// TestRequestToUninitializedRange tests the behavior when a request
// is sent to a node which should be a replica of the correct range
// but has not yet received its initial snapshot. This would
// previously panic due to a malformed error response from the server,
// as seen in https://github.com/cockroachdb/cockroach/issues/6027.
//
// Prior to the other changes in the commit that introduced it, this
// test would reliable trigger the panic from #6027. However, it
// relies on some hacky tricks to both trigger the panic and shut down
// cleanly. If this test needs a lot of maintenance in the future we
// should be willing to get rid of it.
func TestRequestToUninitializedRange(t *testing.T) {
defer leaktest.AfterTest(t)()
s := server.TestServer{StoresPerNode: 2}
if err := s.Start(); err != nil {
t.Fatalf("Could not start server: %v", err)
}
defer s.Stop()
// Choose a range ID that is much larger than any that would be
// created by initial splits.
const rangeID = roachpb.RangeID(1000)
// Set up a range with replicas on two stores of the same node. This
// ensures that the DistSender will consider both replicas healthy
// and will try to talk to both (so we can get a non-retryable error
// from the second store).
replica1 := roachpb.ReplicaDescriptor{
NodeID: 1,
StoreID: 1,
ReplicaID: 1,
}
replica2 := roachpb.ReplicaDescriptor{
NodeID: 1,
StoreID: 2,
ReplicaID: 2,
}
// HACK: remove the second store from the node to generate a
// non-retryable error when we try to talk to it.
store2, err := s.Stores().GetStore(2)
if err != nil {
t.Fatal(err)
}
s.Stores().RemoveStore(store2)
// Create the uninitialized range by sending an isolated raft
// message to the first store.
conn, err := s.RPCContext().GRPCDial(s.ServingAddr())
if err != nil {
t.Fatal(err)
}
raftClient := storage.NewMultiRaftClient(conn)
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
stream, err := raftClient.RaftMessage(ctx)
if err != nil {
t.Fatal(err)
}
msg := storage.RaftMessageRequest{
GroupID: rangeID,
ToReplica: replica1,
FromReplica: replica2,
Message: raftpb.Message{
Type: raftpb.MsgApp,
To: 1,
},
}
if err := stream.Send(&msg); err != nil {
t.Fatal(err)
}
// Make sure the replica was created.
store1, err := s.Stores().GetStore(1)
if err != nil {
t.Fatal(err)
}
util.SucceedsSoon(t, func() error {
if replica, err := store1.GetReplica(rangeID); err != nil {
return util.Errorf("failed to look up replica: %s", err)
} else if replica.IsInitialized() {
return util.Errorf("expected replica to be uninitialized")
}
return nil
})
// Create our own DistSender so we can force some requests to the
// bogus range. The DistSender needs to be in scope for its own
// MockRangeDescriptorDB closure.
var sender *kv.DistSender
sender = kv.NewDistSender(&kv.DistSenderContext{
Clock: s.Clock(),
RPCContext: s.RPCContext(),
RangeDescriptorDB: kv.MockRangeDescriptorDB(
func(key roachpb.RKey, considerIntents, useReverseScan bool,
) ([]roachpb.RangeDescriptor, []roachpb.RangeDescriptor, *roachpb.Error) {
if key.Equal(roachpb.RKeyMin) {
// Pass through requests for the first range to the real sender.
desc, err := sender.FirstRange()
if err != nil {
return nil, nil, roachpb.NewError(err)
}
return []roachpb.RangeDescriptor{*desc}, nil, nil
}
return []roachpb.RangeDescriptor{{
RangeID: rangeID,
StartKey: roachpb.RKey(keys.Meta2Prefix),
EndKey: roachpb.RKeyMax,
Replicas: []roachpb.ReplicaDescriptor{replica1, replica2},
}}, nil, nil
}),
}, s.Gossip())
// Only inconsistent reads triggered the panic in #6027.
hdr := roachpb.Header{
ReadConsistency: roachpb.INCONSISTENT,
}
req := roachpb.NewGet(roachpb.Key("asdf"))
// Repeat the test a few times: due to the randomization between the
// two replicas, each attempt only had a 50% chance of triggering
// the panic.
for i := 0; i < 5; i++ {
_, pErr := client.SendWrappedWith(sender, context.Background(), hdr, req)
// Each attempt fails with "store 2 not found" because that is the
// non-retryable error.
if !testutils.IsPError(pErr, "store 2 not found") {
t.Fatal(pErr)
}
}
}
// processQueue creates a client and sends messages from its designated queue
// via that client, exiting when the client fails or when it idles out. All
// messages remaining in the queue at that point are lost and a new instance of
// processQueue should be started by the next message to be sent.
// TODO(tschottdorf) should let raft know if the node is down;
// need a feedback mechanism for that. Potentially easiest is to arrange for
// the next call to Send() to fail appropriately.
func (t *rpcTransport) processQueue(nodeID roachpb.NodeID, storeID roachpb.StoreID) {
t.mu.Lock()
ch, ok := t.queues[storeID]
t.mu.Unlock()
if !ok {
return
}
// Clean-up when the loop below shuts down.
defer func() {
t.mu.Lock()
delete(t.queues, storeID)
t.mu.Unlock()
}()
addr, err := t.gossip.GetNodeIDAddress(nodeID)
if err != nil {
if log.V(1) {
log.Errorf("could not get address for node %d: %s", nodeID, err)
}
return
}
var dialOpt grpc.DialOption
if t.rpcContext.Insecure {
dialOpt = grpc.WithInsecure()
} else {
tlsConfig, err := t.rpcContext.GetClientTLSConfig()
if err != nil {
log.Error(err)
return
}
dialOpt = grpc.WithTransportCredentials(credentials.NewTLS(tlsConfig))
}
conn, err := grpc.Dial(addr.String(), dialOpt)
if err != nil {
log.Errorf("failed to dial: %v", err)
return
}
defer func() {
if err := conn.Close(); err != nil {
log.Error(err)
}
}()
client := storage.NewMultiRaftClient(conn)
ctx := grpcutil.NewContextWithStopper(context.Background(), t.rpcContext.Stopper)
stream, err := client.RaftMessage(ctx)
if err != nil {
log.Error(err)
return
}
defer func() {
if err := stream.CloseSend(); err != nil {
log.Error(err)
}
}()
var raftIdleTimer util.Timer
defer raftIdleTimer.Stop()
for {
raftIdleTimer.Reset(raftIdleTimeout)
select {
case <-ctx.Done():
return
case <-raftIdleTimer.C:
raftIdleTimer.Read = true
if log.V(1) {
log.Infof("closing Raft transport to %d due to inactivity", nodeID)
}
return
case req := <-ch:
if err := stream.Send(req); err != nil {
log.Error(err)
return
}
}
}
}