// recordJoinEvent begins an asynchronous task which attempts to log a "node
// join" or "node restart" event. This query will retry until it succeeds or the
// server stops.
func (n *Node) recordJoinEvent() {
if !n.ctx.LogRangeEvents {
return
}
logEventType := sql.EventLogNodeRestart
if n.initialBoot {
logEventType = sql.EventLogNodeJoin
}
n.stopper.RunWorker(func() {
for r := retry.Start(retry.Options{Closer: n.stopper.ShouldStop()}); r.Next(); {
if err := n.ctx.DB.Txn(func(txn *client.Txn) *roachpb.Error {
return sql.MakeEventLogger(n.ctx.SQLExecutor.LeaseManager).InsertEventRecord(txn,
logEventType,
int32(n.Descriptor.NodeID),
int32(n.Descriptor.NodeID),
struct {
Descriptor roachpb.NodeDescriptor
ClusterID uuid.UUID
StartedAt int64
}{n.Descriptor, n.ClusterID, n.startedAt},
)
}); err != nil {
log.Warningc(n.context(context.TODO()), "unable to log %s event for node %d: %s", logEventType, n.Descriptor.NodeID, err)
} else {
return
}
}
})
}
// createTestNode creates an rpc server using the specified address,
// gossip instance, KV database and a node using the specified slice
// of engines. The server, clock and node are returned. If gossipBS is
// not nil, the gossip bootstrap address is set to gossipBS.
func createTestNode(addr net.Addr, engines []engine.Engine, gossipBS net.Addr, t *testing.T) (
*grpc.Server, net.Addr, *hlc.Clock, *Node, *stop.Stopper) {
ctx := storage.StoreContext{}
stopper := stop.NewStopper()
ctx.Clock = hlc.NewClock(hlc.UnixNano)
nodeRPCContext := rpc.NewContext(nodeTestBaseContext, ctx.Clock, stopper)
ctx.ScanInterval = 10 * time.Hour
ctx.ConsistencyCheckInterval = 10 * time.Hour
grpcServer := rpc.NewServer(nodeRPCContext)
serverCtx := makeTestContext()
g := gossip.New(
context.Background(),
nodeRPCContext,
grpcServer,
serverCtx.GossipBootstrapResolvers,
stopper,
metric.NewRegistry())
ln, err := netutil.ListenAndServeGRPC(stopper, grpcServer, addr)
if err != nil {
t.Fatal(err)
}
if gossipBS != nil {
// Handle possibility of a :0 port specification.
if gossipBS.Network() == addr.Network() && gossipBS.String() == addr.String() {
gossipBS = ln.Addr()
}
r, err := resolver.NewResolverFromAddress(gossipBS)
if err != nil {
t.Fatalf("bad gossip address %s: %s", gossipBS, err)
}
g.SetResolvers([]resolver.Resolver{r})
g.Start(ln.Addr())
}
ctx.Gossip = g
retryOpts := base.DefaultRetryOptions()
retryOpts.Closer = stopper.ShouldQuiesce()
distSender := kv.NewDistSender(&kv.DistSenderConfig{
Clock: ctx.Clock,
RPCContext: nodeRPCContext,
RPCRetryOptions: &retryOpts,
}, g)
ctx.Ctx = tracing.WithTracer(context.Background(), tracing.NewTracer())
sender := kv.NewTxnCoordSender(ctx.Ctx, distSender, ctx.Clock, false, stopper,
kv.MakeTxnMetrics())
ctx.DB = client.NewDB(sender)
ctx.Transport = storage.NewDummyRaftTransport()
node := NewNode(ctx, status.NewMetricsRecorder(ctx.Clock), metric.NewRegistry(), stopper,
kv.MakeTxnMetrics(), sql.MakeEventLogger(nil))
roachpb.RegisterInternalServer(grpcServer, node)
return grpcServer, ln.Addr(), ctx.Clock, node, stopper
}
// NewServer creates a Server from a server.Context.
func NewServer(srvCtx Context, stopper *stop.Stopper) (*Server, error) {
if _, err := net.ResolveTCPAddr("tcp", srvCtx.Addr); err != nil {
return nil, errors.Errorf("unable to resolve RPC address %q: %v", srvCtx.Addr, err)
}
if srvCtx.Ctx == nil {
srvCtx.Ctx = context.Background()
}
if srvCtx.Ctx.Done() != nil {
panic("context with cancel or deadline")
}
if tracing.TracerFromCtx(srvCtx.Ctx) == nil {
// TODO(radu): instead of modifying srvCtx.Ctx, we should have a separate
// context.Context inside Server. We will need to rename server.Context
// though.
srvCtx.Ctx = tracing.WithTracer(srvCtx.Ctx, tracing.NewTracer())
}
if srvCtx.Insecure {
log.Warning(srvCtx.Ctx, "running in insecure mode, this is strongly discouraged. See --insecure.")
}
// Try loading the TLS configs before anything else.
if _, err := srvCtx.GetServerTLSConfig(); err != nil {
return nil, err
}
if _, err := srvCtx.GetClientTLSConfig(); err != nil {
return nil, err
}
s := &Server{
mux: http.NewServeMux(),
clock: hlc.NewClock(hlc.UnixNano),
stopper: stopper,
}
// Add a dynamic log tag value for the node ID.
//
// We need to pass the server's Ctx as a base context for 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 contexts in the components (various background
// processes will have already started using the contexts).
//
// The dynamic value 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.
s.nodeLogTagVal.Set(log.DynamicIntValueUnknown)
srvCtx.Ctx = log.WithLogTag(srvCtx.Ctx, "n", &s.nodeLogTagVal)
s.ctx = srvCtx
s.clock.SetMaxOffset(srvCtx.MaxOffset)
s.rpcContext = rpc.NewContext(srvCtx.Context, s.clock, s.stopper)
s.rpcContext.HeartbeatCB = func() {
if err := s.rpcContext.RemoteClocks.VerifyClockOffset(); err != nil {
log.Fatal(s.Ctx(), err)
}
}
s.grpc = rpc.NewServer(s.rpcContext)
s.registry = metric.NewRegistry()
s.gossip = gossip.New(
s.Ctx(), s.rpcContext, s.grpc, s.ctx.GossipBootstrapResolvers, s.stopper, s.registry)
s.storePool = storage.NewStorePool(
s.gossip,
s.clock,
s.rpcContext,
srvCtx.ReservationsEnabled,
srvCtx.TimeUntilStoreDead,
s.stopper,
)
// 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{
Ctx: s.Ctx(),
Clock: s.clock,
RPCContext: s.rpcContext,
RPCRetryOptions: &retryOpts,
}
s.distSender = kv.NewDistSender(&distSenderCfg, s.gossip)
txnMetrics := kv.MakeTxnMetrics()
s.registry.AddMetricStruct(txnMetrics)
s.txnCoordSender = kv.NewTxnCoordSender(s.Ctx(), s.distSender, s.clock, srvCtx.Linearizable,
s.stopper, txnMetrics)
s.db = client.NewDB(s.txnCoordSender)
s.raftTransport = storage.NewRaftTransport(storage.GossipAddressResolver(s.gossip), s.grpc, s.rpcContext)
s.kvDB = kv.NewDBServer(s.ctx.Context, s.txnCoordSender, s.stopper)
roachpb.RegisterExternalServer(s.grpc, s.kvDB)
// Set up Lease Manager
var lmKnobs sql.LeaseManagerTestingKnobs
if srvCtx.TestingKnobs.SQLLeaseManager != nil {
lmKnobs = *srvCtx.TestingKnobs.SQLLeaseManager.(*sql.LeaseManagerTestingKnobs)
}
s.leaseMgr = sql.NewLeaseManager(0, *s.db, s.clock, lmKnobs, s.stopper)
s.leaseMgr.RefreshLeases(s.stopper, s.db, s.gossip)
// Set up the DistSQL server
distSQLCfg := distsql.ServerConfig{
Context: s.Ctx(),
DB: s.db,
RPCContext: s.rpcContext,
}
s.distSQLServer = distsql.NewServer(distSQLCfg)
distsql.RegisterDistSQLServer(s.grpc, s.distSQLServer)
// Set up Executor
execCfg := sql.ExecutorConfig{
Context: s.Ctx(),
DB: s.db,
Gossip: s.gossip,
LeaseManager: s.leaseMgr,
Clock: s.clock,
DistSQLSrv: s.distSQLServer,
}
if srvCtx.TestingKnobs.SQLExecutor != nil {
execCfg.TestingKnobs = srvCtx.TestingKnobs.SQLExecutor.(*sql.ExecutorTestingKnobs)
} else {
execCfg.TestingKnobs = &sql.ExecutorTestingKnobs{}
}
s.sqlExecutor = sql.NewExecutor(execCfg, s.stopper)
s.registry.AddMetricStruct(s.sqlExecutor)
s.pgServer = pgwire.MakeServer(s.ctx.Context, s.sqlExecutor)
s.registry.AddMetricStruct(s.pgServer.Metrics())
// TODO(bdarnell): make StoreConfig configurable.
nCtx := storage.StoreContext{
Ctx: s.Ctx(),
Clock: s.clock,
DB: s.db,
Gossip: s.gossip,
Transport: s.raftTransport,
RaftTickInterval: s.ctx.RaftTickInterval,
ScanInterval: s.ctx.ScanInterval,
ScanMaxIdleTime: s.ctx.ScanMaxIdleTime,
ConsistencyCheckInterval: s.ctx.ConsistencyCheckInterval,
ConsistencyCheckPanicOnFailure: s.ctx.ConsistencyCheckPanicOnFailure,
StorePool: s.storePool,
SQLExecutor: sql.InternalExecutor{
LeaseManager: s.leaseMgr,
},
LogRangeEvents: true,
AllocatorOptions: storage.AllocatorOptions{
AllowRebalance: true,
},
}
if srvCtx.TestingKnobs.Store != nil {
nCtx.TestingKnobs = *srvCtx.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(nCtx, s.recorder, s.registry, s.stopper, txnMetrics, sql.MakeEventLogger(s.leaseMgr))
roachpb.RegisterInternalServer(s.grpc, s.node)
storage.RegisterStoresServer(s.grpc, s.node.storesServer)
s.tsDB = ts.NewDB(s.db)
s.tsServer = ts.MakeServer(s.tsDB)
s.admin = makeAdminServer(s)
s.status = newStatusServer(s.db, s.gossip, s.recorder, s.ctx.Context, s.rpcContext, s.node.stores)
for _, gw := range []grpcGatewayServer{&s.admin, s.status, &s.tsServer} {
gw.RegisterService(s.grpc)
}
return s, nil
}
// NewServer creates a Server from a server.Context.
func NewServer(ctx Context, stopper *stop.Stopper) (*Server, error) {
if _, err := net.ResolveTCPAddr("tcp", ctx.Addr); err != nil {
return nil, errors.Errorf("unable to resolve RPC address %q: %v", ctx.Addr, err)
}
if ctx.Insecure {
log.Warning(context.TODO(), "running in insecure mode, this is strongly discouraged. See --insecure.")
}
// Try loading the TLS configs before anything else.
if _, err := ctx.GetServerTLSConfig(); err != nil {
return nil, err
}
if _, err := ctx.GetClientTLSConfig(); err != nil {
return nil, err
}
s := &Server{
Tracer: tracing.NewTracer(),
ctx: ctx,
mux: http.NewServeMux(),
clock: hlc.NewClock(hlc.UnixNano),
stopper: stopper,
}
s.clock.SetMaxOffset(ctx.MaxOffset)
s.rpcContext = rpc.NewContext(ctx.Context, s.clock, s.stopper)
s.rpcContext.HeartbeatCB = func() {
if err := s.rpcContext.RemoteClocks.VerifyClockOffset(); err != nil {
log.Fatal(context.TODO(), err)
}
}
s.grpc = rpc.NewServer(s.rpcContext)
s.registry = metric.NewRegistry()
s.gossip = gossip.New(s.rpcContext, s.grpc, s.ctx.GossipBootstrapResolvers, s.stopper, s.registry)
s.storePool = storage.NewStorePool(
s.gossip,
s.clock,
s.rpcContext,
ctx.ReservationsEnabled,
ctx.TimeUntilStoreDead,
s.stopper,
)
// 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()
s.distSender = kv.NewDistSender(&kv.DistSenderContext{
Clock: s.clock,
RPCContext: s.rpcContext,
RPCRetryOptions: &retryOpts,
}, s.gossip)
txnMetrics := kv.NewTxnMetrics(s.registry)
sender := kv.NewTxnCoordSender(s.distSender, s.clock, ctx.Linearizable, s.Tracer,
s.stopper, txnMetrics)
s.db = client.NewDB(sender)
s.raftTransport = storage.NewRaftTransport(storage.GossipAddressResolver(s.gossip), s.grpc, s.rpcContext)
s.kvDB = kv.NewDBServer(s.ctx.Context, sender, s.stopper)
roachpb.RegisterExternalServer(s.grpc, s.kvDB)
// Set up Lease Manager
var lmKnobs sql.LeaseManagerTestingKnobs
if ctx.TestingKnobs.SQLLeaseManager != nil {
lmKnobs = *ctx.TestingKnobs.SQLLeaseManager.(*sql.LeaseManagerTestingKnobs)
}
s.leaseMgr = sql.NewLeaseManager(0, *s.db, s.clock, lmKnobs, s.stopper)
s.leaseMgr.RefreshLeases(s.stopper, s.db, s.gossip)
// Set up the DistSQL server
distSQLCtx := distsql.ServerContext{
Context: context.Background(),
DB: s.db,
RPCContext: s.rpcContext,
}
s.distSQLServer = distsql.NewServer(distSQLCtx)
distsql.RegisterDistSQLServer(s.grpc, s.distSQLServer)
// Set up Executor
eCtx := sql.ExecutorContext{
Context: context.Background(),
DB: s.db,
Gossip: s.gossip,
LeaseManager: s.leaseMgr,
Clock: s.clock,
DistSQLSrv: s.distSQLServer,
}
if ctx.TestingKnobs.SQLExecutor != nil {
eCtx.TestingKnobs = ctx.TestingKnobs.SQLExecutor.(*sql.ExecutorTestingKnobs)
} else {
eCtx.TestingKnobs = &sql.ExecutorTestingKnobs{}
}
s.sqlExecutor = sql.NewExecutor(eCtx, s.stopper, s.registry)
s.pgServer = pgwire.MakeServer(s.ctx.Context, s.sqlExecutor, s.registry)
// TODO(bdarnell): make StoreConfig configurable.
nCtx := storage.StoreContext{
Clock: s.clock,
DB: s.db,
Gossip: s.gossip,
Transport: s.raftTransport,
RaftTickInterval: s.ctx.RaftTickInterval,
ScanInterval: s.ctx.ScanInterval,
ScanMaxIdleTime: s.ctx.ScanMaxIdleTime,
ConsistencyCheckInterval: s.ctx.ConsistencyCheckInterval,
ConsistencyCheckPanicOnFailure: s.ctx.ConsistencyCheckPanicOnFailure,
Tracer: s.Tracer,
StorePool: s.storePool,
SQLExecutor: sql.InternalExecutor{
LeaseManager: s.leaseMgr,
},
LogRangeEvents: true,
AllocatorOptions: storage.AllocatorOptions{
AllowRebalance: true,
},
}
if ctx.TestingKnobs.Store != nil {
nCtx.TestingKnobs = *ctx.TestingKnobs.Store.(*storage.StoreTestingKnobs)
}
s.recorder = status.NewMetricsRecorder(s.clock)
s.rpcContext.RemoteClocks.RegisterMetrics(s.registry)
s.runtime = status.MakeRuntimeStatSampler(s.clock, s.registry)
s.node = NewNode(nCtx, s.recorder, s.registry, s.stopper, txnMetrics, sql.MakeEventLogger(s.leaseMgr))
roachpb.RegisterInternalServer(s.grpc, s.node)
roachpb.RegisterInternalStoresServer(s.grpc, s.node.InternalStoresServer)
s.tsDB = ts.NewDB(s.db)
s.tsServer = ts.MakeServer(s.tsDB)
s.admin = makeAdminServer(s)
s.status = newStatusServer(s.db, s.gossip, s.recorder, s.ctx.Context, s.rpcContext, s.node.stores)
for _, gw := range []grpcGatewayServer{&s.admin, s.status, &s.tsServer} {
gw.RegisterService(s.grpc)
}
return s, nil
}