React to changes in watch cache initialization

Remove some complexity in RESTOptionsGetter and add default watch cache
sizes for resources that are read by nodes.

pkg/cmd/server/kubernetes/master/master_config.go

@@ -507,6 +507,20 @@ func buildKubeApiserverConfig(
 		return originLongRunningRequestRE.MatchString(r.URL.Path) || kubeLongRunningFunc(r, requestInfo)
 	}
 
+	if apiserverOptions.Etcd.EnableWatchCache {
+		glog.V(2).Infof("Initializing cache sizes based on %dMB limit", apiserverOptions.GenericServerRunOptions.TargetRAMMB)
+		sizes := cachesize.NewHeuristicWatchCacheSizes(apiserverOptions.GenericServerRunOptions.TargetRAMMB)
+		if userSpecified, err := genericoptions.ParseWatchCacheSizes(apiserverOptions.Etcd.WatchCacheSizes); err == nil {
+			for resource, size := range userSpecified {
+				sizes[resource] = size
+			}
+		}
+		apiserverOptions.Etcd.WatchCacheSizes, err = genericoptions.WriteWatchCacheSizes(sizes)
+		if err != nil {
+			return nil, err
+		}
+	}
+
 	if err := apiserverOptions.Etcd.ApplyWithStorageFactoryTo(storageFactory, genericConfig); err != nil {
 		return nil, err
 	}
@@ -566,12 +580,6 @@ func buildKubeApiserverConfig(
 		EnableCoreControllers: true,
 	}
 
-	if apiserverOptions.Etcd.EnableWatchCache {
-		// TODO(rebase): upstream also does the following:
-		// cachesize.InitializeWatchCacheSizes(s.GenericServerRunOptions.TargetRAMMB)
-		cachesize.SetWatchCacheSizes(apiserverOptions.GenericServerRunOptions.WatchCacheSizes)
-	}
-
 	if kubeApiserverConfig.EnableCoreControllers {
 		ttl := masterConfig.KubernetesMasterConfig.MasterEndpointReconcileTTL
 		interval := ttl * 2 / 3

pkg/security/registry/securitycontextconstraints/etcd/etcd.go

@@ -6,7 +6,6 @@ import (
 	"k8s.io/apiserver/pkg/registry/generic/registry"
 	"k8s.io/apiserver/pkg/registry/rest"
 	"k8s.io/kubernetes/pkg/api"
-	"k8s.io/kubernetes/pkg/registry/cachesize"
 
 	securityapi "github.com/openshift/origin/pkg/security/apis/security"
 	"github.com/openshift/origin/pkg/security/registry/securitycontextconstraints"
@@ -30,7 +29,6 @@ func NewREST(optsGetter generic.RESTOptionsGetter) *REST {
 		},
 		PredicateFunc:            securitycontextconstraints.Matcher,
 		DefaultQualifiedResource: securityapi.Resource("securitycontextconstraints"),
-		WatchCacheSize:           cachesize.GetWatchCacheSizeByResource("securitycontextconstraints"),
 
 		CreateStrategy:      securitycontextconstraints.Strategy,
 		UpdateStrategy:      securitycontextconstraints.Strategy,

pkg/util/restoptions/configgetter.go

@@ -1,22 +1,16 @@
 package restoptions
 
 import (
-	"fmt"
-	"strconv"
-	"strings"
 	"sync"
 
-	"k8s.io/apimachinery/pkg/runtime"
+	"github.com/golang/glog"
+
 	"k8s.io/apimachinery/pkg/runtime/schema"
-	kerrors "k8s.io/apimachinery/pkg/util/errors"
 	"k8s.io/apiserver/pkg/registry/generic"
 	"k8s.io/apiserver/pkg/registry/generic/registry"
+	"k8s.io/apiserver/pkg/server/options"
 	serverstorage "k8s.io/apiserver/pkg/server/storage"
-	"k8s.io/apiserver/pkg/storage"
-	"k8s.io/apiserver/pkg/storage/storagebackend"
-	"k8s.io/apiserver/pkg/storage/storagebackend/factory"
 
-	"github.com/golang/glog"
 	configapi "github.com/openshift/origin/pkg/cmd/server/api"
 	kubernetes "github.com/openshift/origin/pkg/cmd/server/kubernetes/master"
 )
@@ -41,7 +35,6 @@ type configRESTOptionsGetter struct {
 }
 
 // NewConfigGetter returns a restoptions.Getter implemented using information from the provided master config.
-// By default, the etcd watch cache is enabled with a size of 1000 per resource type.
 // TODO: this class should either not need to know about configapi.MasterConfig, or not be in pkg/util
 func NewConfigGetter(masterOptions configapi.MasterConfig, defaultResourceConfig *serverstorage.ResourceConfig, resourcePrefixOverrides map[schema.GroupResource]string, enforcedStorageVersions map[schema.GroupResource]schema.GroupVersion, quorumResources map[schema.GroupResource]struct{}) (Getter, error) {
 	apiserverOptions, err := kubernetes.BuildKubeAPIserverOptions(masterOptions)
@@ -55,27 +48,24 @@ func NewConfigGetter(masterOptions configapi.MasterConfig, defaultResourceConfig
 	storageFactory.DefaultResourcePrefixes = resourcePrefixOverrides
 	storageFactory.StorageConfig.Prefix = masterOptions.EtcdStorageConfig.OpenShiftStoragePrefix
 
-	// TODO: refactor vendor/k8s.io/kubernetes/pkg/registry/cachesize to remove our custom cache size code
-	errs := []error{}
-	cacheSizes := map[schema.GroupResource]int{}
-	for _, c := range apiserverOptions.GenericServerRunOptions.WatchCacheSizes {
-		tokens := strings.Split(c, "#")
-		if len(tokens) != 2 {
-			errs = append(errs, fmt.Errorf("invalid watch cache size value '%s', expecting <resource>#<size> format (e.g. builds#100)", c))
-			continue
+	// perform watch cache heuristic like upstream
+	if apiserverOptions.Etcd.EnableWatchCache {
+		glog.V(2).Infof("Initializing cache sizes based on %dMB limit", apiserverOptions.GenericServerRunOptions.TargetRAMMB)
+		sizes := newHeuristicWatchCacheSizes(apiserverOptions.GenericServerRunOptions.TargetRAMMB)
+		if userSpecified, err := options.ParseWatchCacheSizes(apiserverOptions.Etcd.WatchCacheSizes); err == nil {
+			for resource, size := range userSpecified {
+				sizes[resource] = size
+			}
 		}
-
-		resource := schema.ParseGroupResource(tokens[0])
-
-		size, err := strconv.Atoi(tokens[1])
+		apiserverOptions.Etcd.WatchCacheSizes, err = options.WriteWatchCacheSizes(sizes)
 		if err != nil {
-			errs = append(errs, fmt.Errorf("invalid watch cache size value '%s': %v", c, err))
-			continue
+			return nil, err
 		}
-		cacheSizes[resource] = size
 	}
-	if len(errs) > 0 {
-		return nil, kerrors.NewAggregate(errs)
+
+	cacheSizes, err := options.ParseWatchCacheSizes(apiserverOptions.Etcd.WatchCacheSizes)
+	if err != nil {
+		return nil, err
 	}
 
 	return &configRESTOptionsGetter{
@@ -108,41 +98,14 @@ func (g *configRESTOptionsGetter) GetRESTOptions(resource schema.GroupResource)
 		config.Quorum = true
 	}
 
-	configuredCacheSize, specified := g.cacheSizes[resource]
-	if !specified || configuredCacheSize < 0 {
-		configuredCacheSize = g.defaultCacheSize
-	}
-	storageWithCacher := registry.StorageWithCacher(configuredCacheSize)
-
-	decorator := func(
-		copier runtime.ObjectCopier,
-		storageConfig *storagebackend.Config,
-		requestedSize *int,
-		objectType runtime.Object,
-		resourcePrefix string,
-		keyFunc func(obj runtime.Object) (string, error),
-		newListFn func() runtime.Object,
-		getAttrsFunc storage.AttrFunc,
-		triggerFn storage.TriggerPublisherFunc,
-	) (storage.Interface, factory.DestroyFunc) {
-		// use the origin default cache size, not the one in registry.StorageWithCacher
-		capacity := &configuredCacheSize
-		if requestedSize != nil {
-			capacity = requestedSize
-		}
-
-		if *capacity == 0 || !g.cacheEnabled {
-			glog.V(5).Infof("using uncached watch storage for %s (quorum=%t)", resource.String(), storageConfig.Quorum)
-			return generic.UndecoratedStorage(copier, storageConfig, capacity, objectType, resourcePrefix, keyFunc, newListFn, getAttrsFunc, triggerFn)
-		}
-
-		glog.V(5).Infof("using watch cache storage (capacity=%v, quorum=%t) for %s %#v", *capacity, storageConfig.Quorum, resource.String(), storageConfig)
-		return storageWithCacher(copier, storageConfig, capacity, objectType, resourcePrefix, keyFunc, newListFn, getAttrsFunc, triggerFn)
+	cacheSize, ok := g.cacheSizes[resource]
+	if !ok {
+		cacheSize = g.defaultCacheSize
 	}
 
 	resourceOptions := generic.RESTOptions{
 		StorageConfig:           config,
-		Decorator:               decorator,
+		Decorator:               registry.StorageWithCacher(cacheSize),
 		DeleteCollectionWorkers: g.deleteCollectionWorkers,
 		EnableGarbageCollection: g.enableGarbageCollection,
 		ResourcePrefix:          g.storageFactory.ResourcePrefix(resource),
@@ -151,3 +114,25 @@ func (g *configRESTOptionsGetter) GetRESTOptions(resource schema.GroupResource)
 
 	return resourceOptions, nil
 }
+
+// newHeuristicWatchCacheSizes returns a map of suggested watch cache sizes based on total
+// memory. It reuses the upstream heuristic and adds OpenShift specific resources.
+func newHeuristicWatchCacheSizes(expectedRAMCapacityMB int) map[schema.GroupResource]int {
+	// TODO: Revisit this heuristic, copied from upstream
+	clusterSize := expectedRAMCapacityMB / 60
+
+	// default enable watch caches for resources that will have a high number of clients accessing it
+	// and where the write rate may be significant
+	watchCacheSizes := make(map[schema.GroupResource]int)
+	watchCacheSizes[schema.GroupResource{Group: "network.openshift.io", Resource: "hostsubnets"}] = maxInt(5*clusterSize, 100)
+	watchCacheSizes[schema.GroupResource{Group: "network.openshift.io", Resource: "netnamespaces"}] = maxInt(5*clusterSize, 100)
+	watchCacheSizes[schema.GroupResource{Group: "network.openshift.io", Resource: "egressnetworkpolicies"}] = maxInt(10*clusterSize, 100)
+	return watchCacheSizes
+}
+
+func maxInt(a, b int) int {
+	if a > b {
+		return a
+	}
+	return b
+}