Sok3t/descheduler
1
0
Fork 0
mirror of https://github.com/kubernetes-sigs/descheduler.git synced 2026-01-28 22:57:35 +01:00
Code Packages Releases Activity

PodEvictor: turn an exceeded limit into an error

Browse Source 
When checking for node limit getting exceeded the pod eviction
never fails. Thus, ignoring the metric reporting when a pod fails
to be evicted due to node limit constrains.

The error also allows plugin to react on other limits getting
exceeded. E.g. the limit on the number of pods evicted per namespace.
This commit is contained in:
Jan Chaloupka
2024-07-06 20:02:55 +02:00
parent 7657345079
commit 18d0e4a540
17 changed files with 182 additions and 104 deletions
Download Patch File Download Diff File Expand all files Collapse all files

68
pkg/framework/plugins/nodeutilization/nodeutilization.go
View File

@@ -311,42 +311,48 @@ func evictPods(
continue
}
if preEvictionFilterWithOptions(pod) {
if podEvictor.Evict(ctx, pod, evictOptions) {
klog.V(3).InfoS("Evicted pods", "pod", klog.KObj(pod))
if !preEvictionFilterWithOptions(pod) {
continue
}
err = podEvictor.Evict(ctx, pod, evictOptions)
if err == nil {
klog.V(3).InfoS("Evicted pods", "pod", klog.KObj(pod))
for name := range totalAvailableUsage {
if name == v1.ResourcePods {
nodeInfo.usage[name].Sub(*resource.NewQuantity(1, resource.DecimalSI))
totalAvailableUsage[name].Sub(*resource.NewQuantity(1, resource.DecimalSI))
} else {
quantity := utils.GetResourceRequestQuantity(pod, name)
nodeInfo.usage[name].Sub(quantity)
totalAvailableUsage[name].Sub(quantity)
}
}
keysAndValues := []interface{}{
"node", nodeInfo.node.Name,
"CPU", nodeInfo.usage[v1.ResourceCPU].MilliValue(),
"Mem", nodeInfo.usage[v1.ResourceMemory].Value(),
"Pods", nodeInfo.usage[v1.ResourcePods].Value(),
}
for name := range totalAvailableUsage {
if !nodeutil.IsBasicResource(name) {
keysAndValues = append(keysAndValues, string(name), totalAvailableUsage[name].Value())
}
}
klog.V(3).InfoS("Updated node usage", keysAndValues...)
// check if pods can be still evicted
if !continueEviction(nodeInfo, totalAvailableUsage) {
break
for name := range totalAvailableUsage {
if name == v1.ResourcePods {
nodeInfo.usage[name].Sub(*resource.NewQuantity(1, resource.DecimalSI))
totalAvailableUsage[name].Sub(*resource.NewQuantity(1, resource.DecimalSI))
} else {
quantity := utils.GetResourceRequestQuantity(pod, name)
nodeInfo.usage[name].Sub(quantity)
totalAvailableUsage[name].Sub(quantity)
}
}
keysAndValues := []interface{}{
"node", nodeInfo.node.Name,
"CPU", nodeInfo.usage[v1.ResourceCPU].MilliValue(),
"Mem", nodeInfo.usage[v1.ResourceMemory].Value(),
"Pods", nodeInfo.usage[v1.ResourcePods].Value(),
}
for name := range totalAvailableUsage {
if !nodeutil.IsBasicResource(name) {
keysAndValues = append(keysAndValues, string(name), totalAvailableUsage[name].Value())
}
}
klog.V(3).InfoS("Updated node usage", keysAndValues...)
// check if pods can be still evicted
if !continueEviction(nodeInfo, totalAvailableUsage) {
break
}
continue
}
if podEvictor.NodeLimitExceeded(nodeInfo.node) {
switch err.(type) {
case *evictions.EvictionNodeLimitError:
return
default:
klog.Errorf("eviction failed: %v", err)
}
}
}

12
pkg/framework/plugins/podlifetime/pod_lifetime.go
View File

@@ -131,9 +131,17 @@ func (d *PodLifeTime) Deschedule(ctx context.Context, nodes []*v1.Node) *framewo
// in the event that PDB or settings such maxNoOfPodsToEvictPer* prevent too much eviction
podutil.SortPodsBasedOnAge(podsToEvict)
loop:
for _, pod := range podsToEvict {
if !d.handle.Evictor().NodeLimitExceeded(nodeMap[pod.Spec.NodeName]) {
d.handle.Evictor().Evict(ctx, pod, evictions.EvictOptions{StrategyName: PluginName})
err := d.handle.Evictor().Evict(ctx, pod, evictions.EvictOptions{StrategyName: PluginName})
if err == nil {
continue
}
switch err.(type) {
case *evictions.EvictionNodeLimitError:
continue loop
default:
klog.Errorf("eviction failed: %v", err)
}
}

10
pkg/framework/plugins/removeduplicates/removeduplicates.go
View File

@@ -210,9 +210,15 @@ func (r *RemoveDuplicates) Balance(ctx context.Context, nodes []*v1.Node) *frame
// It's assumed all duplicated pods are in the same priority class
// TODO(jchaloup): check if the pod has a different node to lend to
for _, pod := range pods[upperAvg-1:] {
r.handle.Evictor().Evict(ctx, pod, evictions.EvictOptions{StrategyName: PluginName})
if r.handle.Evictor().NodeLimitExceeded(nodeMap[nodeName]) {
err := r.handle.Evictor().Evict(ctx, pod, evictions.EvictOptions{StrategyName: PluginName})
if err == nil {
continue
}
switch err.(type) {
case *evictions.EvictionNodeLimitError:
continue loop
default:
klog.Errorf("eviction failed: %v", err)
}
}
}

13
pkg/framework/plugins/removefailedpods/failedpods.go
View File

@@ -102,10 +102,17 @@ func (d *RemoveFailedPods) Deschedule(ctx context.Context, nodes []*v1.Node) *fr
}
}
totalPods := len(pods)
loop:
for i := 0; i < totalPods; i++ {
d.handle.Evictor().Evict(ctx, pods[i], evictions.EvictOptions{StrategyName: PluginName})
if d.handle.Evictor().NodeLimitExceeded(node) {
break
err := d.handle.Evictor().Evict(ctx, pods[i], evictions.EvictOptions{StrategyName: PluginName})
if err == nil {
continue
}
switch err.(type) {
case *evictions.EvictionNodeLimitError:
break loop
default:
klog.Errorf("eviction failed: %v", err)
}
}
}

13
pkg/framework/plugins/removepodshavingtoomanyrestarts/toomanyrestarts.go
View File

@@ -122,10 +122,17 @@ func (d *RemovePodsHavingTooManyRestarts) Deschedule(ctx context.Context, nodes
}
}
totalPods := len(pods)
loop:
for i := 0; i < totalPods; i++ {
d.handle.Evictor().Evict(ctx, pods[i], evictions.EvictOptions{StrategyName: PluginName})
if d.handle.Evictor().NodeLimitExceeded(node) {
break
err := d.handle.Evictor().Evict(ctx, pods[i], evictions.EvictOptions{StrategyName: PluginName})
if err == nil {
continue
}
switch err.(type) {
case *evictions.EvictionNodeLimitError:
break loop
default:
klog.Errorf("eviction failed: %v", err)
}
}
}

13
pkg/framework/plugins/removepodsviolatinginterpodantiaffinity/pod_antiaffinity.go
View File

@@ -98,7 +98,8 @@ loop:
for i := 0; i < totalPods; i++ {
if utils.CheckPodsWithAntiAffinityExist(pods[i], podsInANamespace, nodeMap) {
if d.handle.Evictor().Filter(pods[i]) && d.handle.Evictor().PreEvictionFilter(pods[i]) {
if d.handle.Evictor().Evict(ctx, pods[i], evictions.EvictOptions{StrategyName: PluginName}) {
err := d.handle.Evictor().Evict(ctx, pods[i], evictions.EvictOptions{StrategyName: PluginName})
if err == nil {
// Since the current pod is evicted all other pods which have anti-affinity with this
// pod need not be evicted.
// Update allPods.
@@ -106,12 +107,16 @@ loop:
pods = append(pods[:i], pods[i+1:]...)
i--
totalPods--
continue
}
switch err.(type) {
case *evictions.EvictionNodeLimitError:
continue loop
default:
klog.Errorf("eviction failed: %v", err)
}
}
}
if d.handle.Evictor().NodeLimitExceeded(node) {
continue loop
}
}
}
return nil

13
pkg/framework/plugins/removepodsviolatingnodeaffinity/node_affinity.go
View File

@@ -134,11 +134,18 @@ func (d *RemovePodsViolatingNodeAffinity) processNodes(ctx context.Context, node
}
}
loop:
for _, pod := range pods {
klog.V(1).InfoS("Evicting pod", "pod", klog.KObj(pod))
d.handle.Evictor().Evict(ctx, pod, evictions.EvictOptions{StrategyName: PluginName})
if d.handle.Evictor().NodeLimitExceeded(node) {
break
err := d.handle.Evictor().Evict(ctx, pod, evictions.EvictOptions{StrategyName: PluginName})
if err == nil {
continue
}
switch err.(type) {
case *evictions.EvictionNodeLimitError:
break loop
default:
klog.Errorf("eviction failed: %v", err)
}
}
}

13
pkg/framework/plugins/removepodsviolatingnodetaints/node_taint.go
View File

@@ -114,6 +114,7 @@ func (d *RemovePodsViolatingNodeTaints) Deschedule(ctx context.Context, nodes []
}
}
totalPods := len(pods)
loop:
for i := 0; i < totalPods; i++ {
if !utils.TolerationsTolerateTaintsWithFilter(
pods[i].Spec.Tolerations,
@@ -121,9 +122,15 @@ func (d *RemovePodsViolatingNodeTaints) Deschedule(ctx context.Context, nodes []
d.taintFilterFnc,
) {
klog.V(2).InfoS("Not all taints with NoSchedule effect are tolerated after update for pod on node", "pod", klog.KObj(pods[i]), "node", klog.KObj(node))
d.handle.Evictor().Evict(ctx, pods[i], evictions.EvictOptions{StrategyName: PluginName})
if d.handle.Evictor().NodeLimitExceeded(node) {
break
err := d.handle.Evictor().Evict(ctx, pods[i], evictions.EvictOptions{StrategyName: PluginName})
if err == nil {
continue
}
switch err.(type) {
case *evictions.EvictionNodeLimitError:
break loop
default:
klog.Errorf("eviction failed: %v", err)
}
}
}

14
pkg/framework/plugins/removepodsviolatingtopologyspreadconstraint/topologyspreadconstraint.go
View File

@@ -235,10 +235,16 @@ func (d *RemovePodsViolatingTopologySpreadConstraint) Balance(ctx context.Contex
}
if d.handle.Evictor().PreEvictionFilter(pod) {
d.handle.Evictor().Evict(ctx, pod, evictions.EvictOptions{StrategyName: PluginName})
}
if d.handle.Evictor().NodeLimitExceeded(nodeMap[pod.Spec.NodeName]) {
nodeLimitExceeded[pod.Spec.NodeName] = true
err := d.handle.Evictor().Evict(ctx, pod, evictions.EvictOptions{StrategyName: PluginName})
if err == nil {
continue
}
switch err.(type) {
case *evictions.EvictionNodeLimitError:
nodeLimitExceeded[pod.Spec.NodeName] = true
default:
klog.Errorf("eviction failed: %v", err)
}
}
}