etcd/raft/node.go

274 lines
6.0 KiB
Go

package raft
import (
"encoding/binary"
"encoding/json"
"log"
"math/rand"
"sort"
"time"
)
type Interface interface {
Step(m Message) bool
Msgs() []Message
}
type tick int64
type Config struct {
NodeId int64
Addr string
Context []byte
}
type Node struct {
sm *stateMachine
elapsed tick
electionRand tick
election tick
heartbeat tick
// TODO: it needs garbage collection later
rmNodes map[int64]struct{}
removed bool
}
func New(id int64, heartbeat, election tick) *Node {
if election < heartbeat*3 {
panic("election is least three times as heartbeat [election: %d, heartbeat: %d]")
}
rand.Seed(time.Now().UnixNano())
n := &Node{
heartbeat: heartbeat,
election: election,
electionRand: election + tick(rand.Int31())%election,
sm: newStateMachine(id, []int64{id}),
rmNodes: make(map[int64]struct{}),
}
return n
}
func Recover(id int64, s *Snapshot, ents []Entry, state State, heartbeat, election tick) *Node {
n := New(id, heartbeat, election)
if s != nil {
n.sm.restore(*s)
}
n.sm.loadEnts(ents)
if !state.IsEmpty() {
n.sm.loadState(state)
}
return n
}
func (n *Node) Id() int64 { return n.sm.id }
func (n *Node) ClusterId() int64 { return n.sm.clusterId }
func (n *Node) Info() Info {
return Info{Id: n.Id()}
}
func (n *Node) Index() int64 { return n.sm.index.Get() }
func (n *Node) Term() int64 { return n.sm.term.Get() }
func (n *Node) Applied() int64 { return n.sm.raftLog.applied }
func (n *Node) HasLeader() bool { return n.Leader() != none }
func (n *Node) IsLeader() bool { return n.Leader() == n.Id() }
func (n *Node) Leader() int64 { return n.sm.lead.Get() }
func (n *Node) IsRemoved() bool { return n.removed }
func (n *Node) Nodes() []int64 {
nodes := make(int64Slice, 0, len(n.sm.ins))
for k := range n.sm.ins {
nodes = append(nodes, k)
}
sort.Sort(nodes)
return nodes
}
// Propose asynchronously proposes data be applied to the underlying state machine.
func (n *Node) Propose(data []byte) { n.propose(Normal, data) }
func (n *Node) propose(t int64, data []byte) {
n.Step(Message{From: n.sm.id, ClusterId: n.ClusterId(), Type: msgProp, Entries: []Entry{{Type: t, Data: data}}})
}
func (n *Node) Campaign() { n.Step(Message{From: n.sm.id, ClusterId: n.ClusterId(), Type: msgHup}) }
func (n *Node) InitCluster(clusterId int64) {
d := make([]byte, 10)
wn := binary.PutVarint(d, clusterId)
n.propose(ClusterInit, d[:wn])
}
func (n *Node) Add(id int64, addr string, context []byte) {
n.UpdateConf(AddNode, &Config{NodeId: id, Addr: addr, Context: context})
}
func (n *Node) Remove(id int64) {
n.UpdateConf(RemoveNode, &Config{NodeId: id})
}
func (n *Node) Msgs() []Message { return n.sm.Msgs() }
func (n *Node) Step(m Message) bool {
if m.Type == msgDenied {
n.removed = true
return false
}
if n.ClusterId() != none && m.ClusterId != none && m.ClusterId != n.ClusterId() {
log.Printf("deny message from=%d cluster=%d", m.From, m.ClusterId)
n.sm.send(Message{To: m.From, ClusterId: n.ClusterId(), Type: msgDenied})
return true
}
if _, ok := n.rmNodes[m.From]; ok {
if m.From != n.sm.id {
n.sm.send(Message{To: m.From, ClusterId: n.ClusterId(), Type: msgDenied})
}
return true
}
l := len(n.sm.msgs)
if !n.sm.Step(m) {
return false
}
for _, m := range n.sm.msgs[l:] {
switch m.Type {
case msgAppResp:
// We just heard from the leader of the same term.
n.elapsed = 0
case msgVoteResp:
// We just heard from the candidate the node voted for.
if m.Index >= 0 {
n.elapsed = 0
}
}
}
return true
}
// Next returns all the appliable entries
func (n *Node) Next() []Entry {
ents := n.sm.nextEnts()
for i := range ents {
switch ents[i].Type {
case Normal:
case ClusterInit:
cid, nr := binary.Varint(ents[i].Data)
if nr <= 0 {
panic("init cluster failed: cannot read clusterId")
}
if n.ClusterId() != -1 {
panic("cannot init a started cluster")
}
n.sm.clusterId = cid
case AddNode:
c := new(Config)
if err := json.Unmarshal(ents[i].Data, c); err != nil {
log.Printf("raft: err=%q", err)
continue
}
n.sm.addNode(c.NodeId)
delete(n.rmNodes, c.NodeId)
case RemoveNode:
c := new(Config)
if err := json.Unmarshal(ents[i].Data, c); err != nil {
log.Printf("raft: err=%q", err)
continue
}
n.sm.removeNode(c.NodeId)
n.rmNodes[c.NodeId] = struct{}{}
if c.NodeId == n.sm.id {
n.removed = true
}
default:
panic("unexpected entry type")
}
}
return ents
}
// Tick triggers the node to do a tick.
// If the current elapsed is greater or equal than the timeout,
// node will send corresponding message to the statemachine.
func (n *Node) Tick() {
if !n.sm.promotable {
return
}
timeout, msgType := n.electionRand, msgHup
if n.sm.state == stateLeader {
timeout, msgType = n.heartbeat, msgBeat
}
if n.elapsed >= timeout {
n.Step(Message{From: n.sm.id, ClusterId: n.ClusterId(), Type: msgType})
n.elapsed = 0
if n.sm.state != stateLeader {
n.electionRand = n.election + tick(rand.Int31())%n.election
}
} else {
n.elapsed++
}
}
// IsEmpty returns ture if the log of the node is empty.
func (n *Node) IsEmpty() bool {
return n.sm.raftLog.isEmpty()
}
func (n *Node) UpdateConf(t int64, c *Config) {
data, err := json.Marshal(c)
if err != nil {
panic(err)
}
n.propose(t, data)
}
// UnstableEnts retuens all the entries that need to be persistent.
// The first return value is offset, and the second one is unstable entries.
func (n *Node) UnstableEnts() []Entry {
return n.sm.raftLog.unstableEnts()
}
func (n *Node) UnstableState() State {
if n.sm.unstableState.IsEmpty() {
return EmptyState
}
s := n.sm.unstableState
n.sm.clearState()
return s
}
func (n *Node) UnstableSnapshot() Snapshot {
if n.sm.raftLog.unstableSnapshot.IsEmpty() {
return emptySnapshot
}
s := n.sm.raftLog.unstableSnapshot
n.sm.raftLog.unstableSnapshot = emptySnapshot
return s
}
func (n *Node) GetSnap() Snapshot {
return n.sm.raftLog.snapshot
}
func (n *Node) Compact(d []byte) {
n.sm.compact(d)
}
func (n *Node) EntsLen() int {
return len(n.sm.raftLog.ents)
}