redhat.com>
This commit is contained in:
Swati Sehgal 2023-04-27 11:26:50 +01:00
parent 3dbb741c97
commit d727df1741
1 changed files with 119 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

119
test/e2e_node/device_plugin_test.go
View File

@ -340,6 +340,125 @@ func testDevicePlugin(f *framework.Framework, pluginSockDir string) {
framework.ExpectNoError(err, "inconsistent device assignment after pod restart")
})
// simulate kubelet and container restart, *but not* device plugin re-registration.
// The device assignment should be kept and be stable across the kubelet and container restart, because it's the kubelet which
// performs the device allocation, and both the device plugin is stable.
ginkgo.It("Keeps device plugin assignments across pod and kubelet restarts (no device plugin re-registration)", func(ctx context.Context) {
podRECMD := fmt.Sprintf("devs=$(ls /tmp/ | egrep '^Dev-[0-9]+$') && echo stub devices: $devs && sleep %s", sleepIntervalWithRestart)
pod1 := e2epod.NewPodClient(f).CreateSync(ctx, makeBusyboxPod(SampleDeviceResourceName, podRECMD))
deviceIDRE := "stub devices: (Dev-[0-9]+)"
devID1, err := parseLog(ctx, f, pod1.Name, pod1.Name, deviceIDRE)
framework.ExpectNoError(err, "getting logs for pod %q", pod1.Name)
gomega.Expect(devID1).To(gomega.Not(gomega.Equal("")), "pod1 requested a device but started successfully without")
pod1, err = e2epod.NewPodClient(f).Get(ctx, pod1.Name, metav1.GetOptions{})
framework.ExpectNoError(err)
ginkgo.By("Wait for node to be ready again")
e2enode.WaitForAllNodesSchedulable(ctx, f.ClientSet, 5*time.Minute)
ginkgo.By("Waiting for container to restart")
ensurePodContainerRestart(ctx, f, pod1.Name, pod1.Name)
ginkgo.By("Confirming that after a container restart, fake-device assignment is kept")
devIDRestart1, err := parseLog(ctx, f, pod1.Name, pod1.Name, deviceIDRE)
framework.ExpectNoError(err, "getting logs for pod %q", pod1.Name)
framework.ExpectEqual(devIDRestart1, devID1)
ginkgo.By("Restarting Kubelet")
restartKubelet(true)
ginkgo.By("Wait for node to be ready again")
e2enode.WaitForAllNodesSchedulable(ctx, f.ClientSet, 5*time.Minute)
ginkgo.By("Waiting for the pod to fail with admission error as device plugin hasn't re-registered yet")
gomega.Eventually(ctx, getPod).
WithArguments(f, pod1.Name).
WithTimeout(time.Minute).
Should(HaveFailedWithAdmissionError(),
"the pod succeeded to start, when it should fail with the admission error")
// crosscheck from the device assignment is preserved and stable from perspective of the kubelet.
// note we don't check again the logs of the container: the check is done at startup, the container
// never restarted (runs "forever" from this test timescale perspective) hence re-doing this check
// is useless.
ginkgo.By("Verifying the device assignment after kubelet restart using podresources API")
gomega.Eventually(ctx, func() error {
v1PodResources, err = getV1NodeDevices(ctx)
return err
}, 30*time.Second, framework.Poll).ShouldNot(gomega.HaveOccurred(), "cannot fetch the compute resource assignment after kubelet restart")
err = checkPodResourcesAssignment(v1PodResources, pod1.Namespace, pod1.Name, pod1.Spec.Containers[0].Name, SampleDeviceResourceName, []string{devID1})
framework.ExpectNoError(err, "inconsistent device assignment after pod restart")
})
// simulate device plugin re-registration, *but not* container and kubelet restart.
// After the device plugin has re-registered, the list healthy devices is repopulated based on the devices discovered.
// Once Pod2 is running we determine the device that was allocated it. As long as the device allocation succeeds the
// test should pass.
ginkgo.It("Keeps device plugin assignments after the device plugin has been re-registered (no kubelet, pod restart)", func(ctx context.Context) {
podRECMD := fmt.Sprintf("devs=$(ls /tmp/ | egrep '^Dev-[0-9]+$') && echo stub devices: $devs && sleep %s", sleepIntervalForever)
pod1 := e2epod.NewPodClient(f).CreateSync(ctx, makeBusyboxPod(SampleDeviceResourceName, podRECMD))
deviceIDRE := "stub devices: (Dev-[0-9]+)"
devID1, err := parseLog(ctx, f, pod1.Name, pod1.Name, deviceIDRE)
framework.ExpectNoError(err, "getting logs for pod %q", pod1.Name)
gomega.Expect(devID1).To(gomega.Not(gomega.Equal("")), "pod1 requested a device but started successfully without")
pod1, err = e2epod.NewPodClient(f).Get(ctx, pod1.Name, metav1.GetOptions{})
framework.ExpectNoError(err)
ginkgo.By("Wait for node to be ready again")
e2enode.WaitForAllNodesSchedulable(ctx, f.ClientSet, 5*time.Minute)
ginkgo.By("Re-Register resources and delete the plugin pod")
gp := int64(0)
deleteOptions := metav1.DeleteOptions{
GracePeriodSeconds: &gp,
}
e2epod.NewPodClient(f).DeleteSync(ctx, devicePluginPod.Name, deleteOptions, time.Minute)
waitForContainerRemoval(ctx, devicePluginPod.Spec.Containers[0].Name, devicePluginPod.Name, devicePluginPod.Namespace)
ginkgo.By("Recreating the plugin pod")
devicePluginPod = e2epod.NewPodClient(f).CreateSync(ctx, dptemplate)
err = e2epod.WaitTimeoutForPodRunningInNamespace(ctx, f.ClientSet, devicePluginPod.Name, devicePluginPod.Namespace, 1*time.Minute)
framework.ExpectNoError(err)
ginkgo.By("Waiting for resource to become available on the local node after re-registration")
gomega.Eventually(ctx, func() bool {
node, ready := getLocalTestNode(ctx, f)
return ready &&
CountSampleDeviceCapacity(node) == expectedSampleDevsAmount &&
CountSampleDeviceAllocatable(node) == expectedSampleDevsAmount
}, 30*time.Second, framework.Poll).Should(gomega.BeTrue())
// crosscheck that after device plugin re-registration the device assignment is preserved and
// stable from the kubelet's perspective.
// note we don't check again the logs of the container: the check is done at startup, the container
// never restarted (runs "forever" from this test timescale perspective) hence re-doing this check
// is useless.
ginkgo.By("Verifying the device assignment after device plugin re-registration using podresources API")
gomega.Eventually(ctx, func() error {
v1PodResources, err = getV1NodeDevices(ctx)
return err
}, 30*time.Second, framework.Poll).ShouldNot(gomega.HaveOccurred(), "cannot fetch the compute resource assignment after kubelet restart")
err = checkPodResourcesAssignment(v1PodResources, pod1.Namespace, pod1.Name, pod1.Spec.Containers[0].Name, SampleDeviceResourceName, []string{devID1})
framework.ExpectNoError(err, "inconsistent device assignment after pod restart")
ginkgo.By("Creating another pod")
pod2 := e2epod.NewPodClient(f).CreateSync(ctx, makeBusyboxPod(SampleDeviceResourceName, podRECMD))
err = e2epod.WaitTimeoutForPodRunningInNamespace(ctx, f.ClientSet, pod2.Name, f.Namespace.Name, 1*time.Minute)
framework.ExpectNoError(err)
ginkgo.By("Checking that pod got a fake device")
devID2, err := parseLog(ctx, f, pod2.Name, pod2.Name, deviceIDRE)
framework.ExpectNoError(err, "getting logs for pod %q", pod2.Name)
gomega.Expect(devID2).To(gomega.Not(gomega.Equal("")), "pod2 requested a device but started successfully without")
})
// simulate kubelet restart *and* device plugin re-registration, while the pod and the container stays running.
// The device assignment should be kept and be stable across the kubelet/device plugin restart, as both the aforementioned components
// orchestrate the device allocation: the actual consumer (container) is stable.