Merge branch 'master' into rolling_updates_with_statefulset_annotations

# Conflicts:
#	pkg/cluster/k8sres.go
This commit is contained in:
Oleksii Kliukin 2018-05-15 15:33:34 +02:00
commit 0c616a802f
11 changed files with 212 additions and 102 deletions

View File

@ -6,32 +6,32 @@
The Postgres operator manages PostgreSQL clusters on Kubernetes using the [operator pattern](https://coreos.com/blog/introducing-operators.html).
During the initial run it registers the [Custom Resource Definition (CRD)](https://kubernetes.io/docs/concepts/api-extension/custom-resources/#customresourcedefinitions) for Postgres.
The PostgreSQL CRD is essentially the schema that describes the contents of the manifests for deploying individual
PostgreSQL clusters using StatefulSets and [Patroni](https://github.com/zalando/patroni).
The `postgresql` CRD is essentially the schema that describes the contents of the manifests for deploying individual
Postgres clusters using StatefulSets and [Patroni](https://github.com/zalando/patroni).
Once the operator is running, it performs the following actions:
* watches for new PostgreSQL cluster manifests and deploys corresponding clusters
* watches for new `postgresql` manifests and deploys new clusters
* watches for updates to existing manifests and changes corresponding properties of the running clusters
* watches for deletes of the existing manifests and deletes corresponding clusters
* acts on an update to the operator definition itself and changes the running clusters when necessary
(i.e. when the docker image inside the operator definition has been updated)
* periodically checks running clusters against the manifests and acts on the differences found
* acts on an update to the operator configuration itself and changes the running clusters when necessary
(i.e. the Docker image changes for a minor release update)
* periodically checks running clusters against the manifests and syncs changes
For instance, when the user creates a new custom object of type ``postgresql`` by submitting a new manifest with
``kubectl``, the operator fetches that object and creates the corresponding Kubernetes structures
(StatefulSets, Services, Secrets) according to its definition.
Example: When a user creates a new custom object of type ``postgresql`` by submitting a new manifest with
``kubectl``, the operator fetches that object and creates the required Kubernetes entities to spawn a new Postgres cluster
(StatefulSets, Services, Secrets).
Another example is changing the docker image inside the operator. In this case, the operator first goes to all StatefulSets
it manages and updates them with the new docker images; afterwards, all pods from each StatefulSet are killed one by one
(rolling upgrade) and the replacements are spawned automatically by each StatefulSet with the new docker image.
Update example: After changing the Docker image inside the operator's configuration, the operator first goes to all StatefulSets
it manages and updates them with the new Docker image; afterwards, all pods from each StatefulSet are killed one by one
and the replacements are spawned automatically by each StatefulSet with the new Docker image. This is called the Rolling update.
## Scope
The scope of the postgres operator is on provisioning, modifying configuration and cleaning up Postgres clusters that use Patroni, basically to make it easy and convenient to run Patroni based clusters on Kubernetes.
The provisioning and modifying includes Kubernetes resources on one side but also e.g. database and role provisioning once the cluster is up and running.
We try to leave as much work as possible to Kubernetes and to Patroni where it fits, especially the cluster bootstrap and high availability.
The operator is however involved in some overarching orchestration, like rolling upgrades to improve the user experience.
The operator is however involved in some overarching orchestration, like rolling updates to improve the user experience.
Monitoring of clusters is not in scope, for this good tools already exist from ZMON to Prometheus and more Postgres specific options.
@ -147,9 +147,9 @@ We can use the generated secret of the `postgres` robot user to connect to our `
The `manifests/operator-rbac.yaml` defines cluster roles and bindings needed for the operator to function under access control restrictions. To deploy the operator with this RBAC policy use:
```bash
kubectl create -f manifests/configmap.yaml
kubectl create -f manifests/configmap.yaml
kubectl create -f manifests/operator-rbac.yaml
kubectl create -f manifests/postgres-operator.yaml
kubectl create -f manifests/postgres-operator.yaml
kubectl create -f manifests/minimal-postgres-manifest.yaml
```
@ -158,7 +158,7 @@ the `operator` default that is created in the `serviceaccount.yaml`. So you will
This is done intentionally, as to avoid breaking those setups that
already work with the default `operator` account. In the future the operator should ideally be run under the
`zalando-postgres-operator` service account.
`zalando-postgres-operator` service account.
The service account defined in `operator-rbac.yaml` acquires some privileges not really
used by the operator (i.e. we only need list and watch on configmaps),
@ -274,13 +274,23 @@ As a preventive measure, one can restrict the minimum and the maximum number of
If either `min_instances` or `max_instances` is set to a non-zero value, the operator may adjust the number of instances specified in the cluster manifest to match either the min or the max boundary.
For instance, of a cluster manifest has 1 instance and the min_instances is set to 3, the cluster will be created with 3 instances. By default, both parameters are set to -1.
### Load balancers
### Load balancers
For any Postgresql/Spilo cluster an operator creates two separate k8s services: one for the master pod and one for replica pods. To expose these services to an outer network, one can attach load balancers to them by setting `enableMasterLoadBalancer` and/or `enableReplicaLoadBalancer` to `true` in the cluster manifest. In the case any of these variables is omitted from the manifest, the operator configmap's settings `enable_master_load_balancer` and `enable_replica_load_balancer` apply. Note that the operator settings affect all Postgresql services running in a namespace watched by the operator.
For any Postgresql/Spilo cluster, the operator creates two separate k8s services: one for the master pod and one for
replica pods. To expose these services to an outer network, one can attach load balancers to them by setting
`enableMasterLoadBalancer` and/or `enableReplicaLoadBalancer` to `true` in the cluster manifest. In the case any of
these variables are omitted from the manifest, the operator configmap's settings `enable_master_load_balancer` and
`enable_replica_load_balancer` apply. Note that the operator settings affect all Postgresql services running in a
namespace watched by the operator.
For backward compatibility with already configured clusters we maintain in a cluster manifest older parameter names, namely `useLoadBalancer` for enabling the master service's load balancer and `replicaLoadBalancer` for the replica service. If set, these params take precedence over the newer `enableMasterLoadBalancer` and `enableReplicaLoadBalancer`. Note that in older versions of the operator (before PR #258) `replicaLoadBalancer` was responsible for both creating the replica service and attaching an LB to it; now the service is always created (since k8s service typically is free in the cloud setting), and this param only attaches an LB (that typically costs money).
###### Deprecated parameters
For the same reason of compatibility, we maintain the `enable_load_balancer` setting in the operator config map that was previously used to attach a LB to the master service. Its value is examined after the deprecated `useLoadBalancer` setting from the Postgresql manifest but before the recommended `enableMasterLoadBalancer`. There is no equivalent option for the replica service since the service used to be always created with a load balancer.
Parameters `useLoadBalancer` and `replicaLoadBalancer` in the PostgreSQL manifest are deprecated. To retain
compatibility with the old manifests they take affect in the absense of new `enableMasterLoadBalancer` and
`enableReplicaLoadBalancer` parameters (that is, if either of the new ones is present - all deprecated parameters are
ignored). The operator configuration parameter `enable_load_balancer` is ignored in all cases.
`
# Setup development environment

View File

@ -2,26 +2,41 @@ apiVersion: v1
kind: ConfigMap
metadata:
name: postgres-operator
data:
# the env var with the same name in the operator pod may overwrite this value
# if neither is set or evaluates to the empty string, listen to the operator's own namespace
data:
# if set to the "*", listen to all namespaces
# watched_namespace: development
cluster_labels: application:spilo
cluster_name_label: version
pod_role_label: spilo-role
db_hosted_zone: db.example.com
debug_logging: "true"
master_dns_name_format: '{cluster}.{team}.staging.{hostedzone}'
replica_dns_name_format: '{cluster}-repl.{team}.staging.{hostedzone}'
workers: "4"
docker_image: registry.opensource.zalan.do/acid/demospilo-10:1.3-p3
secret_name_template: '{username}.{cluster}.credentials'
etcd_host: ""
infrastructure_roles_secret_name: postgresql-infrastructure-roles
oauth_token_secret_name: postgresql-operator
pam_configuration: |
https://info.example.com/oauth2/tokeninfo?access_token= uid realm=/employees
pam_role_name: zalandos
# etcd_host: ""
super_username: postgres
enable_teams_api: "false"
# enable_team_superuser: "false"
# team_admin_role: "admin"
# teams_api_url: http://fake-teams-api.default.svc.cluster.local
# team_api_role_configuration: "log_statement:all"
# infrastructure_roles_secret_name: postgresql-infrastructure-roles
# oauth_token_secret_name: postgresql-operator
# pam_role_name: zalandos
# pam_configuration: |
# https://info.example.com/oauth2/tokeninfo?access_token= uid realm=/employees
db_hosted_zone: db.example.com
master_dns_name_format: '{cluster}.{team}.staging.{hostedzone}'
replica_dns_name_format: '{cluster}-repl.{team}.staging.{hostedzone}'
enable_master_load_balancer: "false"
enable_replica_load_balancer: "false"
pdb_name_format: "postgres-{cluster}-pdb"
api_port: "8080"
ring_log_lines: "100"
cluster_history_entries: "1000"
pod_terminate_grace_period: 5m
pod_deletion_wait_timeout: 10m
pod_label_wait_timeout: 10m
ready_wait_interval: 3s
@ -30,21 +45,3 @@ data:
resource_check_interval: 3s
resource_check_timeout: 10m
resync_period: 5m
super_username: postgres
enable_teams_api: "false"
enable_team_superuser: "false"
team_admin_role: "admin"
teams_api_url: http://fake-teams-api.default.svc.cluster.local
workers: "4"
# turn on/off load balancers for all Postgres clusters managed by the operator
# LB settings in cluster manifests take priority over these settings
enable_master_load_balancer: "true"
enable_replica_load_balancer: "false"
api_port: "8080"
ring_log_lines: "100"
cluster_history_entries: "1000"
pod_terminate_grace_period: 5m
pdb_name_format: "postgres-{cluster}-pdb"
node_eol_label: "lifecycle-status:pending-decommission"
node_readiness_label: ""
team_api_role_configuration: "log_statement:all"

View File

@ -12,15 +12,9 @@ spec:
serviceAccountName: operator
containers:
- name: postgres-operator
image: registry.opensource.zalan.do/acid/postgres-operator:0f392c2
image: registry.opensource.zalan.do/acid/postgres-operator:4c8dfd7
imagePullPolicy: IfNotPresent
env:
# uncomment to overwrite a similar setting from operator configmap
# if set to the empty string, watch the operator's own namespace
# if set to the "*", listen to all namespaces
# - name: WATCHED_NAMESPACE
# valueFrom:
# fieldRef:
# fieldPath: metadata.namespace
# provided additional ENV vars can overwrite individual config map entries
- name: CONFIG_MAP_NAME
value: "postgres-operator"

View File

@ -170,6 +170,10 @@ func (c *Cluster) setStatus(status spec.PostgresStatus) {
}
}
func (c *Cluster) isNewCluster() bool {
return c.Status == spec.ClusterStatusCreating
}
// initUsers populates c.systemUsers and c.pgUsers maps.
func (c *Cluster) initUsers() error {
c.setProcessName("initializing users")
@ -255,11 +259,15 @@ func (c *Cluster) Create() error {
if c.Endpoints[role] != nil {
return fmt.Errorf("%s endpoint already exists in the cluster", role)
}
ep, err = c.createEndpoint(role)
if err != nil {
return fmt.Errorf("could not create %s endpoint: %v", role, err)
if role == Master {
// replica endpoint will be created by the replica service. Master endpoint needs to be created by us,
// since the corresponding master service doesn't define any selectors.
ep, err = c.createEndpoint(role)
if err != nil {
return fmt.Errorf("could not create %s endpoint: %v", role, err)
}
c.logger.Infof("endpoint %q has been successfully created", util.NameFromMeta(ep.ObjectMeta))
}
c.logger.Infof("endpoint %q has been successfully created", util.NameFromMeta(ep.ObjectMeta))
if c.Services[role] != nil {
return fmt.Errorf("service already exists in the cluster")

View File

@ -684,12 +684,6 @@ func (c *Cluster) shouldCreateLoadBalancerForService(role PostgresRole, spec *sp
case Replica:
// deprecated option takes priority for backward compatibility
if spec.ReplicaLoadBalancer != nil {
c.logger.Debugf("The Postgres manifest for the cluster %v sets the deprecated `replicaLoadBalancer` param. Consider using the `enableReplicaLoadBalancer` instead.", c.Name)
return *spec.ReplicaLoadBalancer
}
// if the value is explicitly set in a Postgresql manifest, follow this setting
if spec.EnableReplicaLoadBalancer != nil {
return *spec.EnableReplicaLoadBalancer
@ -700,23 +694,10 @@ func (c *Cluster) shouldCreateLoadBalancerForService(role PostgresRole, spec *sp
case Master:
if spec.UseLoadBalancer != nil {
c.logger.Debugf("The Postgres manifest for the cluster %v sets the deprecated `useLoadBalancer` param. Consider using the `enableMasterLoadBalancer` instead.", c.Name)
return *spec.UseLoadBalancer
}
// if the value is explicitly set in a Postgresql manifest, follow this setting
if spec.EnableMasterLoadBalancer != nil {
return *spec.EnableMasterLoadBalancer
}
// `enable_load_balancer`` governs LB for a master service
// there is no equivalent deprecated operator option for the replica LB
if c.OpConfig.EnableLoadBalancer != nil {
c.logger.Debugf("The operator configmap sets the deprecated `enable_load_balancer` param. Consider using the `enable_master_load_balancer` or `enable_replica_load_balancer` instead.")
return *c.OpConfig.EnableLoadBalancer
}
return c.OpConfig.EnableMasterLoadBalancer
default:

View File

@ -65,23 +65,6 @@ func TestCreateLoadBalancerLogic(t *testing.T) {
opConfig: config.Config{EnableReplicaLoadBalancer: false},
result: false,
},
{
subtest: "old format, load balancer is enabled for replica",
role: Replica,
spec: &spec.PostgresSpec{ReplicaLoadBalancer: True()},
opConfig: config.Config{},
result: true,
},
{
subtest: "old format has priority",
role: Replica,
spec: &spec.PostgresSpec{
ReplicaLoadBalancer: True(),
EnableReplicaLoadBalancer: False(),
},
opConfig: config.Config{},
result: true,
},
}
for _, tt := range tests {
cluster.OpConfig = tt.opConfig

View File

@ -449,10 +449,16 @@ func (c *Cluster) deleteService(role PostgresRole) error {
}
func (c *Cluster) createEndpoint(role PostgresRole) (*v1.Endpoints, error) {
var (
subsets []v1.EndpointSubset
)
c.setProcessName("creating endpoint")
subsets := make([]v1.EndpointSubset, 0)
if role == Master {
//TODO: set subsets to the master
if !c.isNewCluster() {
subsets = c.generateEndpointSubsets(role)
} else {
// Patroni will populate the master endpoint for the new cluster
// The replica endpoint will be filled-in by the service selector.
subsets = make([]v1.EndpointSubset, 0)
}
endpointsSpec := c.generateEndpoint(role, subsets)
@ -466,6 +472,34 @@ func (c *Cluster) createEndpoint(role PostgresRole) (*v1.Endpoints, error) {
return endpoints, nil
}
func (c *Cluster) generateEndpointSubsets(role PostgresRole) []v1.EndpointSubset {
result := make([]v1.EndpointSubset, 0)
pods, err := c.getRolePods(role)
if err != nil {
if role == Master {
c.logger.Warningf("could not obtain the address for %s pod: %v", role, err)
} else {
c.logger.Warningf("could not obtain the addresses for %s pods: %v", role, err)
}
return result
}
endPointAddresses := make([]v1.EndpointAddress, 0)
for _, pod := range pods {
endPointAddresses = append(endPointAddresses, v1.EndpointAddress{IP: pod.Status.PodIP})
}
if len(endPointAddresses) > 0 {
result = append(result, v1.EndpointSubset{
Addresses: endPointAddresses,
Ports: []v1.EndpointPort{{"postgresql", 5432, "TCP"}},
})
} else if role == Master {
c.logger.Warningf("master is not running, generated master endpoint does not contain any addresses")
}
return result
}
func (c *Cluster) createPodDisruptionBudget() (*policybeta1.PodDisruptionBudget, error) {
podDisruptionBudgetSpec := c.generatePodDisruptionBudget()
podDisruptionBudget, err := c.KubeClient.

View File

@ -111,6 +111,7 @@ func (c *Controller) initOperatorConfig() {
}
c.opConfig = config.NewFromMap(configMapData)
c.warnOnDeprecatedOperatorParameters()
scalyrAPIKey := os.Getenv("SCALYR_API_KEY")
if scalyrAPIKey != "" {
@ -119,6 +120,14 @@ func (c *Controller) initOperatorConfig() {
}
// warningOnDeprecatedParameters emits warnings upon finding deprecated parmaters
func (c *Controller) warnOnDeprecatedOperatorParameters() {
if c.opConfig.EnableLoadBalancer != nil {
c.logger.Warningf("Operator configuration parameter 'enable_load_balancer' is deprecated and takes no effect. " +
"Consider using the 'enable_master_load_balancer' or 'enable_replica_load_balancer' instead.")
}
}
func (c *Controller) initPodServiceAccount() {
if c.opConfig.PodServiceAccountDefinition == "" {

View File

@ -166,6 +166,17 @@ func (c *Controller) processEvent(event spec.ClusterEvent) {
defer c.curWorkerCluster.Store(event.WorkerID, nil)
if event.EventType == spec.EventAdd || event.EventType == spec.EventUpdate || event.EventType == spec.EventSync {
// handle deprecated parameters by possibly assigning their values to the new ones.
if event.OldSpec != nil {
c.mergeDeprecatedPostgreSQLSpecParameters(&event.OldSpec.Spec)
}
if event.NewSpec != nil {
c.mergeDeprecatedPostgreSQLSpecParameters(&event.NewSpec.Spec)
}
c.warnOnDeprecatedPostgreSQLSpecParameters(&event.NewSpec.Spec)
}
switch event.EventType {
case spec.EventAdd:
if clusterFound {
@ -287,6 +298,46 @@ func (c *Controller) processClusterEventsQueue(idx int, stopCh <-chan struct{},
}
}
func (c *Controller) warnOnDeprecatedPostgreSQLSpecParameters(spec *spec.PostgresSpec) {
deprecate := func(deprecated, replacement string) {
c.logger.Warningf("Parameter %q is deprecated. Consider setting %q instead", deprecated, replacement)
}
noeffect := func(param string, explanation string) {
c.logger.Warningf("Parameter %q takes no effect. %s", param, explanation)
}
if spec.UseLoadBalancer != nil {
deprecate("useLoadBalancer", "enableMasterLoadBalancer")
}
if spec.ReplicaLoadBalancer != nil {
deprecate("replicaLoadBalancer", "enableReplicaLoadBalancer")
}
if len(spec.MaintenanceWindows) > 0 {
noeffect("maintenanceWindows", "Not implemented.")
}
}
func (c *Controller) mergeDeprecatedPostgreSQLSpecParameters(spec *spec.PostgresSpec) *spec.PostgresSpec {
if spec.UseLoadBalancer != nil || spec.ReplicaLoadBalancer != nil {
if spec.EnableReplicaLoadBalancer != nil || spec.EnableMasterLoadBalancer != nil {
c.logger.Warnf("Both old and new load balancer options are present, ignoring old ones")
} else {
if spec.UseLoadBalancer != nil {
spec.EnableMasterLoadBalancer = new(bool)
*spec.EnableMasterLoadBalancer = *spec.UseLoadBalancer
}
if spec.ReplicaLoadBalancer != nil {
spec.EnableReplicaLoadBalancer = new(bool)
*spec.EnableReplicaLoadBalancer = *spec.ReplicaLoadBalancer
}
}
}
return spec
}
func (c *Controller) queueClusterEvent(old, new *spec.Postgresql, eventType spec.EventType) {
var (
uid types.UID

View File

@ -0,0 +1,43 @@
package controller
import (
"github.com/zalando-incubator/postgres-operator/pkg/spec"
"reflect"
"testing"
)
var (
True bool = true
False bool = false
)
func TestMergeDeprecatedPostgreSQLSpecParameters(t *testing.T) {
c := NewController(&spec.ControllerConfig{})
tests := []struct {
name string
in *spec.PostgresSpec
out *spec.PostgresSpec
error string
}{
{
"Check that old parameters propagate values to the new ones",
&spec.PostgresSpec{UseLoadBalancer: &True, ReplicaLoadBalancer: &True},
&spec.PostgresSpec{UseLoadBalancer: &True, ReplicaLoadBalancer: &True,
EnableMasterLoadBalancer: &True, EnableReplicaLoadBalancer: &True},
"New parameters should be set from the values of old ones",
},
{
"Check that new parameters are not set when both old and new ones are present",
&spec.PostgresSpec{UseLoadBalancer: &True, EnableReplicaLoadBalancer: &True},
&spec.PostgresSpec{UseLoadBalancer: &True, EnableReplicaLoadBalancer: &True},
"New parameters should remain unchanged when both old and new are present",
},
}
for _, tt := range tests {
result := c.mergeDeprecatedPostgreSQLSpecParameters(tt.in)
if !reflect.DeepEqual(result, tt.out) {
t.Errorf("%s: %v", tt.name, tt.error)
}
}
}

View File

@ -88,7 +88,7 @@ type Config struct {
EnableMasterLoadBalancer bool `name:"enable_master_load_balancer" default:"true"`
EnableReplicaLoadBalancer bool `name:"enable_replica_load_balancer" default:"false"`
// deprecated and kept for backward compatibility
EnableLoadBalancer *bool `name:"enable_load_balancer" default:"true"`
EnableLoadBalancer *bool `name:"enable_load_balancer"`
MasterDNSNameFormat stringTemplate `name:"master_dns_name_format" default:"{cluster}.{team}.{hostedzone}"`
ReplicaDNSNameFormat stringTemplate `name:"replica_dns_name_format" default:"{cluster}-repl.{team}.{hostedzone}"`
PDBNameFormat stringTemplate `name:"pdb_name_format" default:"postgres-{cluster}-pdb"`