feat: audit and remove etcd members if machines no longer exist

This PR introduces some auditing functionality for etcd in the workload
clusters. If there are members in the workload etcd that don't
correspond to any CP node, the members will be removed. This helps in
the case where a machine is deleted out from under us and a new CP
machine comes up but can't join etcd.

Signed-off-by: Spencer Smith <[email protected]>

Author: rsmitty

controllers/etcd.go

@@ -6,6 +6,8 @@ package controllers
 
 import (
 	"context"
+	"fmt"
+	"strings"
 
 	"github.com/talos-systems/talos/pkg/machinery/api/machine"
 	talosclient "github.com/talos-systems/talos/pkg/machinery/client"
@@ -45,3 +47,108 @@ func (r *TalosControlPlaneReconciler) gracefulEtcdLeave(ctx context.Context, c *
 
 	return nil
 }
+
+// forceEtcdLeave removes a given machine from the etcd cluster by telling another CP node to remove the member.
+// This is used in times when the machine was deleted out from under us.
+func (r *TalosControlPlaneReconciler) forceEtcdLeave(ctx context.Context, c *talosclient.Client, cluster client.ObjectKey, memberName string) error {
+	r.Log.Info("Removing etcd member", "memberName", memberName)
+
+	err := c.EtcdRemoveMember(
+		ctx,
+		&machine.EtcdRemoveMemberRequest{
+			Member: memberName,
+		},
+	)
+	if err != nil {
+		return err
+	}
+
+	return nil
+}
+
+// auditEtcd rolls through all etcd members to see if there's a matching controlplane machine
+// It uses the first controlplane node returned as the etcd endpoint
+func (r *TalosControlPlaneReconciler) auditEtcd(ctx context.Context, cluster client.ObjectKey, cpName string) error {
+	machines, err := r.getControlPlaneMachinesForCluster(ctx, cluster, cpName)
+	if err != nil {
+		return err
+	}
+
+	if len(machines) == 0 {
+		return nil
+	}
+
+	for _, machine := range machines {
+		// nb: we'll assume any machine that doesn't have a noderef is new and we can audit later because
+		//     otherwise a new etcd member can get removed before even getting the noderef set by the CAPI controllers.
+		if machine.Status.NodeRef == nil {
+			return fmt.Errorf("some CP machines do not have a noderef")
+		}
+	}
+	// Select the first CP machine that's not being deleted and has a noderef
+	var designatedCPMachine capiv1.Machine
+
+	for _, machine := range machines {
+		if machine.ObjectMeta.DeletionTimestamp.IsZero() && machine.Status.NodeRef != nil {
+			designatedCPMachine = machine
+			break
+		}
+	}
+
+	clientset, err := r.kubeconfigForCluster(ctx, cluster)
+	if err != nil {
+		return err
+	}
+
+	c, err := r.talosconfigForMachine(ctx, clientset, designatedCPMachine)
+	if err != nil {
+		return err
+	}
+
+	// Save the first internal IP of the designated machine to use as our node target
+	// and setup the ctx to target it
+	var firstIntAddr string
+
+	for _, addr := range designatedCPMachine.Status.Addresses {
+		if addr.Type == capiv1.MachineInternalIP {
+			firstIntAddr = addr.Address
+			break
+		}
+	}
+
+	nodeCtx := talosclient.WithNodes(ctx, firstIntAddr)
+
+	response, err := c.EtcdMemberList(nodeCtx, &machine.EtcdMemberListRequest{})
+	if err != nil {
+		return err
+	}
+
+	// Only querying one CP node, so only 1 message should return.
+	memberList := response.Messages[0]
+
+	if len(memberList.Members) == 0 {
+		return nil
+	}
+
+	// For each etcd member, look through the list of machines and see if noderef matches
+	for _, member := range memberList.Members {
+		present := false
+		for _, machine := range machines {
+			// break apart the noderef name in case it's an fqdn (like in AWS)
+			machineNodeNameExploded := strings.Split(machine.Status.NodeRef.Name, ".")
+
+			if machineNodeNameExploded[0] == member {
+				present = true
+				break
+			}
+		}
+
+		if !present {
+			r.Log.Info("found etcd member that doesn't exist as controlplane machine", "member", member)
+
+			r.forceEtcdLeave(nodeCtx, c, cluster, member)
+		}
+	}
+
+	return nil
+}

controllers/taloscontrolplane_controller.go

@@ -245,6 +245,12 @@ func (r *TalosControlPlaneReconciler) Reconcile(req ctrl.Request) (res ctrl.Resu
 		return ctrl.Result{}, err
 	}
 
+	// Audit the etcd member list to remove any nodes that no longer exist
+	if err := r.auditEtcd(ctx, util.ObjectKey(cluster), controlPlane.TCP.Name); err != nil {
+		logger.Info("failed to check etcd membership list", "error", err)
+		return ctrl.Result{Requeue: true}, nil
+	}
+
 	return ctrl.Result{}, nil
 }
 

go.mod

@@ -5,11 +5,11 @@ go 1.16
 require (
 	cloud.google.com/go v0.47.0 // indirect
 	github.com/go-logr/logr v0.1.0
-	github.com/onsi/ginkgo v1.12.1
+	github.com/onsi/ginkgo v1.15.0
 	github.com/onsi/gomega v1.10.1
 	github.com/pkg/errors v0.9.1
-	github.com/talos-systems/cluster-api-bootstrap-provider-talos v0.2.0-alpha.10
-	github.com/talos-systems/talos/pkg/machinery v0.0.0-20210218160848-32d25885288f
+	github.com/talos-systems/cluster-api-bootstrap-provider-talos v0.2.0-alpha.12
+	github.com/talos-systems/talos/pkg/machinery v0.0.0-20210520203624-828772cec9a3
 	k8s.io/api v0.17.9
 	k8s.io/apimachinery v0.17.9
 	k8s.io/apiserver v0.17.9