etcd.go

/*
Copyright 2014 Google Inc. All rights reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/

package etcd

import (
	"fmt"
	"strconv"

	"github.com/GoogleCloudPlatform/kubernetes/pkg/api"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/api/errors"
	etcderr "github.com/GoogleCloudPlatform/kubernetes/pkg/api/errors/etcd"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/api/validation"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/constraint"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/labels"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/registry/pod"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/runtime"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/tools"
	"github.com/GoogleCloudPlatform/kubernetes/pkg/watch"

	"github.com/golang/glog"
)

// TODO: Need to add a reconciler loop that makes sure that things in pods are reflected into
//       kubelet (and vice versa)

// Registry implements PodRegistry, ControllerRegistry, ServiceRegistry and MinionRegistry, backed by etcd.
type Registry struct {
	tools.EtcdHelper
	manifestFactory pod.ManifestFactory
}

// NewRegistry creates an etcd registry.
func NewRegistry(helper tools.EtcdHelper, manifestFactory pod.ManifestFactory) *Registry {
	registry := &Registry{
		EtcdHelper: helper,
	}
	registry.manifestFactory = manifestFactory
	return registry
}

func makePodKey(podID string) string {
	return "/registry/pods/" + podID
}

// parseWatchResourceVersion takes a resource version argument and converts it to
// the etcd version we should pass to helper.Watch(). Because resourceVersion is
// an opaque value, the default watch behavior for non-zero watch is to watch
// the next value (if you pass "1", you will see updates from "2" onwards).
func parseWatchResourceVersion(resourceVersion, kind string) (uint64, error) {
	if resourceVersion == "" || resourceVersion == "0" {
		return 0, nil
	}
	version, err := strconv.ParseUint(resourceVersion, 10, 64)
	if err != nil {
		return 0, etcderr.InterpretResourceVersionError(err, kind, resourceVersion)
	}
	return version + 1, nil
}

// ListPods obtains a list of pods with labels that match selector.
func (r *Registry) ListPods(ctx api.Context, selector labels.Selector) (*api.PodList, error) {
	return r.ListPodsPredicate(ctx, func(pod *api.Pod) bool {
		return selector.Matches(labels.Set(pod.Labels))
	})
}

// ListPodsPredicate obtains a list of pods that match filter.
func (r *Registry) ListPodsPredicate(ctx api.Context, filter func(*api.Pod) bool) (*api.PodList, error) {
	allPods := api.PodList{}
	err := r.ExtractToList("/registry/pods", &allPods)
	if err != nil {
		return nil, err
	}
	filtered := []api.Pod{}
	for _, pod := range allPods.Items {
		if filter(&pod) {
			// TODO: Currently nothing sets CurrentState.Host. We need a feedback loop that sets
			// the CurrentState.Host and Status fields. Here we pretend that reality perfectly
			// matches our desires.
			pod.CurrentState.Host = pod.DesiredState.Host
			filtered = append(filtered, pod)
		}
	}
	allPods.Items = filtered
	return &allPods, nil
}

// WatchPods begins watching for new, changed, or deleted pods.
func (r *Registry) WatchPods(ctx api.Context, resourceVersion string, filter func(*api.Pod) bool) (watch.Interface, error) {
	version, err := parseWatchResourceVersion(resourceVersion, "pod")
	if err != nil {
		return nil, err
	}
	return r.WatchList("/registry/pods", version, func(obj runtime.Object) bool {
		switch t := obj.(type) {
		case *api.Pod:
			return filter(t)
		default:
			// Must be an error
			return true
		}
	})
}

// GetPod gets a specific pod specified by its ID.
func (r *Registry) GetPod(ctx api.Context, podID string) (*api.Pod, error) {
	var pod api.Pod
	if err := r.ExtractObj(makePodKey(podID), &pod, false); err != nil {
		return nil, etcderr.InterpretGetError(err, "pod", podID)
	}
	// TODO: Currently nothing sets CurrentState.Host. We need a feedback loop that sets
	// the CurrentState.Host and Status fields. Here we pretend that reality perfectly
	// matches our desires.
	pod.CurrentState.Host = pod.DesiredState.Host
	return &pod, nil
}

func makeContainerKey(machine string) string {
	return "/registry/hosts/" + machine + "/kubelet"
}

// CreatePod creates a pod based on a specification.
func (r *Registry) CreatePod(ctx api.Context, pod *api.Pod) error {
	// Set current status to "Waiting".
	pod.CurrentState.Status = api.PodWaiting
	pod.CurrentState.Host = ""
	// DesiredState.Host == "" is a signal to the scheduler that this pod needs scheduling.
	pod.DesiredState.Status = api.PodRunning
	pod.DesiredState.Host = ""
	err := r.CreateObj(makePodKey(pod.ID), pod, 0)
	return etcderr.InterpretCreateError(err, "pod", pod.ID)
}

// ApplyBinding implements binding's registry
func (r *Registry) ApplyBinding(ctx api.Context, binding *api.Binding) error {
	return etcderr.InterpretCreateError(r.assignPod(binding.PodID, binding.Host), "binding", "")
}

// setPodHostTo sets the given pod's host to 'machine' iff it was previously 'oldMachine'.
// Returns the current state of the pod, or an error.
func (r *Registry) setPodHostTo(podID, oldMachine, machine string) (finalPod *api.Pod, err error) {
	podKey := makePodKey(podID)
	err = r.AtomicUpdate(podKey, &api.Pod{}, func(obj runtime.Object) (runtime.Object, error) {
		pod, ok := obj.(*api.Pod)
		if !ok {
			return nil, fmt.Errorf("unexpected object: %#v", obj)
		}
		if pod.DesiredState.Host != oldMachine {
			return nil, fmt.Errorf("pod %v is already assigned to host %v", pod.ID, pod.DesiredState.Host)
		}
		pod.DesiredState.Host = machine
		finalPod = pod
		return pod, nil
	})
	return finalPod, err
}

// assignPod assigns the given pod to the given machine.
func (r *Registry) assignPod(podID string, machine string) error {
	finalPod, err := r.setPodHostTo(podID, "", machine)
	if err != nil {
		return err
	}
	// TODO: move this to a watch/rectification loop.
	manifest, err := r.manifestFactory.MakeManifest(machine, *finalPod)
	if err != nil {
		return err
	}
	contKey := makeContainerKey(machine)
	err = r.AtomicUpdate(contKey, &api.ContainerManifestList{}, func(in runtime.Object) (runtime.Object, error) {
		manifests := *in.(*api.ContainerManifestList)
		manifests.Items = append(manifests.Items, manifest)
		if !constraint.Allowed(manifests.Items) {
			return nil, fmt.Errorf("The assignment would cause a constraint violation")
		}
		return &manifests, nil
	})
	if err != nil {
		// Put the pod's host back the way it was. This is a terrible hack that
		// won't be needed if we convert this to a rectification loop.
		if _, err2 := r.setPodHostTo(podID, machine, ""); err2 != nil {
			glog.Errorf("Stranding pod %v; couldn't clear host after previous error: %v", podID, err2)
		}
	}
	return err
}

func (r *Registry) UpdatePod(ctx api.Context, pod *api.Pod) error {
	var podOut api.Pod
	podKey := makePodKey(pod.ID)
	err := r.EtcdHelper.ExtractObj(podKey, &podOut, false)
	if err != nil {
		return err
	}
	scheduled := podOut.DesiredState.Host != ""
	if scheduled {
		pod.DesiredState.Host = podOut.DesiredState.Host
		// If it's already been scheduled, limit the types of updates we'll accept.
		errs := validation.ValidatePodUpdate(pod, &podOut)
		if len(errs) != 0 {
			return errors.NewInvalid("Pod", pod.ID, errs)
		}
	}
	// There's no race with the scheduler, because either this write will fail because the host
	// has been updated, or the host update will fail because this pod has been updated.
	err = r.EtcdHelper.SetObj(podKey, pod)
	if err != nil {
		return err
	}
	if !scheduled {
		// never scheduled, just update.
		return nil
	}
	containerKey := makeContainerKey(podOut.DesiredState.Host)
	return r.AtomicUpdate(containerKey, &api.ContainerManifestList{}, func(in runtime.Object) (runtime.Object, error) {
		manifests := in.(*api.ContainerManifestList)
		for ix := range manifests.Items {
			if manifests.Items[ix].ID == pod.ID {
				manifests.Items[ix] = pod.DesiredState.Manifest
				return manifests, nil
			}
		}
		// This really shouldn't happen
		glog.Warningf("Couldn't find: %s in %#v", pod.ID, manifests)
		return manifests, fmt.Errorf("Failed to update pod, couldn't find %s in %#v", pod.ID, manifests)
	})
}

// DeletePod deletes an existing pod specified by its ID.
func (r *Registry) DeletePod(ctx api.Context, podID string) error {
	var pod api.Pod
	podKey := makePodKey(podID)
	err := r.ExtractObj(podKey, &pod, false)
	if err != nil {
		return etcderr.InterpretDeleteError(err, "pod", podID)
	}
	// First delete the pod, so a scheduler doesn't notice it getting removed from the
	// machine and attempt to put it somewhere.
	err = r.Delete(podKey, true)
	if err != nil {
		return etcderr.InterpretDeleteError(err, "pod", podID)
	}
	machine := pod.DesiredState.Host
	if machine == "" {
		// Pod was never scheduled anywhere, just return.
		return nil
	}
	// Next, remove the pod from the machine atomically.
	contKey := makeContainerKey(machine)
	return r.AtomicUpdate(contKey, &api.ContainerManifestList{}, func(in runtime.Object) (runtime.Object, error) {
		manifests := in.(*api.ContainerManifestList)
		newManifests := make([]api.ContainerManifest, 0, len(manifests.Items))
		found := false
		for _, manifest := range manifests.Items {
			if manifest.ID != podID {
				newManifests = append(newManifests, manifest)
			} else {
				found = true
			}
		}
		if !found {
			// This really shouldn't happen, it indicates something is broken, and likely
			// there is a lost pod somewhere.
			// However it is "deleted" so log it and move on
			glog.Warningf("Couldn't find: %s in %#v", podID, manifests)
		}
		manifests.Items = newManifests
		return manifests, nil
	})
}

// ListControllers obtains a list of ReplicationControllers.
func (r *Registry) ListControllers(ctx api.Context) (*api.ReplicationControllerList, error) {
	controllers := &api.ReplicationControllerList{}
	err := r.ExtractToList("/registry/controllers", controllers)
	return controllers, err
}

// WatchControllers begins watching for new, changed, or deleted controllers.
func (r *Registry) WatchControllers(ctx api.Context, resourceVersion string) (watch.Interface, error) {
	version, err := parseWatchResourceVersion(resourceVersion, "replicationControllers")
	if err != nil {
		return nil, err
	}
	return r.WatchList("/registry/controllers", version, tools.Everything)
}

func makeControllerKey(id string) string {
	return "/registry/controllers/" + id
}

// GetController gets a specific ReplicationController specified by its ID.
func (r *Registry) GetController(ctx api.Context, controllerID string) (*api.ReplicationController, error) {
	var controller api.ReplicationController
	key := makeControllerKey(controllerID)
	err := r.ExtractObj(key, &controller, false)
	if err != nil {
		return nil, etcderr.InterpretGetError(err, "replicationController", controllerID)
	}
	return &controller, nil
}

// CreateController creates a new ReplicationController.
func (r *Registry) CreateController(ctx api.Context, controller *api.ReplicationController) error {
	err := r.CreateObj(makeControllerKey(controller.ID), controller, 0)
	return etcderr.InterpretCreateError(err, "replicationController", controller.ID)
}

// UpdateController replaces an existing ReplicationController.
func (r *Registry) UpdateController(ctx api.Context, controller *api.ReplicationController) error {
	err := r.SetObj(makeControllerKey(controller.ID), controller)
	return etcderr.InterpretUpdateError(err, "replicationController", controller.ID)
}

// DeleteController deletes a ReplicationController specified by its ID.
func (r *Registry) DeleteController(ctx api.Context, controllerID string) error {
	key := makeControllerKey(controllerID)
	err := r.Delete(key, false)
	return etcderr.InterpretDeleteError(err, "replicationController", controllerID)
}

func makeServiceKey(name string) string {
	return "/registry/services/specs/" + name
}

// ListServices obtains a list of Services.
func (r *Registry) ListServices(ctx api.Context) (*api.ServiceList, error) {
	list := &api.ServiceList{}
	err := r.ExtractToList("/registry/services/specs", list)
	return list, err
}

// CreateService creates a new Service.
func (r *Registry) CreateService(ctx api.Context, svc *api.Service) error {
	err := r.CreateObj(makeServiceKey(svc.ID), svc, 0)
	return etcderr.InterpretCreateError(err, "service", svc.ID)
}

// GetService obtains a Service specified by its name.
func (r *Registry) GetService(ctx api.Context, name string) (*api.Service, error) {
	key := makeServiceKey(name)
	var svc api.Service
	err := r.ExtractObj(key, &svc, false)
	if err != nil {
		return nil, etcderr.InterpretGetError(err, "service", name)
	}
	return &svc, nil
}

// GetEndpoints obtains the endpoints for the service identified by 'name'.
func (r *Registry) GetEndpoints(ctx api.Context, name string) (*api.Endpoints, error) {
	key := makeServiceEndpointsKey(name)
	var endpoints api.Endpoints
	err := r.ExtractObj(key, &endpoints, false)
	if err != nil {
		return nil, etcderr.InterpretGetError(err, "endpoints", name)
	}
	return &endpoints, nil
}

func makeServiceEndpointsKey(name string) string {
	return "/registry/services/endpoints/" + name
}

// DeleteService deletes a Service specified by its name.
func (r *Registry) DeleteService(ctx api.Context, name string) error {
	key := makeServiceKey(name)
	err := r.Delete(key, true)
	if err != nil {
		return etcderr.InterpretDeleteError(err, "service", name)
	}

	// TODO: can leave dangling endpoints, and potentially return incorrect
	// endpoints if a new service is created with the same name
	key = makeServiceEndpointsKey(name)
	if err := r.Delete(key, true); err != nil && !tools.IsEtcdNotFound(err) {
		return etcderr.InterpretDeleteError(err, "endpoints", name)
	}
	return nil
}

// UpdateService replaces an existing Service.
func (r *Registry) UpdateService(ctx api.Context, svc *api.Service) error {
	err := r.SetObj(makeServiceKey(svc.ID), svc)
	return etcderr.InterpretUpdateError(err, "service", svc.ID)
}

// WatchServices begins watching for new, changed, or deleted service configurations.
func (r *Registry) WatchServices(ctx api.Context, label, field labels.Selector, resourceVersion string) (watch.Interface, error) {
	version, err := parseWatchResourceVersion(resourceVersion, "service")
	if err != nil {
		return nil, err
	}
	if !label.Empty() {
		return nil, fmt.Errorf("label selectors are not supported on services")
	}
	if value, found := field.RequiresExactMatch("ID"); found {
		return r.Watch(makeServiceKey(value), version), nil
	}
	if field.Empty() {
		return r.WatchList("/registry/services/specs", version, tools.Everything)
	}
	return nil, fmt.Errorf("only the 'ID' and default (everything) field selectors are supported")
}

// ListEndpoints obtains a list of Services.
func (r *Registry) ListEndpoints(ctx api.Context) (*api.EndpointsList, error) {
	list := &api.EndpointsList{}
	err := r.ExtractToList("/registry/services/endpoints", list)
	return list, err
}

// UpdateEndpoints update Endpoints of a Service.
func (r *Registry) UpdateEndpoints(ctx api.Context, e *api.Endpoints) error {
	// TODO: this is a really bad misuse of AtomicUpdate, need to compute a diff inside the loop.
	err := r.AtomicUpdate(makeServiceEndpointsKey(e.ID), &api.Endpoints{},
		func(input runtime.Object) (runtime.Object, error) {
			// TODO: racy - label query is returning different results for two simultaneous updaters
			return e, nil
		})
	return etcderr.InterpretUpdateError(err, "endpoints", e.ID)
}

// WatchEndpoints begins watching for new, changed, or deleted endpoint configurations.
func (r *Registry) WatchEndpoints(ctx api.Context, label, field labels.Selector, resourceVersion string) (watch.Interface, error) {
	version, err := parseWatchResourceVersion(resourceVersion, "endpoints")
	if err != nil {
		return nil, err
	}
	if !label.Empty() {
		return nil, fmt.Errorf("label selectors are not supported on endpoints")
	}
	if value, found := field.RequiresExactMatch("ID"); found {
		return r.Watch(makeServiceEndpointsKey(value), version), nil
	}
	if field.Empty() {
		return r.WatchList("/registry/services/endpoints", version, tools.Everything)
	}
	return nil, fmt.Errorf("only the 'ID' and default (everything) field selectors are supported")
}

func makeMinionKey(minionID string) string {
	return "/registry/minions/" + minionID
}

func (r *Registry) ListMinions(ctx api.Context) (*api.MinionList, error) {
	minions := &api.MinionList{}
	err := r.ExtractToList("/registry/minions", minions)
	return minions, err
}

func (r *Registry) CreateMinion(ctx api.Context, minion *api.Minion) error {
	// TODO: Add some validations.
	err := r.CreateObj(makeMinionKey(minion.ID), minion, 0)
	return etcderr.InterpretCreateError(err, "minion", minion.ID)
}

func (r *Registry) GetMinion(ctx api.Context, minionID string) (*api.Minion, error) {
	var minion api.Minion
	key := makeMinionKey(minionID)
	err := r.ExtractObj(key, &minion, false)
	if err != nil {
		return nil, etcderr.InterpretGetError(err, "minion", minion.ID)
	}
	return &minion, nil
}

func (r *Registry) DeleteMinion(ctx api.Context, minionID string) error {
	key := makeMinionKey(minionID)
	err := r.Delete(key, true)
	if err != nil {
		return etcderr.InterpretDeleteError(err, "minion", minionID)
	}
	return nil
}