// Index returns a list of items that match on the index function
// Index is thread-safe so long as you treat all items as immutable
func (c *threadSafeMap) Index(indexName string, obj interface{}) ([]interface{}, error) {
c.lock.RLock()
defer c.lock.RUnlock()
indexFunc := c.indexers[indexName]
if indexFunc == nil {
return nil, fmt.Errorf("Index with name %s does not exist", indexName)
}
indexKeys, err := indexFunc(obj)
if err != nil {
return nil, err
}
index := c.indices[indexName]
// need to de-dupe the return list. Since multiple keys are allowed, this can happen.
returnKeySet := sets.String{}
for _, indexKey := range indexKeys {
set := index[indexKey]
for _, key := range set.UnsortedList() {
returnKeySet.Insert(key)
}
}
list := make([]interface{}, 0, returnKeySet.Len())
for absoluteKey := range returnKeySet {
list = append(list, c.items[absoluteKey])
}
return list, nil
}
func runResourceTrackingTest(f *framework.Framework, podsPerNode int, nodeNames sets.String, rm *framework.ResourceMonitor,
expectedCPU map[string]map[float64]float64, expectedMemory framework.ResourceUsagePerContainer) {
numNodes := nodeNames.Len()
totalPods := podsPerNode * numNodes
By(fmt.Sprintf("Creating a RC of %d pods and wait until all pods of this RC are running", totalPods))
rcName := fmt.Sprintf("resource%d-%s", totalPods, string(uuid.NewUUID()))
// TODO: Use a more realistic workload
Expect(framework.RunRC(testutils.RCConfig{
Client: f.ClientSet,
InternalClient: f.InternalClientset,
Name: rcName,
Namespace: f.Namespace.Name,
Image: framework.GetPauseImageName(f.ClientSet),
Replicas: totalPods,
})).NotTo(HaveOccurred())
// Log once and flush the stats.
rm.LogLatest()
rm.Reset()
By("Start monitoring resource usage")
// Periodically dump the cpu summary until the deadline is met.
// Note that without calling framework.ResourceMonitor.Reset(), the stats
// would occupy increasingly more memory. This should be fine
// for the current test duration, but we should reclaim the
// entries if we plan to monitor longer (e.g., 8 hours).
deadline := time.Now().Add(monitoringTime)
for time.Now().Before(deadline) {
timeLeft := deadline.Sub(time.Now())
framework.Logf("Still running...%v left", timeLeft)
if timeLeft < reportingPeriod {
time.Sleep(timeLeft)
} else {
time.Sleep(reportingPeriod)
}
logPodsOnNodes(f.ClientSet, nodeNames.List())
}
By("Reporting overall resource usage")
logPodsOnNodes(f.ClientSet, nodeNames.List())
usageSummary, err := rm.GetLatest()
Expect(err).NotTo(HaveOccurred())
// TODO(random-liu): Remove the original log when we migrate to new perfdash
framework.Logf("%s", rm.FormatResourceUsage(usageSummary))
// Log perf result
framework.PrintPerfData(framework.ResourceUsageToPerfData(rm.GetMasterNodeLatest(usageSummary)))
verifyMemoryLimits(f.ClientSet, expectedMemory, usageSummary)
cpuSummary := rm.GetCPUSummary()
framework.Logf("%s", rm.FormatCPUSummary(cpuSummary))
// Log perf result
framework.PrintPerfData(framework.CPUUsageToPerfData(rm.GetMasterNodeCPUSummary(cpuSummary)))
verifyCPULimits(expectedCPU, cpuSummary)
By("Deleting the RC")
framework.DeleteRCAndPods(f.ClientSet, f.InternalClientset, f.Namespace.Name, rcName)
}
// DialFromContainers executes a curl via kubectl exec in a test container,
// which might then translate to a tcp or udp request based on the protocol
// argument in the url.
// - minTries is the minimum number of curl attempts required before declaring
// success. Set to 0 if you'd like to return as soon as all endpoints respond
// at least once.
// - maxTries is the maximum number of curl attempts. If this many attempts pass
// and we don't see all expected endpoints, the test fails.
// - expectedEps is the set of endpointnames to wait for. Typically this is also
// the hostname reported by each pod in the service through /hostName.
// maxTries == minTries will confirm that we see the expected endpoints and no
// more for maxTries. Use this if you want to eg: fail a readiness check on a
// pod and confirm it doesn't show up as an endpoint.
func (config *NetworkingTestConfig) DialFromContainer(protocol, containerIP, targetIP string, containerHttpPort, targetPort, maxTries, minTries int, expectedEps sets.String) {
cmd := fmt.Sprintf("curl -q -s 'http://%s:%d/dial?request=hostName&protocol=%s&host=%s&port=%d&tries=1'",
containerIP,
containerHttpPort,
protocol,
targetIP,
targetPort)
eps := sets.NewString()
for i := 0; i < maxTries; i++ {
stdout, stderr, err := config.f.ExecShellInPodWithFullOutput(config.HostTestContainerPod.Name, cmd)
if err != nil {
// A failure to kubectl exec counts as a try, not a hard fail.
// Also note that we will keep failing for maxTries in tests where
// we confirm unreachability.
Logf("Failed to execute %q: %v, stdout: %q, stderr %q", cmd, err, stdout, stderr)
} else {
var output map[string][]string
if err := json.Unmarshal([]byte(stdout), &output); err != nil {
Logf("WARNING: Failed to unmarshal curl response. Cmd %v run in %v, output: %s, err: %v",
cmd, config.HostTestContainerPod.Name, stdout, err)
continue
}
for _, hostName := range output["responses"] {
trimmed := strings.TrimSpace(hostName)
if trimmed != "" {
eps.Insert(trimmed)
}
}
}
Logf("Waiting for endpoints: %v", expectedEps.Difference(eps))
// Check against i+1 so we exit if minTries == maxTries.
if (eps.Equal(expectedEps) || eps.Len() == 0 && expectedEps.Len() == 0) && i+1 >= minTries {
return
}
// TODO: get rid of this delay #36281
time.Sleep(hitEndpointRetryDelay)
}
config.diagnoseMissingEndpoints(eps)
Failf("Failed to find expected endpoints:\nTries %d\nCommand %v\nretrieved %v\nexpected %v\n", minTries, cmd, eps, expectedEps)
}
// DialFromNode executes a tcp or udp request based on protocol via kubectl exec
// in a test container running with host networking.
// - minTries is the minimum number of curl attempts required before declaring
// success. Set to 0 if you'd like to return as soon as all endpoints respond
// at least once.
// - maxTries is the maximum number of curl attempts. If this many attempts pass
// and we don't see all expected endpoints, the test fails.
// maxTries == minTries will confirm that we see the expected endpoints and no
// more for maxTries. Use this if you want to eg: fail a readiness check on a
// pod and confirm it doesn't show up as an endpoint.
func (config *NetworkingTestConfig) DialFromNode(protocol, targetIP string, targetPort, maxTries, minTries int, expectedEps sets.String) {
var cmd string
if protocol == "udp" {
// TODO: It would be enough to pass 1s+epsilon to timeout, but unfortunately
// busybox timeout doesn't support non-integer values.
cmd = fmt.Sprintf("echo 'hostName' | timeout -t 2 nc -w 1 -u %s %d", targetIP, targetPort)
} else {
cmd = fmt.Sprintf("timeout -t 15 curl -q -s --connect-timeout 1 http://%s:%d/hostName", targetIP, targetPort)
}
// TODO: This simply tells us that we can reach the endpoints. Check that
// the probability of hitting a specific endpoint is roughly the same as
// hitting any other.
eps := sets.NewString()
filterCmd := fmt.Sprintf("%s | grep -v '^\\s*$'", cmd)
for i := 0; i < maxTries; i++ {
stdout, stderr, err := config.f.ExecShellInPodWithFullOutput(config.HostTestContainerPod.Name, filterCmd)
if err != nil || len(stderr) > 0 {
// A failure to exec command counts as a try, not a hard fail.
// Also note that we will keep failing for maxTries in tests where
// we confirm unreachability.
Logf("Failed to execute %q: %v, stdout: %q, stderr: %q", filterCmd, err, stdout, stderr)
} else {
trimmed := strings.TrimSpace(stdout)
if trimmed != "" {
eps.Insert(trimmed)
}
}
Logf("Waiting for %+v endpoints, got endpoints %+v", expectedEps.Difference(eps), eps)
// Check against i+1 so we exit if minTries == maxTries.
if (eps.Equal(expectedEps) || eps.Len() == 0 && expectedEps.Len() == 0) && i+1 >= minTries {
return
}
// TODO: get rid of this delay #36281
time.Sleep(hitEndpointRetryDelay)
}
config.diagnoseMissingEndpoints(eps)
Failf("Failed to find expected endpoints:\nTries %d\nCommand %v\nretrieved %v\nexpected %v\n", minTries, cmd, eps, expectedEps)
}
func TestHammerController(t *testing.T) {
// This test executes a bunch of requests through the fake source and
// controller framework to make sure there's no locking/threading
// errors. If an error happens, it should hang forever or trigger the
// race detector.
// source simulates an apiserver object endpoint.
source := fcache.NewFakeControllerSource()
// Let's do threadsafe output to get predictable test results.
outputSetLock := sync.Mutex{}
// map of key to operations done on the key
outputSet := map[string][]string{}
recordFunc := func(eventType string, obj interface{}) {
key, err := DeletionHandlingMetaNamespaceKeyFunc(obj)
if err != nil {
t.Errorf("something wrong with key: %v", err)
key = "oops something went wrong with the key"
}
// Record some output when items are deleted.
outputSetLock.Lock()
defer outputSetLock.Unlock()
outputSet[key] = append(outputSet[key], eventType)
}
// Make a controller which just logs all the changes it gets.
_, controller := NewInformer(
source,
&v1.Pod{},
time.Millisecond*100,
ResourceEventHandlerFuncs{
AddFunc: func(obj interface{}) { recordFunc("add", obj) },
UpdateFunc: func(oldObj, newObj interface{}) { recordFunc("update", newObj) },
DeleteFunc: func(obj interface{}) { recordFunc("delete", obj) },
},
)
if controller.HasSynced() {
t.Errorf("Expected HasSynced() to return false before we started the controller")
}
// Run the controller and run it until we close stop.
stop := make(chan struct{})
go controller.Run(stop)
// Let's wait for the controller to do its initial sync
wait.Poll(100*time.Millisecond, wait.ForeverTestTimeout, func() (bool, error) {
return controller.HasSynced(), nil
})
if !controller.HasSynced() {
t.Errorf("Expected HasSynced() to return true after the initial sync")
}
wg := sync.WaitGroup{}
const threads = 3
wg.Add(threads)
for i := 0; i < threads; i++ {
go func() {
defer wg.Done()
// Let's add a few objects to the source.
currentNames := sets.String{}
rs := rand.NewSource(rand.Int63())
f := fuzz.New().NilChance(.5).NumElements(0, 2).RandSource(rs)
r := rand.New(rs) // Mustn't use r and f concurrently!
for i := 0; i < 100; i++ {
var name string
var isNew bool
if currentNames.Len() == 0 || r.Intn(3) == 1 {
f.Fuzz(&name)
isNew = true
} else {
l := currentNames.List()
name = l[r.Intn(len(l))]
}
pod := &v1.Pod{}
f.Fuzz(pod)
pod.ObjectMeta.Name = name
pod.ObjectMeta.Namespace = "default"
// Add, update, or delete randomly.
// Note that these pods are not valid-- the fake source doesn't
// call validation or perform any other checking.
if isNew {
currentNames.Insert(name)
source.Add(pod)
continue
}
switch r.Intn(2) {
case 0:
currentNames.Insert(name)
source.Modify(pod)
case 1:
currentNames.Delete(name)
source.Delete(pod)
}
}
}()
}
wg.Wait()
// Let's wait for the controller to finish processing the things we just added.
// TODO: look in the queue to see how many items need to be processed.
time.Sleep(100 * time.Millisecond)
close(stop)
// TODO: Verify that no goroutines were leaked here and that everything shut
// down cleanly.
outputSetLock.Lock()
t.Logf("got: %#v", outputSet)
}
