Пример #1
0
// TestNodeLivenessEpochIncrement verifies that incrementing the epoch
// of a node requires the node to be considered not-live and that on
// increment, no other nodes believe the epoch-incremented node to be
// live.
func TestNodeLivenessEpochIncrement(t *testing.T) {
	defer leaktest.AfterTest(t)()
	mtc := startMultiTestContext(t, 2)
	defer mtc.Stop()

	verifyLiveness(t, mtc)
	stopNodeLivenessHeartbeats(mtc)

	// First try to increment the epoch of a known-live node.
	deadNodeID := mtc.gossips[1].NodeID.Get()
	if err := mtc.nodeLivenesses[0].IncrementEpoch(
		context.Background(), deadNodeID); !testutils.IsError(err, "cannot increment epoch on live node") {
		t.Fatalf("expected error incrementing a live node: %v", err)
	}

	// Advance clock past liveness threshold & increment epoch.
	oldLiveness, err := mtc.nodeLivenesses[0].GetLiveness(deadNodeID)
	if err != nil {
		t.Fatal(err)
	}
	active, _ := storage.RangeLeaseDurations(
		storage.RaftElectionTimeout(base.DefaultRaftTickInterval, 0))
	mtc.manualClock.Increment(active.Nanoseconds() + 1)
	if err := mtc.nodeLivenesses[0].IncrementEpoch(context.Background(), deadNodeID); err != nil {
		t.Fatalf("unexpected error incrementing a live node: %s", err)
	}

	// Verify that the epoch has been advanced.
	util.SucceedsSoon(t, func() error {
		newLiveness, err := mtc.nodeLivenesses[0].GetLiveness(deadNodeID)
		if err != nil {
			return err
		}
		if newLiveness.Epoch != oldLiveness.Epoch+1 {
			return errors.Errorf("expected epoch to increment")
		}
		if newLiveness.Expiration != oldLiveness.Expiration {
			return errors.Errorf("expected expiration to remain unchanged")
		}
		if live, err := mtc.nodeLivenesses[0].IsLive(deadNodeID); live || err != nil {
			return errors.Errorf("expected dead node to remain dead after epoch increment %t: %v", live, err)
		}
		return nil
	})

	// Verify epoch increment metric count.
	if c := mtc.nodeLivenesses[0].Metrics().EpochIncrements.Count(); c != 1 {
		t.Errorf("expected epoch increment == 1; got %d", c)
	}
}
Пример #2
0
func TestNodeLiveness(t *testing.T) {
	defer leaktest.AfterTest(t)()
	t.Skip("#9973")
	mtc := startMultiTestContext(t, 3)
	defer mtc.Stop()

	// Verify liveness of all nodes for all nodes.
	verifyLiveness(t, mtc)
	stopNodeLivenessHeartbeats(mtc)

	// Advance clock past the liveness threshold to verify IsLive becomes false.
	active, _ := storage.RangeLeaseDurations(
		storage.RaftElectionTimeout(base.DefaultRaftTickInterval, 0))
	mtc.manualClock.Increment(active.Nanoseconds() + 1)
	for idx, nl := range mtc.nodeLivenesses {
		nodeID := mtc.gossips[idx].NodeID.Get()
		live, err := nl.IsLive(nodeID)
		if err != nil {
			t.Error(err)
		} else if live {
			t.Errorf("expected node %d to be considered not-live after advancing node clock", nodeID)
		}
	}
	// Trigger a manual heartbeat and verify liveness is reestablished.
	for _, nl := range mtc.nodeLivenesses {
		if err := nl.ManualHeartbeat(); err != nil {
			t.Fatal(err)
		}
	}
	verifyLiveness(t, mtc)

	// Verify metrics counts.
	for _, nl := range mtc.nodeLivenesses {
		if c := nl.Metrics().HeartbeatSuccesses.Count(); c != 2 {
			t.Errorf("expected metrics count == 2; got %d", c)
		}
	}
}
Пример #3
0
// NewServer creates a Server from a server.Context.
func NewServer(cfg Config, stopper *stop.Stopper) (*Server, error) {
	if _, err := net.ResolveTCPAddr("tcp", cfg.AdvertiseAddr); err != nil {
		return nil, errors.Errorf("unable to resolve RPC address %q: %v", cfg.AdvertiseAddr, err)
	}

	if cfg.AmbientCtx.Tracer == nil {
		cfg.AmbientCtx.Tracer = tracing.NewTracer()
	}

	// Try loading the TLS configs before anything else.
	if _, err := cfg.GetServerTLSConfig(); err != nil {
		return nil, err
	}
	if _, err := cfg.GetClientTLSConfig(); err != nil {
		return nil, err
	}

	s := &Server{
		mux:     http.NewServeMux(),
		clock:   hlc.NewClock(hlc.UnixNano, cfg.MaxOffset),
		stopper: stopper,
		cfg:     cfg,
	}
	// Add a dynamic log tag value for the node ID.
	//
	// We need to pass an ambient context to the various server components, but we
	// won't know the node ID until we Start(). At that point it's too late to
	// change the ambient contexts in the components (various background processes
	// will have already started using them).
	//
	// NodeIDContainer allows us to add the log tag to the context now and update
	// the value asynchronously. It's not significantly more expensive than a
	// regular tag since it's just doing an (atomic) load when a log/trace message
	// is constructed. The node ID is set by the Store if this host was
	// bootstrapped; otherwise a new one is allocated in Node.
	s.cfg.AmbientCtx.AddLogTag("n", &s.nodeIDContainer)

	ctx := s.AnnotateCtx(context.Background())
	if s.cfg.Insecure {
		log.Warning(ctx, "running in insecure mode, this is strongly discouraged. See --insecure.")
	}

	s.rpcContext = rpc.NewContext(s.cfg.AmbientCtx, s.cfg.Config, s.clock, s.stopper)
	s.rpcContext.HeartbeatCB = func() {
		if err := s.rpcContext.RemoteClocks.VerifyClockOffset(); err != nil {
			log.Fatal(ctx, err)
		}
	}
	s.grpc = rpc.NewServer(s.rpcContext)

	s.registry = metric.NewRegistry()
	s.gossip = gossip.New(
		s.cfg.AmbientCtx,
		&s.nodeIDContainer,
		s.rpcContext,
		s.grpc,
		s.cfg.GossipBootstrapResolvers,
		s.stopper,
		s.registry,
	)
	s.storePool = storage.NewStorePool(
		s.cfg.AmbientCtx,
		s.gossip,
		s.clock,
		s.rpcContext,
		s.cfg.TimeUntilStoreDead,
		s.stopper,
		/* deterministic */ false,
	)

	// A custom RetryOptions is created which uses stopper.ShouldQuiesce() as
	// the Closer. This prevents infinite retry loops from occurring during
	// graceful server shutdown
	//
	// Such a loop loop occurs with the DistSender attempts a connection to the
	// local server during shutdown, and receives an internal server error (HTTP
	// Code 5xx). This is the correct error for a server to return when it is
	// shutting down, and is normally retryable in a cluster environment.
	// However, on a single-node setup (such as a test), retries will never
	// succeed because the only server has been shut down; thus, thus the
	// DistSender needs to know that it should not retry in this situation.
	retryOpts := base.DefaultRetryOptions()
	retryOpts.Closer = s.stopper.ShouldQuiesce()
	distSenderCfg := kv.DistSenderConfig{
		AmbientCtx:      s.cfg.AmbientCtx,
		Clock:           s.clock,
		RPCContext:      s.rpcContext,
		RPCRetryOptions: &retryOpts,
	}
	s.distSender = kv.NewDistSender(distSenderCfg, s.gossip)

	txnMetrics := kv.MakeTxnMetrics(s.cfg.MetricsSampleInterval)
	s.registry.AddMetricStruct(txnMetrics)
	s.txnCoordSender = kv.NewTxnCoordSender(
		s.cfg.AmbientCtx,
		s.distSender,
		s.clock,
		s.cfg.Linearizable,
		s.stopper,
		txnMetrics,
	)
	s.db = client.NewDB(s.txnCoordSender)

	// Use the range lease expiration and renewal durations as the node
	// liveness expiration and heartbeat interval.
	active, renewal := storage.RangeLeaseDurations(
		storage.RaftElectionTimeout(s.cfg.RaftTickInterval, s.cfg.RaftElectionTimeoutTicks))
	s.nodeLiveness = storage.NewNodeLiveness(
		s.cfg.AmbientCtx, s.clock, s.db, s.gossip, active, renewal,
	)
	s.registry.AddMetricStruct(s.nodeLiveness.Metrics())

	s.raftTransport = storage.NewRaftTransport(
		s.cfg.AmbientCtx, storage.GossipAddressResolver(s.gossip), s.grpc, s.rpcContext,
	)

	s.kvDB = kv.NewDBServer(s.cfg.Config, s.txnCoordSender, s.stopper)
	roachpb.RegisterExternalServer(s.grpc, s.kvDB)

	// Set up internal memory metrics for use by internal SQL executors.
	s.internalMemMetrics = sql.MakeMemMetrics("internal")
	s.registry.AddMetricStruct(s.internalMemMetrics)

	// Set up Lease Manager
	var lmKnobs sql.LeaseManagerTestingKnobs
	if cfg.TestingKnobs.SQLLeaseManager != nil {
		lmKnobs = *s.cfg.TestingKnobs.SQLLeaseManager.(*sql.LeaseManagerTestingKnobs)
	}
	s.leaseMgr = sql.NewLeaseManager(&s.nodeIDContainer, *s.db, s.clock, lmKnobs,
		s.stopper, &s.internalMemMetrics)
	s.leaseMgr.RefreshLeases(s.stopper, s.db, s.gossip)

	// Set up the DistSQL server
	distSQLCfg := distsql.ServerConfig{
		AmbientContext: s.cfg.AmbientCtx,
		DB:             s.db,
		RPCContext:     s.rpcContext,
		Stopper:        s.stopper,
	}
	s.distSQLServer = distsql.NewServer(distSQLCfg)
	distsql.RegisterDistSQLServer(s.grpc, s.distSQLServer)

	// Set up admin memory metrics for use by admin SQL executors.
	s.adminMemMetrics = sql.MakeMemMetrics("admin")
	s.registry.AddMetricStruct(s.adminMemMetrics)

	// Set up Executor
	execCfg := sql.ExecutorConfig{
		AmbientCtx:            s.cfg.AmbientCtx,
		NodeID:                &s.nodeIDContainer,
		DB:                    s.db,
		Gossip:                s.gossip,
		LeaseManager:          s.leaseMgr,
		Clock:                 s.clock,
		DistSQLSrv:            s.distSQLServer,
		MetricsSampleInterval: s.cfg.MetricsSampleInterval,
	}
	if s.cfg.TestingKnobs.SQLExecutor != nil {
		execCfg.TestingKnobs = s.cfg.TestingKnobs.SQLExecutor.(*sql.ExecutorTestingKnobs)
	} else {
		execCfg.TestingKnobs = &sql.ExecutorTestingKnobs{}
	}
	if s.cfg.TestingKnobs.SQLSchemaChanger != nil {
		execCfg.SchemaChangerTestingKnobs =
			s.cfg.TestingKnobs.SQLSchemaChanger.(*sql.SchemaChangerTestingKnobs)
	} else {
		execCfg.SchemaChangerTestingKnobs = &sql.SchemaChangerTestingKnobs{}
	}
	s.sqlExecutor = sql.NewExecutor(execCfg, s.stopper, &s.adminMemMetrics)
	s.registry.AddMetricStruct(s.sqlExecutor)

	s.pgServer = pgwire.MakeServer(
		s.cfg.AmbientCtx, s.cfg.Config, s.sqlExecutor, &s.internalMemMetrics, s.cfg.SQLMemoryPoolSize,
	)
	s.registry.AddMetricStruct(s.pgServer.Metrics())

	s.tsDB = ts.NewDB(s.db)
	s.tsServer = ts.MakeServer(s.cfg.AmbientCtx, s.tsDB, s.cfg.TimeSeriesServerConfig, s.stopper)

	// TODO(bdarnell): make StoreConfig configurable.
	storeCfg := storage.StoreConfig{
		AmbientCtx:                     s.cfg.AmbientCtx,
		Clock:                          s.clock,
		DB:                             s.db,
		Gossip:                         s.gossip,
		NodeLiveness:                   s.nodeLiveness,
		Transport:                      s.raftTransport,
		RaftTickInterval:               s.cfg.RaftTickInterval,
		ScanInterval:                   s.cfg.ScanInterval,
		ScanMaxIdleTime:                s.cfg.ScanMaxIdleTime,
		ConsistencyCheckInterval:       s.cfg.ConsistencyCheckInterval,
		ConsistencyCheckPanicOnFailure: s.cfg.ConsistencyCheckPanicOnFailure,
		MetricsSampleInterval:          s.cfg.MetricsSampleInterval,
		StorePool:                      s.storePool,
		SQLExecutor: sql.InternalExecutor{
			LeaseManager: s.leaseMgr,
		},
		LogRangeEvents: s.cfg.EventLogEnabled,
		AllocatorOptions: storage.AllocatorOptions{
			AllowRebalance: true,
		},
		RangeLeaseActiveDuration:  active,
		RangeLeaseRenewalDuration: renewal,
		TimeSeriesDataStore:       s.tsDB,
	}
	if s.cfg.TestingKnobs.Store != nil {
		storeCfg.TestingKnobs = *s.cfg.TestingKnobs.Store.(*storage.StoreTestingKnobs)
	}

	s.recorder = status.NewMetricsRecorder(s.clock)
	s.registry.AddMetricStruct(s.rpcContext.RemoteClocks.Metrics())

	s.runtime = status.MakeRuntimeStatSampler(s.clock)
	s.registry.AddMetricStruct(s.runtime)

	s.node = NewNode(storeCfg, s.recorder, s.registry, s.stopper, txnMetrics, sql.MakeEventLogger(s.leaseMgr))
	roachpb.RegisterInternalServer(s.grpc, s.node)
	storage.RegisterConsistencyServer(s.grpc, s.node.storesServer)
	storage.RegisterFreezeServer(s.grpc, s.node.storesServer)

	s.admin = newAdminServer(s)
	s.status = newStatusServer(
		s.cfg.AmbientCtx, s.db, s.gossip, s.recorder, s.rpcContext, s.node.stores,
	)
	for _, gw := range []grpcGatewayServer{s.admin, s.status, &s.tsServer} {
		gw.RegisterService(s.grpc)
	}

	return s, nil
}