built from CA v1.29.4

ADOBE_CHANGELOG.md

@@ -0,0 +1,102 @@
+# About
+
+We maintain a set of patches on top of upstream K8s autoscaler. They need to be
+semantically versioned as well so we understand, for example, that we're
+upgrading a fix version of autoscaler vs. a fix version of our patches.
+
+The scheme is quite simple. The product is first, followed by the upstream
+version, followed by our version:
+
+```
+Product is cluster-autoscaler or vertical-pod-autoscaler or addon-resizer
+K stands for K8s upstream version
+A stands for Adobe's changes
+
+{Product}-{K major}.{K minor}.{K fix}-{A major}.{A minor}.{A fix}-adobe
+```
+
+# Log
+
+## cluster-autoscaler-1.29.4-1.6.0-adobe
+- built from CA v1.29.4
+
+## cluster-autoscaler-1.28.6-1.6.0-adobe
+- built from CA v1.28.6
+
+## cluster-autoscaler-1.27.7-1.6.0-adobe
+- built from CA v1.27.7
+
+## cluster-autoscaler-1.27.5-1.6.0-adobe
+- built from CA v1.27.5
+
+## cluster-autoscaler-1.26.4-1.6.0-adobe
+- built from CA v1.26.4
+- backport https://github.com/kubernetes/autoscaler/pull/5550
+
+## cluster-autoscaler-1.25.2-1.6.0-adobe
+- backport https://github.com/kubernetes/autoscaler/pull/5550
+
+## cluster-autoscaler-1.25.2-1.5.0-adobe
+- built from CA v1.25.2
+
+## cluster-autoscaler-1.24.2-1.6.0-adobe
+- backport https://github.com/kubernetes/autoscaler/pull/5550
+
+## cluster-autoscaler-1.24.2-1.5.0-adobe
+- built from CA v1.24.2
+
+## cluster-autoscaler-1.24.1-1.5.0-adobe
+- built from CA v1.24.1
+
+## cluster-autoscaler-1.23.1-1.5.0-adobe
+- skip deletion of unregistered nodes if there are more than 50% of discovered nodes
+
+## cluster-autoscaler-1.23.1-1.4.0-adobe
+- built from CA v1.23.1
+- backport https://github.com/kubernetes/autoscaler/pull/4970
+- remove code from 1.0.0-adobe and 1.2.0-adobe
+
+## cluster-autoscaler-1.22.3-1.3.0-adobe
+- built from CA v1.22.3
+- backport https://github.com/kubernetes/autoscaler/pull/4970
+
+## cluster-autoscaler-1.22.2-1.2.1-adobe
+- built from CA v1.22.2
+
+## cluster-autoscaler-1.21.2-1.2.1-adobe
+- built from CA v1.21.2
+
+## cluster-autoscaler-1.20.1-1.2.1-adobe
+- also use `topology.kubernetes.io/zone` when evaluating topology spread constraints
+
+## cluster-autoscaler-1.20.1-1.2.0-adobe
+- improve scale out time when using topology spread constraints
+
+## cluster-autoscaler-1.20.1-1.1.0-adobe
+- upgraded ca
+
+## cluster-autoscaler-1.20.0-1.1.0-adobe
+- fix uneven scaling with topology contraints
+- backport https://github.com/kubernetes/autoscaler/pull/3883
+
+## cluster-autoscaler-1.20.0-1.0.0-adobe
+- on scale-down, evict DaemonSet pods after all the other pods
+
+# Procedure
+
+```bash
+# Assuming cluster-autoscaler-1.21-adobe is the latest "Adobe" release, built
+# "on top" of of 1.21.0, and we want to upgrade to 1.21.2
+git checkout cluster-autoscaler-1.21-adobe
+# Reset to the upstream version we originally built on
+git reset cluster-autoscaler-1.21.0
+# Stash "Adobe" changes
+git stash
+# Get the latest stuff from upstream
+git fetch upstream
+# Update the branch to upstream 1.21.2
+git merge -s recursive -X theirs cluster-autoscaler-1.21.2
+# Apply "Adobe" changes, on top of 1.21.2
+git stash pop
+# Resolve conflicts if needed, then git add/git commit all the files (including this one)
+```

cluster-autoscaler/clusterstate/clusterstate_adobe.go

@@ -0,0 +1,8 @@
+package clusterstate
+
+// ClearUnregisteredNodes clears the unregistered node cache
+func (csr *ClusterStateRegistry) ClearUnregisteredNodes() {
+	csr.Lock()
+	defer csr.Unlock()
+	csr.unregisteredNodes = make(map[string]UnregisteredNode)
+}

cluster-autoscaler/core/static_autoscaler.go

@@ -411,7 +411,14 @@ func (a *StaticAutoscaler) RunOnce(currentTime time.Time) caerrors.AutoscalerErr
 	// Check if there are any nodes that failed to register in Kubernetes
 	// master.
 	unregisteredNodes := a.clusterStateRegistry.GetUnregisteredNodes()
-	if len(unregisteredNodes) > 0 {
+	if float64(len(unregisteredNodes)) > float64(len(allNodes))*50.0/100.0 {
+		klog.Warningf("%d unregistered nodes present - skipping deletion (only %d nodes discovered)", len(unregisteredNodes), len(allNodes))
+		// clear the state so that the UnregisteredSince timestamp is reset
+		a.clusterStateRegistry.ClearUnregisteredNodes()
+		for _, n := range unregisteredNodes {
+			klog.V(0).Infof("--> %v", n.Node.GetObjectMeta().GetName())
+		}
+	} else if len(unregisteredNodes) > 0 {
 		klog.V(1).Infof("%d unregistered nodes present", len(unregisteredNodes))
 		removedAny, err := a.removeOldUnregisteredNodes(unregisteredNodes, autoscalingContext,
 			a.clusterStateRegistry, currentTime, autoscalingContext.LogRecorder)
@@ -422,6 +429,8 @@ func (a *StaticAutoscaler) RunOnce(currentTime time.Time) caerrors.AutoscalerErr
 		if removedAny {
 			klog.V(0).Infof("Some unregistered nodes were removed")
 		}
+	} else {
+		klog.V(0).Infof("noop")
 	}
 
 	if !a.clusterStateRegistry.IsClusterHealthy() {