Ensure pods both running and ready before starting e2e tests

hack/e2e-internal/build-release.sh

@@ -21,7 +21,7 @@ set -o pipefail
 KUBE_ROOT=$(dirname "${BASH_SOURCE}")/../..
 
 : ${KUBE_VERSION_ROOT:=${KUBE_ROOT}}
-: ${KUBECTL:="${KUBE_VERSION_ROOT}/cluster/kubectl.sh"}
+: ${KUBECTL:=${KUBE_VERSION_ROOT}/cluster/kubectl.sh}
 : ${KUBE_CONFIG_FILE:="config-test.sh"}
 
 export KUBECTL KUBE_CONFIG_FILE

hack/e2e-internal/e2e-cluster-size.sh

@@ -21,7 +21,7 @@ set -o pipefail
 KUBE_ROOT=$(dirname "${BASH_SOURCE}")/../..
 
 : ${KUBE_VERSION_ROOT:=${KUBE_ROOT}}
-: ${KUBECTL:="${KUBE_VERSION_ROOT}/cluster/kubectl.sh"}
+: ${KUBECTL:=${KUBE_VERSION_ROOT}/cluster/kubectl.sh}
 : ${KUBE_CONFIG_FILE:="config-test.sh"}
 
 export KUBECTL KUBE_CONFIG_FILE

hack/e2e-internal/e2e-down.sh

@@ -21,7 +21,7 @@ set -o pipefail
 KUBE_ROOT=$(dirname "${BASH_SOURCE}")/../..
 
 : ${KUBE_VERSION_ROOT:=${KUBE_ROOT}}
-: ${KUBECTL:="${KUBE_VERSION_ROOT}/cluster/kubectl.sh"}
+: ${KUBECTL:=${KUBE_VERSION_ROOT}/cluster/kubectl.sh}
 : ${KUBE_CONFIG_FILE:="config-test.sh"}
 
 export KUBECTL KUBE_CONFIG_FILE

hack/e2e-internal/e2e-push.sh

@@ -21,7 +21,7 @@ set -o pipefail
 KUBE_ROOT=$(dirname "${BASH_SOURCE}")/../..
 
 : ${KUBE_VERSION_ROOT:=${KUBE_ROOT}}
-: ${KUBECTL:="${KUBE_VERSION_ROOT}/cluster/kubectl.sh"}
+: ${KUBECTL:=${KUBE_VERSION_ROOT}/cluster/kubectl.sh}
 : ${KUBE_CONFIG_FILE:="config-test.sh"}
 
 export KUBECTL KUBE_CONFIG_FILE

hack/e2e-internal/e2e-status.sh

@@ -14,8 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# e2e-status checks that the status of a cluster is acceptable for running
-# e2e tests.
 set -o errexit
 set -o nounset
 set -o pipefail
@@ -23,7 +21,7 @@ set -o pipefail
 KUBE_ROOT=$(dirname "${BASH_SOURCE}")/../..
 
 : ${KUBE_VERSION_ROOT:=${KUBE_ROOT}}
-: ${KUBECTL:="${KUBE_VERSION_ROOT}/cluster/kubectl.sh"}
+: ${KUBECTL:=${KUBE_VERSION_ROOT}/cluster/kubectl.sh}
 : ${KUBE_CONFIG_FILE:="config-test.sh"}
 
 export KUBECTL KUBE_CONFIG_FILE
@@ -34,49 +32,3 @@ source "${KUBE_VERSION_ROOT}/cluster/${KUBERNETES_PROVIDER}/util.sh"
 prepare-e2e
 
 ${KUBECTL} version
-
-# Before running tests, ensure that all pods are 'Running'. Tests can timeout
-# and fail because the test pods don't run in time. The problem is that the pods
-# that a cluster runs on startup take too long to start running, with sequential
-# Docker pulls of large images being the culprit. These startup pods block the
-# test pods from running.
-
-# Settings:
-# timeout is in seconds; 1200 = 20 minutes.
-timeout=1200
-# pause is how many seconds to sleep between pod get calls.
-pause=5
-# min_pods is the minimum number of pods we require.
-min_pods=1
-
-# Check pod statuses.
-deadline=$(($(date '+%s')+${timeout}))
-echo "Waiting at most ${timeout} seconds for all pods to be 'Running'" >&2
-all_running=0
-until [[ ${all_running} == 1 ]]; do
-  if [[ "$(date '+%s')" -ge "${deadline}" ]]; then
-    echo "All pods never 'Running' in time." >&2
-    exit 1
-  fi
-  statuses=($(${KUBECTL} get pods --template='{{range.items}}{{.status.phase}} {{end}}' --api-version=v1beta3))
-
-  # Ensure that we have enough pods.
-  echo "Found ${#statuses[@]} pods with statuses: ${statuses[@]}" >&2
-  if [[ ${#statuses[@]} -lt ${min_pods} ]]; then
-    continue
-  fi
-
-  # Then, ensure all pods found are 'Running'.
-  found_running=1
-  for status in "${statuses[@]}"; do
-    if [[ "${status}" != "Running" ]]; then
-      # If we find a pod that isn't 'Running', sleep here to avoid delaying
-      # other code paths (where all pods are 'Running').
-      found_running=0
-      sleep ${pause}
-      break
-    fi
-  done
-  all_running=${found_running}
-done
-echo "All pods are 'Running'" >&2

hack/e2e-internal/e2e-up.sh

@@ -21,7 +21,7 @@ set -o pipefail
 KUBE_ROOT=$(dirname "${BASH_SOURCE}")/../..
 
 : ${KUBE_VERSION_ROOT:=${KUBE_ROOT}}
-: ${KUBECTL:="${KUBE_VERSION_ROOT}/cluster/kubectl.sh"}
+: ${KUBECTL:=${KUBE_VERSION_ROOT}/cluster/kubectl.sh}
 : ${KUBE_CONFIG_FILE:="config-test.sh"}
 
 export KUBECTL KUBE_CONFIG_FILE

test/e2e/e2e_test.go

@@ -22,7 +22,9 @@ import (
 	"path"
 	"strings"
 	"testing"
+	"time"
 
+	"github.com/GoogleCloudPlatform/kubernetes/pkg/api"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/client/clientcmd"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/cloudprovider"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/util"
@@ -33,6 +35,28 @@ import (
 	"github.com/onsi/gomega"
 )
 
+const (
+	// podStartupTimeout is the time to allow all pods in the cluster to become
+	// running and ready before any e2e tests run. It includes pulling all of
+	// the pods (as of 5/18/15 this is 8 pods).
+	podStartupTimeout = 10 * time.Minute
+
+	// minStartupPods is the minimum number of pods that will allow
+	// wiatForPodsRunningReady(...) to succeed. More verbosely, that function
+	// checks that all pods in the cluster are both in a phase of "running" and
+	// have a condition of "ready": "true". It aims to ensure that the cluster's
+	// pods are fully healthy before beginning e2e tests. However, if there were
+	// only 0 pods, it would technically pass if there wasn't a required minimum
+	// number of pods. We expect every cluster to come up with some number of
+	// pods (which in practice is more than this number), so we have this
+	// minimum here as a sanity check to make sure that there are actually pods
+	// on the cluster (i.e. preventing a possible race with kube-addons). This
+	// does *not* mean that the function will succeed as soon as minStartupPods
+	// are found to be running and ready; it ensures that *all* pods it finds
+	// are running and ready. This is the minimum number it must find.
+	minStartupPods = 1
+)
+
 var (
 	cloudConfig = &testContext.CloudConfig
 
@@ -92,6 +116,15 @@ func TestE2E(t *testing.T) {
 	}
 
 	gomega.RegisterFailHandler(ginkgo.Fail)
+
+	// Ensure all pods are running and ready before starting tests (otherwise,
+	// cluster infrastructure pods that are being pulled or started can block
+	// test pods from running, and tests that ensure all pods are running and
+	// ready will fail).
+	if err := waitForPodsRunningReady(api.NamespaceDefault, minStartupPods, podStartupTimeout); err != nil {
+		glog.Fatalf("Error waiting for all pods to be running and ready: %v", err)
+	}
+
 	// Run tests through the Ginkgo runner with output to console + JUnit for Jenkins
 	var r []ginkgo.Reporter
 	if *reportDir != "" {

test/e2e/reboot.go

@@ -191,25 +191,6 @@ func rebootNode(c *client.Client, provider, name string, result chan bool) {
 	result <- true
 }
 
-// podRunningReady is the checker function passed to waitForPodCondition(...)
-// (found in util.go). It ensures that the pods' phase is running and that the
-// ready condition is true.
-func podRunningReady(p *api.Pod) (bool, error) {
-	// Check the phase is running.
-	if p.Status.Phase != api.PodRunning {
-		return false, fmt.Errorf("want pod %s on %s to be %v but was %v",
-			p.ObjectMeta.Name, p.Spec.Host, api.PodRunning, p.Status.Phase)
-	}
-	// Check the ready condition is true.
-	for _, cond := range p.Status.Conditions {
-		if cond.Type == api.PodReady && cond.Status == api.ConditionTrue {
-			return true, nil
-		}
-	}
-	return false, fmt.Errorf("pod %s on %s didn't have condition %v, %v; conditions: %v",
-		p.ObjectMeta.Name, p.Spec.Host, api.PodReady, api.ConditionTrue, p.Status.Conditions)
-}
-
 // checkPodsRunning returns whether all pods whose names are listed in podNames
 // are running.
 func checkPodsRunning(c *client.Client, podNames []string, timeout time.Duration) bool {

test/e2e/util.go

@@ -106,6 +106,83 @@ func providerIs(providers ...string) bool {
 
 type podCondition func(pod *api.Pod) (bool, error)
 
+// podReady returns whether pod has a condition of Ready with a status of true.
+func podReady(pod *api.Pod) bool {
+	for _, cond := range pod.Status.Conditions {
+		if cond.Type == api.PodReady && cond.Status == api.ConditionTrue {
+			return true
+		}
+	}
+	return false
+}
+
+// logPodStates logs all pod states for debugging.
+func logPodStates(c *client.Client, ns string) {
+	podList, err := c.Pods(ns).List(labels.Everything(), fields.Everything())
+	if err != nil {
+		Logf("Error getting pods for logPodStates(...): %v", err)
+		return
+	}
+	Logf("Phase and conditions for all pods in namespace '%s':", ns)
+	for _, pod := range podList.Items {
+		Logf("- pod '%s' on '%s' has phase '%v' and conditions %v",
+			pod.ObjectMeta.Name, pod.Spec.Host, pod.Status.Phase, pod.Status.Conditions)
+	}
+}
+
+// podRunningReady checks whether pod p's phase is running and it has a ready
+// condition of status true.
+func podRunningReady(p *api.Pod) (bool, error) {
+	// Check the phase is running.
+	if p.Status.Phase != api.PodRunning {
+		return false, fmt.Errorf("want pod '%s' on '%s' to be '%v' but was '%v'",
+			p.ObjectMeta.Name, p.Spec.Host, api.PodRunning, p.Status.Phase)
+	}
+	// Check the ready condition is true.
+	if !podReady(p) {
+		return false, fmt.Errorf("pod '%s' on '%s' didn't have condition {%v %v}; conditions: %v",
+			p.ObjectMeta.Name, p.Spec.Host, api.PodReady, api.ConditionTrue, p.Status.Conditions)
+
+	}
+	return true, nil
+}
+
+// waitForPodsRunningReady waits up to timeout to ensure that all pods in
+// namespace ns are running and ready, requiring that it finds at least minPods.
+// It has separate behavior from other 'wait for' pods functions in that it re-
+// queries the list of pods on every iteration. This is useful, for example, in
+// cluster startup, because the number of pods increases while waiting.
+func waitForPodsRunningReady(ns string, minPods int, timeout time.Duration) error {
+	c, err := loadClient()
+	if err != nil {
+		return err
+	}
+	Logf("Waiting up to %v for all pods (need at least %d) in namespace '%s' to be running and ready",
+		timeout, minPods, ns)
+	for start := time.Now(); time.Since(start) < timeout; time.Sleep(podPoll) {
+		// We get the new list of pods in every iteration beause more pods come
+		// online during startup and we want to ensure they are also checked.
+		podList, err := c.Pods(ns).List(labels.Everything(), fields.Everything())
+		if err != nil {
+			Logf("Error getting pods in namespace '%s': %v", ns, err)
+			continue
+		}
+		nOk := 0
+		for _, pod := range podList.Items {
+			if res, err := podRunningReady(&pod); res && err == nil {
+				nOk++
+			}
+		}
+		Logf("%d / %d pods in namespace '%s' are running and ready (%v elapsed)",
+			nOk, len(podList.Items), ns, time.Since(start))
+		if nOk == len(podList.Items) && nOk >= minPods {
+			return nil
+		}
+	}
+	logPodStates(c, ns)
+	return fmt.Errorf("Not all pods in namespace '%s' running and ready within %v", ns, timeout)
+}
+
 func waitForPodCondition(c *client.Client, ns, podName, desc string, poll, timeout time.Duration, condition podCondition) error {
 	Logf("Waiting up to %v for pod %s status to be %s", timeout, podName, desc)
 	for start := time.Now(); time.Since(start) < timeout; time.Sleep(poll) {
@@ -118,9 +195,10 @@ func waitForPodCondition(c *client.Client, ns, podName, desc string, poll, timeo
 		if done {
 			return err
 		}
-		Logf("Waiting for pod %s in namespace %s status to be %q (found %q) (%v)", podName, ns, desc, pod.Status.Phase, time.Since(start))
+		Logf("Waiting for pod '%s' in namespace '%s' status to be '%q' (found phase: '%q', readiness: %t) (%v)",
+			podName, ns, desc, pod.Status.Phase, podReady(pod), time.Since(start))
 	}
-	return fmt.Errorf("gave up waiting for pod %s to be %s after %v", podName, desc, timeout)
+	return fmt.Errorf("gave up waiting for pod '%s' to be '%s' after %v", podName, desc, timeout)
 }
 
 // createNS should be used by every test, note that we append a common prefix to the provided test name.
@@ -150,7 +228,7 @@ func waitForPodRunning(c *client.Client, podName string) error {
 func waitForPodNotPending(c *client.Client, ns, podName string) error {
 	return waitForPodCondition(c, ns, podName, "!pending", podPoll, podStartTimeout, func(pod *api.Pod) (bool, error) {
 		if pod.Status.Phase != api.PodPending {
-			Logf("Saw pod %s in namespace %s out of pending state (found %q)", podName, ns, pod.Status.Phase)
+			Logf("Saw pod '%s' in namespace '%s' out of pending state (found '%q')", podName, ns, pod.Status.Phase)
 			return true, nil
 		}
 		return false, nil
@@ -163,17 +241,17 @@ func waitForPodSuccessInNamespace(c *client.Client, podName string, contName str
 		// Cannot use pod.Status.Phase == api.PodSucceeded/api.PodFailed due to #2632
 		ci, ok := api.GetContainerStatus(pod.Status.ContainerStatuses, contName)
 		if !ok {
-			Logf("No Status.Info for container %s in pod %s yet", contName, podName)
+			Logf("No Status.Info for container '%s' in pod '%s' yet", contName, podName)
 		} else {
 			if ci.State.Termination != nil {
 				if ci.State.Termination.ExitCode == 0 {
 					By("Saw pod success")
 					return true, nil
 				} else {
-					return true, fmt.Errorf("pod %s terminated with failure: %+v", podName, ci.State.Termination)
+					return true, fmt.Errorf("pod '%s' terminated with failure: %+v", podName, ci.State.Termination)
 				}
 			} else {
-				Logf("Nil State.Termination for container %s in pod %s in namespace %s so far", contName, podName, namespace)
+				Logf("Nil State.Termination for container '%s' in pod '%s' in namespace '%s' so far", contName, podName, namespace)
 			}
 		}
 		return false, nil