storage/cacher: ensure the cache is at the Most Recent ResourceVersion when streaming was requested

staging/src/k8s.io/apiserver/pkg/storage/cacher/cacher.go

@@ -513,6 +513,7 @@ func (c *Cacher) Watch(ctx context.Context, key string, opts storage.ListOptions
 	if !utilfeature.DefaultFeatureGate.Enabled(features.WatchList) && opts.SendInitialEvents != nil {
 		opts.SendInitialEvents = nil
 	}
+	// TODO: we should eventually get rid of this legacy case
 	if opts.SendInitialEvents == nil && opts.ResourceVersion == "" {
 		return c.storage.Watch(ctx, key, opts)
 	}
@@ -557,14 +558,14 @@ func (c *Cacher) Watch(ctx context.Context, key string, opts storage.ListOptions
 	//   watchers on our watcher having a processing hiccup
 	chanSize := c.watchCache.suggestedWatchChannelSize(c.indexedTrigger != nil, triggerSupported)
 
-	// Determine a function that computes the bookmarkAfterResourceVersion
+	// Determine the ResourceVersion to which the watch cache must be synchronized
-	bookmarkAfterResourceVersionFn, err := c.getBookmarkAfterResourceVersionLockedFunc(ctx, requestedWatchRV, opts)
+	requiredResourceVersion, err := c.getWatchCacheResourceVersion(ctx, requestedWatchRV, opts)
 	if err != nil {
 		return newErrWatcher(err), nil
 	}
 
-	// Determine a function that computes the watchRV we should start from
+	// Determine a function that computes the bookmarkAfterResourceVersion
-	startWatchResourceVersionFn, err := c.getStartResourceVersionForWatchLockedFunc(ctx, requestedWatchRV, opts)
+	bookmarkAfterResourceVersionFn, err := c.getBookmarkAfterResourceVersionLockedFunc(requestedWatchRV, requiredResourceVersion, opts)
 	if err != nil {
 		return newErrWatcher(err), nil
 	}
@@ -596,7 +597,7 @@ func (c *Cacher) Watch(ctx context.Context, key string, opts storage.ListOptions
 	// moreover even though the c.waitUntilWatchCacheFreshAndForceAllEvents acquires a lock
 	// it is safe to release the lock after the method finishes because we don't require
 	// any atomicity between the call to the method and further calls that actually get the events.
-	forceAllEvents, err := c.waitUntilWatchCacheFreshAndForceAllEvents(ctx, requestedWatchRV, opts)
+	err = c.waitUntilWatchCacheFreshAndForceAllEvents(ctx, requiredResourceVersion, opts)
 	if err != nil {
 		return newErrWatcher(err), nil
 	}
@@ -609,13 +610,8 @@ func (c *Cacher) Watch(ctx context.Context, key string, opts storage.ListOptions
 	c.watchCache.RLock()
 	defer c.watchCache.RUnlock()
 
-	startWatchRV := startWatchResourceVersionFn()
 	var cacheInterval *watchCacheInterval
-	if forceAllEvents {
+	cacheInterval, err = c.watchCache.getAllEventsSinceLocked(requiredResourceVersion, opts)
-		cacheInterval, err = c.watchCache.getIntervalFromStoreLocked()
-	} else {
-		cacheInterval, err = c.watchCache.getAllEventsSinceLocked(startWatchRV)
-	}
 	if err != nil {
 		// To match the uncached watch implementation, once we have passed authn/authz/admission,
 		// and successfully parsed a resource version, other errors must fail with a watch event of type ERROR,
@@ -657,7 +653,7 @@ func (c *Cacher) Watch(ctx context.Context, key string, opts storage.ListOptions
 		return newImmediateCloseWatcher(), nil
 	}
 
-	go watcher.processInterval(ctx, cacheInterval, startWatchRV)
+	go watcher.processInterval(ctx, cacheInterval, requiredResourceVersion)
 	return watcher, nil
 }
 
@@ -1249,59 +1245,61 @@ func (c *Cacher) LastSyncResourceVersion() (uint64, error) {
 // spits a ResourceVersion after which the bookmark event will be delivered.
 //
 // The returned function must be called under the watchCache lock.
-func (c *Cacher) getBookmarkAfterResourceVersionLockedFunc(ctx context.Context, parsedResourceVersion uint64, opts storage.ListOptions) (func() uint64, error) {
+func (c *Cacher) getBookmarkAfterResourceVersionLockedFunc(parsedResourceVersion, requiredResourceVersion uint64, opts storage.ListOptions) (func() uint64, error) {
 	if opts.SendInitialEvents == nil || !*opts.SendInitialEvents || !opts.Predicate.AllowWatchBookmarks {
 		return func() uint64 { return 0 }, nil
 	}
-	return c.getCommonResourceVersionLockedFunc(ctx, parsedResourceVersion, opts)
-}
 
-// getStartResourceVersionForWatchLockedFunc returns a function that
-// spits a ResourceVersion the watch will be started from.
-// Depending on the input parameters the semantics of the returned ResourceVersion are:
-//   - start at Exact (return parsedWatchResourceVersion)
-//   - start at Most Recent (return an RV from etcd)
-//   - start at Any (return the current watchCache's RV)
-//
-// The returned function must be called under the watchCache lock.
-func (c *Cacher) getStartResourceVersionForWatchLockedFunc(ctx context.Context, parsedWatchResourceVersion uint64, opts storage.ListOptions) (func() uint64, error) {
-	if opts.SendInitialEvents == nil || *opts.SendInitialEvents {
-		return func() uint64 { return parsedWatchResourceVersion }, nil
-	}
-	return c.getCommonResourceVersionLockedFunc(ctx, parsedWatchResourceVersion, opts)
-}
-
-// getCommonResourceVersionLockedFunc a helper that simply computes a ResourceVersion
-// based on the input parameters. Please examine callers of this method to get more context.
-//
-// The returned function must be called under the watchCache lock.
-func (c *Cacher) getCommonResourceVersionLockedFunc(ctx context.Context, parsedWatchResourceVersion uint64, opts storage.ListOptions) (func() uint64, error) {
 	switch {
 	case len(opts.ResourceVersion) == 0:
-		rv, err := storage.GetCurrentResourceVersionFromStorage(ctx, c.storage, c.newListFunc, c.resourcePrefix, c.objectType.String())
+		return func() uint64 { return requiredResourceVersion }, nil
-		if err != nil {
+	case parsedResourceVersion == 0:
-			return nil, err
-		}
-		return func() uint64 { return rv }, nil
-	case parsedWatchResourceVersion == 0:
 		// here we assume that watchCache locked is already held
 		return func() uint64 { return c.watchCache.resourceVersion }, nil
 	default:
-		return func() uint64 { return parsedWatchResourceVersion }, nil
+		return func() uint64 { return parsedResourceVersion }, nil
 	}
 }
 
+// getWatchCacheResourceVersion returns a ResourceVersion to which the watch cache must be synchronized to
+//
+// Depending on the input parameters, the semantics of the returned ResourceVersion are:
+//   - must be at Exact RV (when parsedWatchResourceVersion > 0)
+//   - can be at Any RV (when parsedWatchResourceVersion = 0)
+//   - must be at Most Recent RV (return an RV from etcd)
+//
+// note that the above semantic is enforced by the API validation (defined elsewhere):
+//
+//	if SendInitiaEvents != nil => ResourceVersionMatch = NotOlderThan
+//	if ResourceVersionmatch != nil => ResourceVersionMatch = NotOlderThan & SendInitialEvents != nil
+//
+// to satisfy the legacy case (SendInitialEvents = true, RV="") we skip checking opts.Predicate.AllowWatchBookmarks
+func (c *Cacher) getWatchCacheResourceVersion(ctx context.Context, parsedWatchResourceVersion uint64, opts storage.ListOptions) (uint64, error) {
+	if len(opts.ResourceVersion) != 0 {
+		return parsedWatchResourceVersion, nil
+	}
+	rv, err := storage.GetCurrentResourceVersionFromStorage(ctx, c.storage, c.newListFunc, c.resourcePrefix, c.objectType.String())
+	return rv, err
+}
+
 // waitUntilWatchCacheFreshAndForceAllEvents waits until cache is at least
 // as fresh as given requestedWatchRV if sendInitialEvents was requested.
-// Additionally, it instructs the caller whether it should ask for
+// otherwise, we allow for establishing the connection because the clients
-// all events from the cache (full state) or not.
+// can wait for events without unnecessary blocking.
-func (c *Cacher) waitUntilWatchCacheFreshAndForceAllEvents(ctx context.Context, requestedWatchRV uint64, opts storage.ListOptions) (bool, error) {
+func (c *Cacher) waitUntilWatchCacheFreshAndForceAllEvents(ctx context.Context, requestedWatchRV uint64, opts storage.ListOptions) error {
 	if opts.SendInitialEvents != nil && *opts.SendInitialEvents {
+		// TODO(p0lyn0mial): adapt the following logic once
+		//   https://github.com/kubernetes/kubernetes/pull/123264 merges
+		if utilfeature.DefaultFeatureGate.Enabled(features.ConsistentListFromCache) && c.watchCache.notFresh(requestedWatchRV) {
+			c.watchCache.waitingUntilFresh.Add()
+			defer c.watchCache.waitingUntilFresh.Remove()
+		}
+		// TODO(p0lyn0mial): add a metric to track the number of times we have failed while waiting
 		err := c.watchCache.waitUntilFreshAndBlock(ctx, requestedWatchRV)
 		defer c.watchCache.RUnlock()
-		return err == nil, err
+		return err
 	}
-	return false, nil
+	return nil
 }
 
 // errWatcher implements watch.Interface to return a single error

staging/src/k8s.io/apiserver/pkg/storage/cacher/cacher_whitebox_test.go

@@ -32,6 +32,7 @@ import (
 
 	apiequality "k8s.io/apimachinery/pkg/api/equality"
 	apierrors "k8s.io/apimachinery/pkg/api/errors"
+	"k8s.io/apimachinery/pkg/api/meta"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/fields"
 	"k8s.io/apimachinery/pkg/labels"
@@ -85,12 +86,25 @@ type dummyStorage struct {
 	err       error
 	getListFn func(_ context.Context, _ string, _ storage.ListOptions, listObj runtime.Object) error
 	watchFn   func(_ context.Context, _ string, _ storage.ListOptions) (watch.Interface, error)
+
+	// use getRequestWatchProgressCounter when reading
+	// the value of the counter
+	requestWatchProgressCounter int
 }
 
 func (d *dummyStorage) RequestWatchProgress(ctx context.Context) error {
+	d.Lock()
+	defer d.Unlock()
+	d.requestWatchProgressCounter++
 	return nil
 }
 
+func (d *dummyStorage) getRequestWatchProgressCounter() int {
+	d.RLock()
+	defer d.RUnlock()
+	return d.requestWatchProgressCounter
+}
+
 type dummyWatch struct {
 	ch chan watch.Event
 }
@@ -1605,48 +1619,156 @@ func TestCacheIntervalInvalidationStopsWatch(t *testing.T) {
 
 func TestWaitUntilWatchCacheFreshAndForceAllEvents(t *testing.T) {
 	defer featuregatetesting.SetFeatureGateDuringTest(t, utilfeature.DefaultFeatureGate, features.WatchList, true)()
-	backingStorage := &dummyStorage{}
+	defer featuregatetesting.SetFeatureGateDuringTest(t, utilfeature.DefaultFeatureGate, features.ConsistentListFromCache, true)()
-	cacher, _, err := newTestCacher(backingStorage)
-	if err != nil {
-		t.Fatalf("Couldn't create cacher: %v", err)
-	}
-	defer cacher.Stop()
 
-	opts := storage.ListOptions{
+	scenarios := []struct {
-		Predicate:         storage.Everything,
+		name               string
-		SendInitialEvents: pointer.Bool(true),
+		opts               storage.ListOptions
-		ResourceVersion:   "105",
+		backingStorage     *dummyStorage
-	}
+		verifyBackingStore func(t *testing.T, s *dummyStorage)
-	opts.Predicate.AllowWatchBookmarks = true
+	}{
-
-	w, err := cacher.Watch(context.Background(), "pods/ns", opts)
-	require.NoError(t, err, "failed to create watch: %v")
-	defer w.Stop()
-	verifyEvents(t, w, []watch.Event{
 		{
-			Type: watch.Error,
+			name: "allowWatchBookmarks=true, sendInitialEvents=true, RV=105",
-			Object: &metav1.Status{
+			opts: storage.ListOptions{
-				Status:  metav1.StatusFailure,
+				Predicate: func() storage.SelectionPredicate {
-				Message: storage.NewTooLargeResourceVersionError(105, 100, resourceVersionTooHighRetrySeconds).Error(),
+					p := storage.Everything
-				Details: storage.NewTooLargeResourceVersionError(105, 100, resourceVersionTooHighRetrySeconds).(*apierrors.StatusError).Status().Details,
+					p.AllowWatchBookmarks = true
-				Reason:  metav1.StatusReasonTimeout,
+					return p
-				Code:    504,
+				}(),
+				SendInitialEvents: pointer.Bool(true),
+				ResourceVersion:   "105",
+			},
+			verifyBackingStore: func(t *testing.T, s *dummyStorage) {
+				require.NotEqual(t, 0, s.requestWatchProgressCounter, "expected store.RequestWatchProgressCounter to be > 0. It looks like watch progress wasn't requested!")
 			},
 		},
-	}, true)
 
-	go func() {
-		cacher.watchCache.Add(makeTestPodDetails("pod1", 105, "node1", map[string]string{"label": "value1"}))
-	}()
-	w, err = cacher.Watch(context.Background(), "pods/ns", opts)
-	require.NoError(t, err, "failed to create watch: %v")
-	defer w.Stop()
-	verifyEvents(t, w, []watch.Event{
 		{
-			Type:   watch.Added,
+			name: "legacy: allowWatchBookmarks=false, sendInitialEvents=true, RV=unset",
-			Object: makeTestPodDetails("pod1", 105, "node1", map[string]string{"label": "value1"}),
+			opts: storage.ListOptions{
+				Predicate: func() storage.SelectionPredicate {
+					p := storage.Everything
+					p.AllowWatchBookmarks = false
+					return p
+				}(),
+				SendInitialEvents: pointer.Bool(true),
+			},
+			backingStorage: func() *dummyStorage {
+				hasBeenPrimed := false
+				s := &dummyStorage{}
+				s.getListFn = func(_ context.Context, key string, opts storage.ListOptions, listObj runtime.Object) error {
+					listAccessor, err := meta.ListAccessor(listObj)
+					if err != nil {
+						return err
+					}
+					// the first call to this function
+					// primes the cacher
+					if !hasBeenPrimed {
+						listAccessor.SetResourceVersion("100")
+						hasBeenPrimed = true
+						return nil
+					}
+					listAccessor.SetResourceVersion("105")
+					return nil
+				}
+				return s
+			}(),
+			verifyBackingStore: func(t *testing.T, s *dummyStorage) {
+				require.NotEqual(t, 0, s.getRequestWatchProgressCounter(), "expected store.RequestWatchProgressCounter to be > 0. It looks like watch progress wasn't requested!")
+			},
 		},
-	}, true)
+
+		{
+			name: "allowWatchBookmarks=true, sendInitialEvents=true, RV=unset",
+			opts: storage.ListOptions{
+				Predicate: func() storage.SelectionPredicate {
+					p := storage.Everything
+					p.AllowWatchBookmarks = true
+					return p
+				}(),
+				SendInitialEvents: pointer.Bool(true),
+			},
+			backingStorage: func() *dummyStorage {
+				hasBeenPrimed := false
+				s := &dummyStorage{}
+				s.getListFn = func(_ context.Context, key string, opts storage.ListOptions, listObj runtime.Object) error {
+					listAccessor, err := meta.ListAccessor(listObj)
+					if err != nil {
+						return err
+					}
+					// the first call to this function
+					// primes the cacher
+					if !hasBeenPrimed {
+						listAccessor.SetResourceVersion("100")
+						hasBeenPrimed = true
+						return nil
+					}
+					listAccessor.SetResourceVersion("105")
+					return nil
+				}
+				return s
+			}(),
+			verifyBackingStore: func(t *testing.T, s *dummyStorage) {
+				require.NotEqual(t, 0, s.getRequestWatchProgressCounter(), "expected store.RequestWatchProgressCounter to be > 0. It looks like watch progress wasn't requested!")
+			},
+		},
+	}
+
+	for _, scenario := range scenarios {
+		t.Run(scenario.name, func(t *testing.T) {
+			var backingStorage *dummyStorage
+			if scenario.backingStorage != nil {
+				backingStorage = scenario.backingStorage
+			} else {
+				backingStorage = &dummyStorage{}
+			}
+			cacher, _, err := newTestCacher(backingStorage)
+			if err != nil {
+				t.Fatalf("Couldn't create cacher: %v", err)
+			}
+			defer cacher.Stop()
+			if err := cacher.ready.wait(context.Background()); err != nil {
+				t.Fatalf("unexpected error waiting for the cache to be ready")
+			}
+
+			w, err := cacher.Watch(context.Background(), "pods/ns", scenario.opts)
+			require.NoError(t, err, "failed to create watch: %v")
+			defer w.Stop()
+			var expectedErr *apierrors.StatusError
+			if !errors.As(storage.NewTooLargeResourceVersionError(105, 100, resourceVersionTooHighRetrySeconds), &expectedErr) {
+				t.Fatalf("Unable to convert NewTooLargeResourceVersionError to apierrors.StatusError")
+			}
+			verifyEvents(t, w, []watch.Event{
+				{
+					Type: watch.Error,
+					Object: &metav1.Status{
+						Status:  metav1.StatusFailure,
+						Message: expectedErr.Error(),
+						Details: expectedErr.ErrStatus.Details,
+						Reason:  metav1.StatusReasonTimeout,
+						Code:    504,
+					},
+				},
+			}, true)
+
+			go func(t *testing.T) {
+				err := cacher.watchCache.Add(makeTestPodDetails("pod1", 105, "node1", map[string]string{"label": "value1"}))
+				require.NoError(t, err, "failed adding a pod to the watchCache")
+			}(t)
+			w, err = cacher.Watch(context.Background(), "pods/ns", scenario.opts)
+			require.NoError(t, err, "failed to create watch: %v")
+			defer w.Stop()
+			verifyEvents(t, w, []watch.Event{
+				{
+					Type:   watch.Added,
+					Object: makeTestPodDetails("pod1", 105, "node1", map[string]string{"label": "value1"}),
+				},
+			}, true)
+			if scenario.verifyBackingStore != nil {
+				scenario.verifyBackingStore(t, backingStorage)
+			}
+		})
+	}
 }
 
 type fakeStorage struct {