Use the route informer to detect conflicts

Because the initial sync of the router may take a long time and we need
to detect conflicts caused early in that sync, separate out contention
detection into a separate goroutine fed by the informer which looks for
repeated changes to a given route's status. The status plugin resets the
state to candidate when it makes a write.

pkg/cmd/infra/router/f5.go

@@ -239,10 +239,11 @@ func (o *F5RouterOptions) Run() error {
 	if o.UpdateStatus {
 		lease := writerlease.New(time.Minute, 3*time.Second)
 		go lease.Run(wait.NeverStop)
-		tracker := controller.NewSimpleContentionTracker(o.ResyncInterval / 10)
+		informer := factory.CreateRoutesSharedInformer()
+		tracker := controller.NewSimpleContentionTracker(informer, o.RouterName, o.ResyncInterval/10)
 		tracker.SetConflictMessage(fmt.Sprintf("The router detected another process is writing conflicting updates to route status with name %q. Please ensure that the configuration of all routers is consistent. Route status will not be updated as long as conflicts are detected.", o.RouterName))
 		go tracker.Run(wait.NeverStop)
-		routeLister := routelisters.NewRouteLister(factory.CreateRoutesSharedInformer().GetIndexer())
+		routeLister := routelisters.NewRouteLister(informer.GetIndexer())
 		status := controller.NewStatusAdmitter(plugin, routeclient.Route(), routeLister, o.RouterName, o.RouterCanonicalHostname, lease, tracker)
 		recorder = status
 		plugin = status

pkg/cmd/infra/router/template.go

@@ -420,10 +420,11 @@ func (o *TemplateRouterOptions) Run() error {
 	if o.UpdateStatus {
 		lease := writerlease.New(time.Minute, 3*time.Second)
 		go lease.Run(wait.NeverStop)
-		tracker := controller.NewSimpleContentionTracker(o.ResyncInterval / 10)
+		informer := factory.CreateRoutesSharedInformer()
+		tracker := controller.NewSimpleContentionTracker(informer, o.RouterName, o.ResyncInterval/10)
 		tracker.SetConflictMessage(fmt.Sprintf("The router detected another process is writing conflicting updates to route status with name %q. Please ensure that the configuration of all routers is consistent. Route status will not be updated as long as conflicts are detected.", o.RouterName))
 		go tracker.Run(wait.NeverStop)
-		routeLister := routelisters.NewRouteLister(factory.CreateRoutesSharedInformer().GetIndexer())
+		routeLister := routelisters.NewRouteLister(informer.GetIndexer())
 		status := controller.NewStatusAdmitter(plugin, routeclient.Route(), routeLister, o.RouterName, o.RouterCanonicalHostname, lease, tracker)
 		recorder = status
 		plugin = status

pkg/router/controller/contention.go

@@ -0,0 +1,281 @@
+package controller
+
+import (
+	"sync"
+	"time"
+
+	"github.com/golang/glog"
+
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/client-go/tools/cache"
+
+	routeapi "github.com/openshift/origin/pkg/route/apis/route"
+)
+
+// ContentionTracker records modifications to a particular entry to prevent endless
+// loops when multiple routers are configured with conflicting info. A given router
+// process tracks whether the ingress status is change from a correct value to any
+// other value (by invoking IsContended when the state has diverged).
+type ContentionTracker interface {
+	// IsContended should be invoked when the state of the object in storage differs
+	// from the desired state. It will return true if the provided id was recently
+	// reset from the correct state to an incorrect state. The current ingress is the
+	// expected state of the object at this time and may be used by the tracker to
+	// determine if the most recent update was a contention. This method does not
+	// update the state of the tracker.
+	IsChangeContended(id string, now time.Time, current *routeapi.RouteIngress) bool
+	// Clear informs the tracker that the provided ingress state was confirmed to
+	// match the current state of this process. If a subsequent call to IsChangeContended
+	// is made within the expiration window, the object will be considered as contended.
+	Clear(id string, current *routeapi.RouteIngress)
+}
+
+type elementState int
+
+const (
+	stateCandidate elementState = iota
+	stateContended
+)
+
+type trackerElement struct {
+	at    time.Time
+	state elementState
+	last  *routeapi.RouteIngress
+}
+
+// SimpleContentionTracker tracks whether a given identifier is changed from a correct
+// state (set by Clear) to an incorrect state (inferred by calling IsContended).
+type SimpleContentionTracker struct {
+	informer   cache.SharedInformer
+	routerName string
+
+	expires time.Duration
+	// maxContentions is the number of contentions detected before the entire
+	maxContentions int
+	message        string
+
+	lock        sync.Mutex
+	contentions int
+	ids         map[string]trackerElement
+}
+
+// NewSimpleContentionTracker creates a ContentionTracker that will prevent writing
+// to the same route more often than once per interval. A background process will
+// periodically flush old entries (at twice interval) in order to prevent the list
+// growing unbounded if routes are created and deleted frequently. The informer
+// detects changes to ingress records for routerName and will advance the tracker
+// state from candidate to contended if the host, wildcardPolicy, or canonical host
+// name fields are repeatedly updated.
+func NewSimpleContentionTracker(informer cache.SharedInformer, routerName string, interval time.Duration) *SimpleContentionTracker {
+	return &SimpleContentionTracker{
+		informer:       informer,
+		routerName:     routerName,
+		expires:        interval,
+		maxContentions: 5,
+
+		ids: make(map[string]trackerElement),
+	}
+}
+
+// SetConflictMessage will print message whenever contention with another writer
+// is detected.
+func (t *SimpleContentionTracker) SetConflictMessage(message string) {
+	t.lock.Lock()
+	defer t.lock.Unlock()
+	t.message = message
+}
+
+// Run starts the background cleanup process for expired items.
+func (t *SimpleContentionTracker) Run(stopCh <-chan struct{}) {
+	// Watch the informer for changes to the route ingress we care about (identified
+	// by router name) and if we see it change remember it. This loop can process
+	// changes to routes faster than the router, which means it has more up-to-date
+	// contention info and can detect contention while the main controller is still
+	// syncing.
+	t.informer.AddEventHandler(cache.ResourceEventHandlerFuncs{
+		UpdateFunc: func(oldObj, obj interface{}) {
+			oldRoute, ok := oldObj.(*routeapi.Route)
+			if !ok {
+				return
+			}
+			route, ok := obj.(*routeapi.Route)
+			if !ok {
+				return
+			}
+			if ingress := ingressChanged(oldRoute, route, t.routerName); ingress != nil {
+				t.Changed(string(route.UID), ingress)
+			}
+		},
+	})
+
+	// periodically clean up expired changes
+	ticker := time.NewTicker(t.expires * 2)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-stopCh:
+			return
+		case <-ticker.C:
+			t.flush()
+		}
+	}
+}
+
+func (t *SimpleContentionTracker) flush() {
+	t.lock.Lock()
+	defer t.lock.Unlock()
+
+	// reset conflicts every expiration interval, but remove tracking info less often
+	contentionExpiration := nowFn().Add(-t.expires)
+	trackerExpiration := contentionExpiration.Add(-2 * t.expires)
+
+	removed := 0
+	contentions := 0
+	for id, last := range t.ids {
+		switch last.state {
+		case stateContended:
+			if last.at.Before(contentionExpiration) {
+				delete(t.ids, id)
+				removed++
+				continue
+			}
+			contentions++
+		default:
+			if last.at.Before(trackerExpiration) {
+				delete(t.ids, id)
+				removed++
+				continue
+			}
+		}
+	}
+	if t.contentions > 0 && len(t.message) > 0 {
+		glog.Warning(t.message)
+	}
+	glog.V(5).Infof("Flushed contention tracker (%s): %d out of %d removed, %d total contentions", t.expires*2, removed, removed+len(t.ids), t.contentions)
+	t.contentions = contentions
+}
+
+// Changed records that a change to an ingress value was detected. This is called from
+// a separate goroutine and may have seen newer events than the current route controller
+// plugins, so we don't do direct time comparisons. Instead we count edge transitions on
+// a given id.
+func (t *SimpleContentionTracker) Changed(id string, current *routeapi.RouteIngress) {
+	t.lock.Lock()
+	defer t.lock.Unlock()
+
+	// we have detected a sufficient number of conflicts to skip all updates for this interval
+	if t.contentions > t.maxContentions {
+		glog.V(4).Infof("Reached max contentions, stop tracking changes")
+		return
+	}
+
+	// if we have never recorded this object
+	last, ok := t.ids[id]
+	if !ok {
+		t.ids[id] = trackerElement{
+			at:    nowFn().Time,
+			state: stateCandidate,
+			last:  current,
+		}
+		glog.V(4).Infof("Object %s is a candidate for contention", id)
+		return
+	}
+
+	// the previous state matches the current state, nothing to do
+	if ingressEqual(last.last, current) {
+		glog.V(4).Infof("Object %s is unchanged", id)
+		return
+	}
+
+	if last.state == stateContended {
+		t.contentions++
+		glog.V(4).Infof("Object %s is contended and has been modified by another writer", id)
+		return
+	}
+
+	// if it appears that the state is being changed by another party, mark it as contended
+	if last.state == stateCandidate {
+		t.ids[id] = trackerElement{
+			at:    nowFn().Time,
+			state: stateContended,
+			last:  current,
+		}
+		t.contentions++
+		glog.V(4).Infof("Object %s has been modified by another writer", id)
+		return
+	}
+}
+
+func (t *SimpleContentionTracker) IsChangeContended(id string, now time.Time, current *routeapi.RouteIngress) bool {
+	t.lock.Lock()
+	defer t.lock.Unlock()
+
+	// we have detected a sufficient number of conflicts to skip all updates for this interval
+	if t.contentions > t.maxContentions {
+		glog.V(4).Infof("Reached max contentions, rejecting all update attempts until the next interval")
+		return true
+	}
+
+	// if we have expired or never recorded this object
+	last, ok := t.ids[id]
+	if !ok || last.at.Add(t.expires).Before(now) {
+		return false
+	}
+
+	// if the object is contended, exit early
+	if last.state == stateContended {
+		glog.V(4).Infof("Object %s is being contended by another writer", id)
+		return true
+	}
+
+	return false
+}
+
+func (t *SimpleContentionTracker) Clear(id string, current *routeapi.RouteIngress) {
+	t.lock.Lock()
+	defer t.lock.Unlock()
+
+	last, ok := t.ids[id]
+	if !ok {
+		return
+	}
+	last.last = current
+	last.state = stateCandidate
+	t.ids[id] = last
+}
+
+func ingressEqual(a, b *routeapi.RouteIngress) bool {
+	return a.Host == b.Host && a.RouterCanonicalHostname == b.RouterCanonicalHostname && a.WildcardPolicy == b.WildcardPolicy && a.RouterName == b.RouterName
+}
+
+func ingressConditionTouched(ingress *routeapi.RouteIngress) *metav1.Time {
+	var lastTouch *metav1.Time
+	for _, condition := range ingress.Conditions {
+		if t := condition.LastTransitionTime; t != nil {
+			switch {
+			case lastTouch == nil, t.After(lastTouch.Time):
+				lastTouch = t
+			}
+		}
+	}
+	return lastTouch
+}
+
+func ingressChanged(oldRoute, route *routeapi.Route, routerName string) *routeapi.RouteIngress {
+	var ingress *routeapi.RouteIngress
+	for i := range route.Status.Ingress {
+		if route.Status.Ingress[i].RouterName == routerName {
+			ingress = &route.Status.Ingress[i]
+			for _, old := range oldRoute.Status.Ingress {
+				if old.RouterName == routerName {
+					if !ingressEqual(ingress, &old) {
+						return ingress
+					}
+					return nil
+				}
+			}
+			return nil
+		}
+	}
+	return nil
+}