refactor affinity to use same startegy as flux operator (#59)

* refactor affinity to use same startegy as flux operator

Signed-off-by: vsoch <vsoch@users.noreply.github.com>

Author: vsoch

pkg/jobs/launcher.go

@@ -188,7 +188,7 @@ func (m *LauncherWorker) AddWorkers(
 ) (*jobset.ReplicatedJob, error) {
 
 	numWorkers := spec.Spec.Pods - 1
-	workers, err := metrics.GetReplicatedJob(spec, false, numWorkers, numWorkers, m.WorkerLetter)
+	workers, err := metrics.GetReplicatedJob(spec, false, numWorkers, numWorkers, m.WorkerLetter, m.SoleTenancy)
 	if err != nil {
 		return workers, err
 	}
@@ -221,7 +221,7 @@ func (m *LauncherWorker) ReplicatedJobs(spec *api.MetricSet) ([]jobset.Replicate
 	m.ensureDefaultNames()
 
 	// Generate a replicated job for the launcher (LauncherWorker) and workers
-	launcher, err := metrics.GetReplicatedJob(spec, false, 1, 1, m.LauncherLetter)
+	launcher, err := metrics.GetReplicatedJob(spec, false, 1, 1, m.LauncherLetter, m.SoleTenancy)
 	if err != nil {
 		return js, err
 	}

pkg/jobs/storage.go

@@ -47,6 +47,7 @@ func (m StorageGeneric) Description() string {
 	return m.Summary
 }
 
+// By default assume storage does not have sole tenancy
 func (m StorageGeneric) HasSoleTenancy() bool {
 	return false
 }

pkg/metrics/application.go

@@ -56,7 +56,7 @@ func GetApplicationReplicatedJobs(
 	m := (*metric)
 
 	// This defaults to one replicated job, named "m", no custom replicated job name, and sole tenancy false
-	job, err := GetReplicatedJob(spec, shareProcessNamespace, spec.Spec.Pods, spec.Spec.Completions, "")
+	job, err := GetReplicatedJob(spec, shareProcessNamespace, spec.Spec.Pods, spec.Spec.Completions, "", m.HasSoleTenancy())
 	if err != nil {
 		return rjs, err
 	}

pkg/metrics/jobset.go

@@ -28,6 +28,8 @@ var (
 	soleTenancyValue  = "sole-tenancy"
 )
 
+const podLabelAppName = "app.kubernetes.io/name"
+
 // GetJobSet is called by the controller to return some JobSet based
 // on the type: application, storage, or standalone
 func GetJobSet(
@@ -45,7 +47,7 @@ func GetJobSet(
 		successJobs := getSuccessJobs(set.Metrics())
 
 		// A base JobSet can hold one or more replicated jobs
-		js := getBaseJobSet(spec, successJobs, set.HasSoleTenancy())
+		js := getBaseJobSet(spec, successJobs)
 
 		// Get one or more replicated jobs, depending on the type
 		rjs, err := set.ReplicatedJobs(spec)
@@ -80,7 +82,7 @@ func getSuccessJobs(metrics []*Metric) []string {
 }
 
 // getBaseJobSet shared for either an application or isolated jobset
-func getBaseJobSet(set *api.MetricSet, successSet []string, soleTenancy bool) *jobset.JobSet {
+func getBaseJobSet(set *api.MetricSet, successSet []string) *jobset.JobSet {
 
 	// When suspend is true we have a hard time debugging jobs, so keep false
 	suspend := false
@@ -112,9 +114,6 @@ func getBaseJobSet(set *api.MetricSet, successSet []string, soleTenancy bool) *j
 	}
 
 	// Do we want to assign 1 node: 1 pod? We can use Pod Anti-affinity for that
-	if soleTenancy {
-		js.ObjectMeta.Annotations = map[string]string{jobset.ExclusiveKey: "kubernetes.io/hostname"}
-	}
 	return &js
 }
 
@@ -125,6 +124,7 @@ func GetReplicatedJob(
 	pods int32,
 	completions int32,
 	jobname string,
+	soleTenancy bool,
 ) (*jobset.ReplicatedJob, error) {
 
 	// Default replicated job name, if not set
@@ -183,6 +183,11 @@ func GetReplicatedJob(
 		},
 	}
 
+	// Do we want sole tenancy?
+	if soleTenancy {
+		jobspec.Template.Spec.Affinity = getAffinity(set)
+	}
+
 	// Do we have a pull secret for the application image?
 	if set.Spec.Application.PullSecret != "" {
 		jobspec.Template.Spec.ImagePullSecrets = []corev1.LocalObjectReference{
@@ -193,3 +198,51 @@ func GetReplicatedJob(
 	job.Template.Spec = jobspec
 	return &job, nil
 }
+
+// getAffinity returns to pod affinity to ensure 1 address / node
+func getAffinity(set *api.MetricSet) *corev1.Affinity {
+	return &corev1.Affinity{
+		// Prefer to schedule pods on the same zone
+		PodAffinity: &corev1.PodAffinity{
+			PreferredDuringSchedulingIgnoredDuringExecution: []corev1.WeightedPodAffinityTerm{
+				{
+					Weight: 100,
+					PodAffinityTerm: corev1.PodAffinityTerm{
+						LabelSelector: &metav1.LabelSelector{
+							MatchExpressions: []metav1.LabelSelectorRequirement{
+								{
+									// added in getPodLabels
+									Key:      podLabelAppName,
+									Operator: metav1.LabelSelectorOpIn,
+									Values:   []string{set.Name},
+								},
+							},
+						},
+						TopologyKey: "topology.kubernetes.io/zone",
+					},
+				},
+			},
+		},
+		// Prefer to schedule pods on different nodes
+		PodAntiAffinity: &corev1.PodAntiAffinity{
+			PreferredDuringSchedulingIgnoredDuringExecution: []corev1.WeightedPodAffinityTerm{
+				{
+					Weight: 100,
+					PodAffinityTerm: corev1.PodAffinityTerm{
+						LabelSelector: &metav1.LabelSelector{
+							MatchExpressions: []metav1.LabelSelectorRequirement{
+								{
+									// added in getPodLabels
+									Key:      podLabelAppName,
+									Operator: metav1.LabelSelectorOpIn,
+									Values:   []string{set.Name},
+								},
+							},
+						},
+						TopologyKey: "kubernetes.io/hostname",
+					},
+				},
+			},
+		},
+	}
+}

pkg/metrics/metricset.go

@@ -48,7 +48,6 @@ type MetricSet interface {
 	Metrics() []*Metric
 	EntrypointScripts(*api.MetricSet) []EntrypointScript
 	ReplicatedJobs(*api.MetricSet) ([]jobset.ReplicatedJob, error)
-	HasSoleTenancy() bool
 }
 
 // get an application default entrypoint, if not determined by metric

pkg/metrics/storage.go

@@ -22,7 +22,7 @@ func (m *StorageMetricSet) ReplicatedJobs(spec *api.MetricSet) ([]jobset.Replica
 
 	// Storage metrics do not need to share the process namespace
 	// The jobname empty string will use the default, no custom replicated job name, and sole tenancy false
-	job, err := GetReplicatedJob(spec, false, spec.Spec.Pods, spec.Spec.Completions, "")
+	job, err := GetReplicatedJob(spec, false, spec.Spec.Pods, spec.Spec.Completions, "", m.HasSoleTenancy())
 	if err != nil {
 		return rjs, err
 	}