// TestMultiNodeStart ensures that a node can be started correctly. The node should
// start with correct configuration change entries, and can accept and commit
// proposals.
func TestMultiNodeStart(t *testing.T) {
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
cc := raftpb.ConfChange{Type: raftpb.ConfChangeAddNode, NodeID: 1}
ccdata, err := cc.Marshal()
if err != nil {
t.Fatalf("unexpected marshal error: %v", err)
}
wants := []Ready{
{
SoftState: &SoftState{Lead: 1, RaftState: StateLeader},
HardState: raftpb.HardState{Term: 2, Commit: 2, Vote: 1},
Entries: []raftpb.Entry{
{Type: raftpb.EntryConfChange, Term: 1, Index: 1, Data: ccdata},
{Term: 2, Index: 2},
},
CommittedEntries: []raftpb.Entry{
{Type: raftpb.EntryConfChange, Term: 1, Index: 1, Data: ccdata},
{Term: 2, Index: 2},
},
},
{
HardState: raftpb.HardState{Term: 2, Commit: 3, Vote: 1},
Entries: []raftpb.Entry{{Term: 2, Index: 3, Data: []byte("foo")}},
CommittedEntries: []raftpb.Entry{{Term: 2, Index: 3, Data: []byte("foo")}},
},
}
mn := StartMultiNode(1)
storage := NewMemoryStorage()
mn.CreateGroup(1, newTestConfig(1, nil, 10, 1, storage), []Peer{{ID: 1}})
mn.Campaign(ctx, 1)
gs := <-mn.Ready()
g := gs[1]
if !reflect.DeepEqual(g, wants[0]) {
t.Fatalf("#%d: g = %+v,\n w %+v", 1, g, wants[0])
} else {
storage.Append(g.Entries)
mn.Advance(gs)
}
mn.Propose(ctx, 1, []byte("foo"))
if gs2 := <-mn.Ready(); !reflect.DeepEqual(gs2[1], wants[1]) {
t.Errorf("#%d: g = %+v,\n w %+v", 2, gs2[1], wants[1])
} else {
storage.Append(gs2[1].Entries)
mn.Advance(gs2)
}
select {
case rd := <-mn.Ready():
t.Errorf("unexpected Ready: %+v", rd)
case <-time.After(time.Millisecond):
}
}
func (mn *multiNode) ProposeConfChange(ctx context.Context, group uint64, cc pb.ConfChange) error {
data, err := cc.Marshal()
if err != nil {
return err
}
return mn.Step(ctx, group,
pb.Message{
Type: pb.MsgProp,
Entries: []pb.Entry{
{Type: pb.EntryConfChange, Data: data},
},
})
}
// TestMultiNodeProposeConfig ensures that multiNode.ProposeConfChange
// sends the given configuration proposal to the underlying raft.
func TestMultiNodeProposeConfig(t *testing.T) {
mn := newMultiNode(1)
go mn.run()
s := NewMemoryStorage()
mn.CreateGroup(1, newTestConfig(1, nil, 10, 1, s), []Peer{{ID: 1}})
mn.Campaign(context.TODO(), 1)
proposed := false
var lastIndex uint64
var ccdata []byte
for {
rds := <-mn.Ready()
rd := rds[1]
s.Append(rd.Entries)
// change the step function to appendStep until this raft becomes leader
if !proposed && rd.SoftState.Lead == mn.id {
cc := raftpb.ConfChange{Type: raftpb.ConfChangeAddNode, NodeID: 1}
var err error
ccdata, err = cc.Marshal()
if err != nil {
t.Fatal(err)
}
mn.ProposeConfChange(context.TODO(), 1, cc)
proposed = true
}
mn.Advance(rds)
var err error
lastIndex, err = s.LastIndex()
if err != nil {
t.Fatal(err)
}
if lastIndex >= 3 {
break
}
}
mn.Stop()
entries, err := s.Entries(lastIndex, lastIndex+1, noLimit)
if err != nil {
t.Fatal(err)
}
if len(entries) != 1 {
t.Fatalf("len(entries) = %d, want %d", len(entries), 1)
}
if entries[0].Type != raftpb.EntryConfChange {
t.Fatalf("type = %v, want %v", entries[0].Type, raftpb.EntryConfChange)
}
if !bytes.Equal(entries[0].Data, ccdata) {
t.Errorf("data = %v, want %v", entries[0].Data, ccdata)
}
}
func (mn *multiNode) run() {
groups := map[uint64]*groupState{}
rds := map[uint64]Ready{}
var advancec chan map[uint64]Ready
for {
// Only select readyc if we have something to report and we are not
// currently waiting for an advance.
readyc := mn.readyc
if len(rds) == 0 || advancec != nil {
readyc = nil
}
// group points to the group that was touched on this iteration (if any)
var group *groupState
select {
case gc := <-mn.groupc:
if (gc.config.ID != mn.id) && (gc.config.ID != 0 && mn.id != 0) {
panic("if gc.config.ID and mn.id differ, one of them must be zero")
}
if gc.config.ID == 0 {
gc.config.ID = mn.id
}
r := newRaft(gc.config)
group = &groupState{
id: gc.id,
raft: r,
}
groups[gc.id] = group
lastIndex, err := gc.config.Storage.LastIndex()
if err != nil {
panic(err) // TODO(bdarnell)
}
// If the log is empty, this is a new group (like StartNode); otherwise it's
// restoring an existing group (like RestartNode).
// TODO(bdarnell): rethink group initialization and whether the application needs
// to be able to tell us when it expects the group to exist.
if lastIndex == 0 {
r.becomeFollower(1, None)
ents := make([]pb.Entry, len(gc.peers))
for i, peer := range gc.peers {
cc := pb.ConfChange{Type: pb.ConfChangeAddNode, NodeID: peer.ID, Context: peer.Context}
data, err := cc.Marshal()
if err != nil {
panic("unexpected marshal error")
}
ents[i] = pb.Entry{Type: pb.EntryConfChange, Term: 1, Index: uint64(i + 1), Data: data}
}
r.raftLog.append(ents...)
r.raftLog.committed = uint64(len(ents))
for _, peer := range gc.peers {
r.addNode(peer.ID)
}
}
// Set the initial hard and soft states after performing all initialization.
group.prevSoftSt = r.softState()
group.prevHardSt = r.HardState
close(gc.done)
case gr := <-mn.rmgroupc:
delete(groups, gr.id)
delete(rds, gr.id)
close(gr.done)
case mm := <-mn.propc:
// TODO(bdarnell): single-node impl doesn't read from propc unless the group
// has a leader; we can't do that since we have one propc for many groups.
// We'll have to buffer somewhere on a group-by-group basis, or just let
// raft.Step drop any such proposals on the floor.
var ok bool
if group, ok = groups[mm.group]; ok {
mm.msg.From = group.raft.id
group.raft.Step(mm.msg)
}
case mm := <-mn.recvc:
group = groups[mm.group]
if _, ok := group.raft.prs[mm.msg.From]; ok || !IsResponseMsg(mm.msg) {
group.raft.Step(mm.msg)
}
case mcc := <-mn.confc:
group = groups[mcc.group]
if mcc.msg.NodeID == None {
group.raft.resetPendingConf()
select {
case mcc.ch <- pb.ConfState{Nodes: group.raft.nodes()}:
case <-mn.done:
}
break
}
switch mcc.msg.Type {
case pb.ConfChangeAddNode:
group.raft.addNode(mcc.msg.NodeID)
case pb.ConfChangeRemoveNode:
group.raft.removeNode(mcc.msg.NodeID)
case pb.ConfChangeUpdateNode:
group.raft.resetPendingConf()
default:
panic("unexpected conf type")
}
select {
case mcc.ch <- pb.ConfState{Nodes: group.raft.nodes()}:
case <-mn.done:
}
case <-mn.tickc:
// TODO(bdarnell): instead of calling every group on every tick,
// we should have a priority queue of groups based on their next
// time-based event.
for _, g := range groups {
g.raft.tick()
rd := g.newReady()
if rd.containsUpdates() {
rds[g.id] = rd
}
}
case readyc <- rds:
// Clear outgoing messages as soon as we've passed them to the application.
for g := range rds {
groups[g].raft.msgs = nil
}
rds = map[uint64]Ready{}
advancec = mn.advancec
case advs := <-advancec:
for groupID, rd := range advs {
g, ok := groups[groupID]
if !ok {
continue
}
g.commitReady(rd)
// We've been accumulating new entries in rds which may now be obsolete.
// Drop the old Ready object and create a new one if needed.
delete(rds, groupID)
newRd := g.newReady()
if newRd.containsUpdates() {
rds[groupID] = newRd
}
}
advancec = nil
case ms := <-mn.status:
if g, ok := groups[ms.group]; ok {
s := getStatus(g.raft)
ms.ch <- &s
} else {
ms.ch <- nil
}
case <-mn.stop:
close(mn.done)
return
}
if group != nil {
rd := group.newReady()
if rd.containsUpdates() {
rds[group.id] = rd
}
}
}
}