temporal/client/client_bean.go

//go:generate mockgen -package $GOPACKAGE -source $GOFILE -destination client_bean_mock.go

package client

import (
	"fmt"
	"sync"
	"sync/atomic"

	"go.temporal.io/api/serviceerror"
	"go.temporal.io/api/workflowservice/v1"
	"go.temporal.io/server/api/adminservice/v1"
	"go.temporal.io/server/api/historyservice/v1"
	"go.temporal.io/server/api/matchingservice/v1"
	"go.temporal.io/server/client/admin"
	"go.temporal.io/server/client/frontend"
	"go.temporal.io/server/client/history"
	"go.temporal.io/server/client/matching"
	"go.temporal.io/server/common/cluster"
	"google.golang.org/grpc"
)

type (
	// Bean is a collection of clients
	Bean interface {
		GetHistoryClient() historyservice.HistoryServiceClient
		GetMatchingClient(namespaceIDToName NamespaceIDToNameFunc) (matchingservice.MatchingServiceClient, error)
		GetFrontendClient() workflowservice.WorkflowServiceClient
		GetRemoteAdminClient(string) (adminservice.AdminServiceClient, error)
		GetRemoteFrontendClient(string) (grpc.ClientConnInterface, workflowservice.WorkflowServiceClient, error)
		// Close deterministically releases bean-held resources on shutdown: it
		// stops the history and matching clients (their daemon goroutines and
		// cached gRPC connections) and unregisters the cluster metadata change
		// callback.
		Close()
	}

	frontendClient struct {
		connection grpc.ClientConnInterface
		workflowservice.WorkflowServiceClient
	}

	clientBeanImpl struct {
		sync.Mutex
		historyClient   historyservice.HistoryServiceClient
		matchingClient  atomic.Value
		clusterMetadata cluster.Metadata
		factory         Factory

		adminClientsLock    sync.RWMutex
		adminClients        map[string]adminservice.AdminServiceClient
		frontendClientsLock sync.RWMutex
		frontendClients     map[string]frontendClient
	}
)

// NewClientBean provides a collection of clients
func NewClientBean(factory Factory, clusterMetadata cluster.Metadata) (Bean, error) {

	historyClient, err := factory.NewHistoryClientWithTimeout(history.DefaultTimeout)
	if err != nil {
		return nil, err
	}

	adminClients := map[string]adminservice.AdminServiceClient{}
	frontendClients := map[string]frontendClient{}

	currentClusterName := clusterMetadata.GetCurrentClusterName()
	// Init local cluster client with membership info
	adminClient, err := factory.NewLocalAdminClientWithTimeout(
		admin.DefaultTimeout,
		admin.DefaultLargeTimeout,
	)
	if err != nil {
		return nil, err
	}
	conn, client, err := factory.NewLocalFrontendClientWithTimeout(
		frontend.DefaultTimeout,
		frontend.DefaultLongPollTimeout,
	)
	if err != nil {
		return nil, err
	}
	adminClients[currentClusterName] = adminClient
	frontendClients[currentClusterName] = frontendClient{
		connection:            conn,
		WorkflowServiceClient: client,
	}

	bean := &clientBeanImpl{
		factory:         factory,
		historyClient:   historyClient,
		clusterMetadata: clusterMetadata,
		adminClients:    adminClients,
		frontendClients: frontendClients,
	}
	bean.registerClientEviction()
	return bean, nil
}

func (h *clientBeanImpl) registerClientEviction() {
	currentCluster := h.clusterMetadata.GetCurrentClusterName()
	h.clusterMetadata.RegisterMetadataChangeCallback(
		h,
		func(oldClusterMetadata map[string]*cluster.ClusterInformation, newClusterMetadata map[string]*cluster.ClusterInformation) {
			for clusterName := range newClusterMetadata {
				if clusterName == currentCluster {
					continue
				}
				h.adminClientsLock.Lock()
				delete(h.adminClients, clusterName)
				h.adminClientsLock.Unlock()
				h.frontendClientsLock.Lock()
				delete(h.frontendClients, clusterName)
				h.frontendClientsLock.Unlock()
			}
		})
}

// Close releases the resources held by the bean's clients. See the Bean
// interface for details. It is safe to call more than once.
func (h *clientBeanImpl) Close() {
	h.clusterMetadata.UnRegisterMetadataChangeCallback(h)

	// The history and matching client wrapper chains implement Stop();
	// stopping them releases their daemon goroutines and cached gRPC
	// connections.
	if s, ok := h.historyClient.(interface{ Stop() }); ok {
		s.Stop()
	}
	if mc := h.matchingClient.Load(); mc != nil {
		if s, ok := mc.(interface{ Stop() }); ok {
			s.Stop()
		}
	}
}

func (h *clientBeanImpl) GetHistoryClient() historyservice.HistoryServiceClient {
	return h.historyClient
}

func (h *clientBeanImpl) GetMatchingClient(namespaceIDToName NamespaceIDToNameFunc) (matchingservice.MatchingServiceClient, error) {
	if client := h.matchingClient.Load(); client != nil {
		return client.(matchingservice.MatchingServiceClient), nil
	}
	return h.lazyInitMatchingClient(namespaceIDToName)
}

func (h *clientBeanImpl) GetFrontendClient() workflowservice.WorkflowServiceClient {
	return h.frontendClients[h.clusterMetadata.GetCurrentClusterName()]
}

func (h *clientBeanImpl) GetRemoteAdminClient(cluster string) (adminservice.AdminServiceClient, error) {
	h.adminClientsLock.RLock()
	client, ok := h.adminClients[cluster]
	h.adminClientsLock.RUnlock()
	if ok {
		return client, nil
	}

	clusterInfo, clusterFound := h.clusterMetadata.GetAllClusterInfo()[cluster]
	if !clusterFound {
		// We intentionally return internal error here.
		// This error could only happen with internal mis-configuration.
		// This can happen when a namespace is config for multiple clusters. But those clusters are not connected.
		// We also have logic in task processing to drop tasks when namespace cluster exclude a local cluster.
		return nil, &serviceerror.Internal{
			Message: fmt.Sprintf(
				"Unknown cluster name: %v with given cluster information map: %v.",
				cluster,
				clusterInfo,
			),
		}
	}

	h.adminClientsLock.Lock()
	defer h.adminClientsLock.Unlock()
	client, ok = h.adminClients[cluster]
	if ok {
		return client, nil
	}

	client = h.factory.NewRemoteAdminClientWithTimeout(
		clusterInfo.RPCAddress,
		admin.DefaultTimeout,
		admin.DefaultLargeTimeout,
	)
	h.adminClients[cluster] = client
	return client, nil
}

func (h *clientBeanImpl) GetRemoteFrontendClient(clusterName string) (grpc.ClientConnInterface, workflowservice.WorkflowServiceClient, error) {
	h.frontendClientsLock.RLock()
	client, ok := h.frontendClients[clusterName]
	h.frontendClientsLock.RUnlock()
	if ok {
		return client.connection, client, nil
	}

	clusterInfo, clusterFound := h.clusterMetadata.GetAllClusterInfo()[clusterName]
	if !clusterFound {
		// We intentionally return internal error here.
		// This error could only happen with internal mis-configuration.
		// This can happen when a namespace is config for multiple clusters. But those clusters are not connected.
		// We also have logic in task processing to drop tasks when namespace cluster exclude a local cluster.
		return nil, nil, &serviceerror.Internal{
			Message: fmt.Sprintf(
				"Unknown clusterName name: %v with given clusterName information map: %v.",
				clusterName,
				clusterInfo,
			),
		}
	}

	h.frontendClientsLock.Lock()
	defer h.frontendClientsLock.Unlock()

	client, ok = h.frontendClients[clusterName]
	if ok {
		return client.connection, client, nil
	}

	conn, fClient := h.factory.NewRemoteFrontendClientWithTimeout(
		clusterInfo.RPCAddress,
		frontend.DefaultTimeout,
		frontend.DefaultLongPollTimeout,
	)
	client = frontendClient{
		connection:            conn,
		WorkflowServiceClient: fClient,
	}
	h.frontendClients[clusterName] = client
	return client.connection, client, nil
}

func (h *clientBeanImpl) lazyInitMatchingClient(namespaceIDToName NamespaceIDToNameFunc) (matchingservice.MatchingServiceClient, error) {
	h.Lock()
	defer h.Unlock()
	if cached := h.matchingClient.Load(); cached != nil {
		return cached.(matchingservice.MatchingServiceClient), nil
	}
	client, err := h.factory.NewMatchingClientWithTimeout(namespaceIDToName, matching.DefaultTimeout, matching.DefaultLongPollTimeout)
	if err != nil {
		return nil, err
	}
	h.matchingClient.Store(client)
	return client, nil
}