Set node_readiness_label default to an empty value.

Previously, it was set to the lifecycle-status:ready, breaking a
lot of minikube deployments. Also it was not possible befor to run
with this label set to an empty value.

Document the effect of the label in the new section of the
documentation.

README.md

@@ -32,7 +32,7 @@ This project is currently in active development. It is however already [used int
 
 There is a talk about this project delivered by Josh Berkus on KubeCon 2017: [Kube-native Postgres](https://www.youtube.com/watch?v=Zn1vd7sQ_bc)
 
-Please, report any issues discovered to https://github.com/zalando-incubator/postgres-operator/issues.
+Please, report any issues discovered to httnodeps://github.com/zalando-incubator/postgres-operator/issues.
 
 ## Running and testing the operator
 
@@ -148,6 +148,30 @@ spec:
 Please be aware that the taint and toleration only ensures that no other pod gets scheduled to a PostgreSQL node 
 but not that PostgreSQL pods are placed on such a node. This can be achieved by setting a node affinity rule in the ConfigMap.
 
+### Using the operator to minimize the amount of failovers during the cluster upgrade
+
+To make sure there is only one failover per each PostgreSQL deployment when nodes of the cluster are decomissioned
+during the cluster upgrade, the operator relies on the `node_readiness_label` parameter. It is used in order to determine whether
+a given node is considered ready to run postgres pods. When a node is both `unschedulable` and doesn't have a readiness
+label the operator starts moving master pods out of it. In addition, the operator sets the `PodDisruptionBudget` for the
+postgres pods to make sure pods with the role `master` are not killed by the cluster scale out operations. Each postgres
+pod has a `nodeAffinity` set to avoid being scheduled on the nodes without the readiness label. Together, those measures
+give us the following guarantees:
+
+* no postgres pod runs on the node being not ready
+* no postgres master pod is killed during the scale event before the operator has a chance to move it off the node
+* when a postgres pod is killed by the operator because the node is decomissioned, it will not respawn on another node to-be-decomissioned.
+
+By default the `node_readiness_label` is set to an empty string, disabling this feature altogether. It can be set to a
+string containing one or more key:value parameters, i.e:
+```
+node_readiness_label: "lifecycle-status:ready,disagnostic-checks:ok"
+
+```
+
+when multiple labels are set the operator will require all of them to be present on a node (and set to the specified value) in order to consider
+the node as ready. 
+
 #### Custom Pod Environment Variables
 
 It is possible to configure a config map which is used by the Postgres pods as an additional provider for environment variables.

manifests/minimal-postgres-manifest.yaml

@@ -1,7 +1,7 @@
 apiVersion: "acid.zalan.do/v1"
 kind: postgresql
 metadata:
-  name: acid-minimal-cluster
+  name: acid-minimal-cluster-recent
 spec:
   teamId: "ACID"
   volume:
@@ -21,3 +21,4 @@ spec:
     foo: zalando
   postgresql:
     version: "10"
+  dockerImage: registry.opensource.zalan.do/acid/spilotest-10:improve-support-for-namespaces-v1

pkg/cluster/k8sres.go

@@ -238,6 +238,9 @@ PatroniInitDBParams:
 
 func (c *Cluster) nodeAffinity() *v1.Affinity {
 	matchExpressions := make([]v1.NodeSelectorRequirement, 0)
+	if len(c.OpConfig.NodeReadinessLabel) == 0 {
+		return nil
+	}
 	for k, v := range c.OpConfig.NodeReadinessLabel {
 		matchExpressions = append(matchExpressions, v1.NodeSelectorRequirement{
 			Key:      k,
@@ -431,10 +434,13 @@ func (c *Cluster) generatePodTemplate(
 		ServiceAccountName:            c.OpConfig.ServiceAccountName,
 		TerminationGracePeriodSeconds: &terminateGracePeriodSeconds,
 		Containers:                    []v1.Container{container},
-		Affinity:                      c.nodeAffinity(),
 		Tolerations:                   c.tolerations(tolerationsSpec),
 	}
 
+	if affinity := c.nodeAffinity(); affinity != nil {
+		podSpec.Affinity = affinity
+	}
+
 	if c.OpConfig.ScalyrAPIKey != "" && c.OpConfig.ScalyrImage != "" {
 		podSpec.Containers = append(
 			podSpec.Containers,

pkg/util/config/config.go

@@ -32,7 +32,7 @@ type Resources struct {
 	DefaultCPULimit         string            `name:"default_cpu_limit" default:"3"`
 	DefaultMemoryLimit      string            `name:"default_memory_limit" default:"1Gi"`
 	PodEnvironmentConfigMap string            `name:"pod_environment_configmap" default:""`
-	NodeReadinessLabel      map[string]string `name:"node_readiness_label" default:"lifecycle-status:ready"`
+	NodeReadinessLabel      map[string]string `name:"node_readiness_label" default:""`
 	MaxInstances            int32             `name:"max_instances" default:"-1"`
 	MinInstances            int32             `name:"min_instances" default:"-1"`
 }