519 lines
14 KiB
Markdown
519 lines
14 KiB
Markdown
# MIT 6.824 Lab2: Raft 实验报告
|
||
|
||
> 📅 日期：2026-02-26
|
||
> 🔗 代码仓库：https://git.rc707blog.top/rose_cat707/6.824-golabs-2021-6.824
|
||
> 🌿 分支：answer/20260226
|
||
|
||
---
|
||
|
||
## 1. 实验概述
|
||
|
||
本实验要求实现 Raft 共识算法，这是分布式系统中实现容错的关键组件。Raft 将共识问题分解为三个相对独立的子问题：
|
||
|
||
1. **领导者选举 (Leader Election)** - 2A
|
||
2. **日志复制 (Log Replication)** - 2B
|
||
3. **持久化 (Persistence)** - 2C
|
||
4. **快照 (Snapshots)** - 2D
|
||
|
||
---
|
||
|
||
## 2. 核心数据结构
|
||
|
||
### 2.1 服务器状态
|
||
|
||
```go
|
||
type Raft struct {
|
||
mu sync.Mutex // 保护共享状态的锁
|
||
peers []*labrpc.ClientEnd // 所有节点的 RPC 端点
|
||
persister *Persister // 持久化存储
|
||
me int // 当前节点索引
|
||
dead int32 // 是否被 Kill
|
||
|
||
// 持久状态 (Figure 2)
|
||
currentTerm int // 当前任期
|
||
votedFor int // 投票给谁
|
||
log []LogEntry // 日志条目
|
||
|
||
// 易失状态 - 所有节点
|
||
commitIndex int // 已提交的最高日志索引
|
||
lastApplied int // 已应用的最高日志索引
|
||
|
||
// 易失状态 - 仅领导者
|
||
nextIndex []int // 对每个节点，下一个要发送的日志索引
|
||
matchIndex []int // 对每个节点，已复制的最高日志索引
|
||
|
||
// 服务器状态
|
||
state int // Follower/Candidate/Leader
|
||
lastHeartbeat time.Time // 上次收到心跳的时间
|
||
}
|
||
```
|
||
|
||
### 2.2 日志条目
|
||
|
||
```go
|
||
type LogEntry struct {
|
||
Term int
|
||
Command interface{}
|
||
}
|
||
```
|
||
|
||
---
|
||
|
||
## 3. 2A: Leader Election
|
||
|
||
### 3.1 状态转换
|
||
|
||
```
|
||
┌─────────────┐ ┌─────────────┐
|
||
│ Follower │───超时无心跳──────→│ Candidate │
|
||
└─────────────┘ └─────────────┘
|
||
↑ │
|
||
│ │
|
||
│ ┌─────────────────────────┼─────────────────────────┐
|
||
│ │ │ │
|
||
│ ↓ ↓ ↓
|
||
│ ┌─────────────┐ ┌─────────────┐ 收到更高任期
|
||
└──│ Leader │←──获得多数票──│ Candidate │──────────────────┘
|
||
└─────────────┘ └─────────────┘
|
||
```
|
||
|
||
### 3.2 选举超时
|
||
|
||
```go
|
||
const (
|
||
HeartbeatInterval = 100 * time.Millisecond
|
||
ElectionTimeoutMin = 300 // ms
|
||
ElectionTimeoutMax = 500 // ms
|
||
)
|
||
```
|
||
|
||
随机化选举超时 (300-500ms) 避免同时选举。
|
||
|
||
### 3.3 选举流程
|
||
|
||
```go
|
||
func (rf *Raft) startElection() {
|
||
// 1. 转为 Candidate，增加任期，投自己一票
|
||
rf.state = Candidate
|
||
rf.currentTerm++
|
||
rf.votedFor = rf.me
|
||
rf.persist()
|
||
|
||
// 2. 并行向所有节点发送 RequestVote RPC
|
||
for i := range rf.peers {
|
||
go func(server int) {
|
||
args := RequestVoteArgs{
|
||
Term: term,
|
||
CandidateId: rf.me,
|
||
LastLogIndex: lastLogIndex,
|
||
LastLogTerm: lastLogTerm,
|
||
}
|
||
reply := RequestVoteReply{}
|
||
ok := rf.sendRequestVote(server, &args, &reply)
|
||
// 处理投票响应...
|
||
}(i)
|
||
}
|
||
|
||
// 3. 如果获得多数票，成为 Leader
|
||
if votes > len(rf.peers)/2 {
|
||
rf.state = Leader
|
||
// 初始化 nextIndex 和 matchIndex
|
||
}
|
||
}
|
||
```
|
||
|
||
### 3.4 RequestVote RPC 处理
|
||
|
||
```go
|
||
func (rf *Raft) RequestVote(args *RequestVoteArgs, reply *RequestVoteReply) {
|
||
// 1. 拒绝过期任期的请求
|
||
if args.Term < rf.currentTerm {
|
||
return
|
||
}
|
||
|
||
// 2. 更新任期（如果请求方任期更高）
|
||
if args.Term > rf.currentTerm {
|
||
rf.currentTerm = args.Term
|
||
rf.votedFor = -1
|
||
rf.state = Follower
|
||
rf.persist()
|
||
}
|
||
|
||
// 3. 检查投票条件
|
||
// - 没投给别人，或投的就是这个候选人
|
||
// - 候选人日志至少和自己一样新
|
||
if (rf.votedFor == -1 || rf.votedFor == args.CandidateId) &&
|
||
(args.LastLogTerm > lastLogTerm ||
|
||
(args.LastLogTerm == lastLogTerm && args.LastLogIndex >= lastLogIndex)) {
|
||
reply.VoteGranted = true
|
||
rf.votedFor = args.CandidateId
|
||
rf.lastHeartbeat = time.Now() // 重置选举超时
|
||
rf.persist()
|
||
}
|
||
}
|
||
```
|
||
|
||
---
|
||
|
||
## 4. 2B: Log Replication
|
||
|
||
### 4.1 AppendEntries RPC
|
||
|
||
领导者通过心跳发送 AppendEntries RPC：
|
||
|
||
```go
|
||
type AppendEntriesArgs struct {
|
||
Term int // 领导者任期
|
||
LeaderId int // 领导者 ID
|
||
PrevLogIndex int // 前一个日志索引
|
||
PrevLogTerm int // 前一个日志任期
|
||
Entries []LogEntry // 要追加的日志条目
|
||
LeaderCommit int // 领导者的 commitIndex
|
||
}
|
||
```
|
||
|
||
### 4.2 日志一致性检查
|
||
|
||
Follower 收到 AppendEntries 后：
|
||
|
||
1. 检查 `PrevLogIndex` 处的日志任期是否匹配 `PrevLogTerm`
|
||
2. 如果匹配，追加新条目（删除冲突条目）
|
||
3. 更新 `commitIndex`
|
||
|
||
### 4.3 快速回退优化 (Fast Backup)
|
||
|
||
传统方法：逐个递减 `nextIndex`，效率低。
|
||
|
||
优化方法：Follower 在拒绝时返回额外信息：
|
||
|
||
```go
|
||
type AppendEntriesReply struct {
|
||
Term int
|
||
Success bool
|
||
XTerm int // 冲突条目的任期
|
||
XIndex int // 该任期的第一个索引
|
||
XLen int // 日志长度
|
||
}
|
||
```
|
||
|
||
领导者根据这些信息快速调整 `nextIndex`：
|
||
|
||
```go
|
||
if reply.XTerm != 0 {
|
||
// 在自己日志中找 XTerm
|
||
found := false
|
||
for i := rf.getLastLogIndex(); i > rf.lastIncludedIndex; i-- {
|
||
if rf.getLogTerm(i) == reply.XTerm {
|
||
rf.nextIndex[server] = i + 1
|
||
found = true
|
||
break
|
||
}
|
||
}
|
||
if !found {
|
||
rf.nextIndex[server] = reply.XIndex // 跳到 XTerm 第一个索引
|
||
}
|
||
} else {
|
||
rf.nextIndex[server] = reply.XLen // Follower 日志太短
|
||
}
|
||
```
|
||
|
||
### 4.4 提交规则
|
||
|
||
**关键规则**：只能提交当前任期的日志条目！
|
||
|
||
```go
|
||
func (rf *Raft) updateCommitIndex() {
|
||
for n := rf.getLastLogIndex(); n > rf.commitIndex && n > rf.lastIncludedIndex; n-- {
|
||
if rf.getLogTerm(n) != rf.currentTerm {
|
||
continue // 只提交当前任期的日志
|
||
}
|
||
count := 1 // 自己
|
||
for i := range rf.peers {
|
||
if i != rf.me && rf.matchIndex[i] >= n {
|
||
count++
|
||
}
|
||
}
|
||
if count > len(rf.peers)/2 {
|
||
rf.commitIndex = n
|
||
break
|
||
}
|
||
}
|
||
}
|
||
```
|
||
|
||
---
|
||
|
||
## 5. 2C: Persistence
|
||
|
||
### 5.1 需要持久化的状态
|
||
|
||
根据 Raft 论文 Figure 2：
|
||
|
||
- `currentTerm` - 当前任期
|
||
- `votedFor` - 投票给谁
|
||
- `log[]` - 日志条目
|
||
|
||
### 5.2 持久化实现
|
||
|
||
```go
|
||
func (rf *Raft) persist() {
|
||
w := new(bytes.Buffer)
|
||
e := labgob.NewEncoder(w)
|
||
e.Encode(rf.currentTerm)
|
||
e.Encode(rf.votedFor)
|
||
e.Encode(rf.log)
|
||
e.Encode(rf.lastIncludedIndex) // 2D 新增
|
||
e.Encode(rf.lastIncludedTerm) // 2D 新增
|
||
data := w.Bytes()
|
||
rf.persister.SaveRaftState(data)
|
||
}
|
||
|
||
func (rf *Raft) readPersist(data []byte) {
|
||
if data == nil || len(data) < 1 {
|
||
return
|
||
}
|
||
r := bytes.NewBuffer(data)
|
||
d := labgob.NewDecoder(r)
|
||
var currentTerm, votedFor, lastIncludedIndex, lastIncludedTerm int
|
||
var log []LogEntry
|
||
if d.Decode(¤tTerm) != nil ||
|
||
d.Decode(&votedFor) != nil ||
|
||
d.Decode(&log) != nil ||
|
||
d.Decode(&lastIncludedIndex) != nil ||
|
||
d.Decode(&lastIncludedTerm) != nil {
|
||
// 解码失败
|
||
} else {
|
||
rf.currentTerm = currentTerm
|
||
rf.votedFor = votedFor
|
||
rf.log = log
|
||
rf.lastIncludedIndex = lastIncludedIndex
|
||
rf.lastIncludedTerm = lastIncludedTerm
|
||
}
|
||
}
|
||
```
|
||
|
||
### 5.3 持久化时机
|
||
|
||
每次修改 `currentTerm`、`votedFor` 或 `log` 后都要调用 `persist()`。
|
||
|
||
---
|
||
|
||
## 6. 2D: Snapshots
|
||
|
||
### 6.1 为什么需要快照？
|
||
|
||
随着日志增长，内存和磁盘占用会无限增长。快照机制可以：
|
||
- 截断已应用的日志
|
||
- 减少内存占用
|
||
- 加速新节点同步
|
||
|
||
### 6.2 快照数据结构
|
||
|
||
```go
|
||
type Raft struct {
|
||
// ... 其他字段
|
||
lastIncludedIndex int // 快照包含的最后日志索引
|
||
lastIncludedTerm int // 快照包含的最后日志任期
|
||
}
|
||
```
|
||
|
||
### 6.3 服务端创建快照
|
||
|
||
```go
|
||
func (rf *Raft) Snapshot(index int, snapshot []byte) {
|
||
rf.mu.Lock()
|
||
defer rf.mu.Unlock()
|
||
|
||
// 边界检查
|
||
if index <= rf.lastIncludedIndex || index > rf.getLastLogIndex() {
|
||
return
|
||
}
|
||
|
||
// 先保存 term，再截断日志
|
||
newLastIncludedTerm := rf.getLogTerm(index)
|
||
rf.log = rf.log[index-rf.lastIncludedIndex:]
|
||
rf.lastIncludedTerm = newLastIncludedTerm
|
||
rf.lastIncludedIndex = index
|
||
|
||
// 持久化状态和快照
|
||
rf.persister.SaveStateAndSnapshot(state, snapshot)
|
||
}
|
||
```
|
||
|
||
### 6.4 InstallSnapshot RPC
|
||
|
||
当日志落后于领导者时，领导者发送快照：
|
||
|
||
```go
|
||
type InstallSnapshotArgs struct {
|
||
Term int
|
||
LeaderId int
|
||
LastIncludedIndex int
|
||
LastIncludedTerm int
|
||
Data []byte
|
||
}
|
||
|
||
func (rf *Raft) InstallSnapshot(args *InstallSnapshotArgs, reply *InstallSnapshotReply) {
|
||
// 1. 检查任期
|
||
if args.Term < rf.currentTerm {
|
||
return
|
||
}
|
||
|
||
// 2. 更新状态
|
||
if args.Term > rf.currentTerm {
|
||
rf.currentTerm = args.Term
|
||
rf.votedFor = -1
|
||
rf.persist()
|
||
}
|
||
|
||
// 3. 截断日志
|
||
if args.LastIncludedIndex > rf.lastIncludedIndex {
|
||
if args.LastIncludedIndex <= rf.getLastLogIndex() &&
|
||
rf.getLogTerm(args.LastIncludedIndex) == args.LastIncludedTerm {
|
||
// 保留后面的日志
|
||
rf.log = rf.log[args.LastIncludedIndex-rf.lastIncludedIndex:]
|
||
} else {
|
||
rf.log = make([]LogEntry, 0)
|
||
}
|
||
rf.lastIncludedIndex = args.LastIncludedIndex
|
||
rf.lastIncludedTerm = args.LastIncludedTerm
|
||
rf.persister.SaveStateAndSnapshot(state, args.Data)
|
||
|
||
// 发送到 applyCh
|
||
rf.applyCh <- ApplyMsg{
|
||
SnapshotValid: true,
|
||
Snapshot: args.Data,
|
||
SnapshotIndex: args.LastIncludedIndex,
|
||
SnapshotTerm: args.LastIncludedTerm,
|
||
}
|
||
}
|
||
}
|
||
```
|
||
|
||
---
|
||
|
||
## 7. 关键实现细节
|
||
|
||
### 7.1 并发安全
|
||
|
||
- 使用 `sync.Mutex` 保护所有共享状态
|
||
- 使用 `sync.Cond` 实现应用器的等待/唤醒机制
|
||
- 使用 `atomic` 实现 `Kill()` 和 `killed()`
|
||
|
||
### 7.2 避免活锁
|
||
|
||
- 随机化选举超时 (300-500ms)
|
||
- 心跳间隔 100ms，远小于选举超时
|
||
|
||
### 7.3 正确处理任期
|
||
|
||
- 任期是 Raft 的核心，所有 RPC 都要检查任期
|
||
- 收到更高任期时，立即转为 Follower 并更新任期
|
||
|
||
### 7.4 日志索引处理
|
||
|
||
- 快照后的日志索引需要偏移计算
|
||
- `getLogTerm()` 和 `getLogEntry()` 统一处理索引转换
|
||
|
||
---
|
||
|
||
## 8. 测试结果
|
||
|
||
### 8.1 2A: Leader Election
|
||
|
||
```
|
||
=== RUN TestInitialElection2A
|
||
... Passed -- 3.1 3 54 14568 0
|
||
=== RUN TestReElection2A
|
||
... Passed -- 8.9 3 176 34056 0
|
||
=== RUN TestManyElections2A
|
||
... Passed -- 5.5 7 408 86416 0
|
||
PASS
|
||
```
|
||
|
||
### 8.2 2B: Log Replication
|
||
|
||
```
|
||
=== RUN TestBasicAgree2B
|
||
... Passed -- 1.0 3 16 4310 3
|
||
=== RUN TestRPCBytes2B
|
||
... Passed -- 2.8 3 48 113570 11
|
||
=== RUN TestFailAgree2B
|
||
... Passed -- 6.3 3 116 30885 8
|
||
=== RUN TestFailNoAgree2B
|
||
... Passed -- 3.8 5 148 34328 4
|
||
=== RUN TestConcurrentStarts2B
|
||
... Passed -- 0.8 3 14 3552 6
|
||
=== RUN TestRejoin2B
|
||
... Passed -- 4.5 3 138 31037 4
|
||
=== RUN TestBackup2B
|
||
... Passed -- 33.6 5 2120 1888317 102
|
||
=== RUN TestCount2B
|
||
... Passed -- 2.1 3 36 10066 12
|
||
PASS
|
||
```
|
||
|
||
### 8.3 2C: Persistence
|
||
|
||
```
|
||
=== RUN TestPersist12C
|
||
... Passed -- 4.1 3 76 18768 6
|
||
=== RUN TestPersist22C
|
||
... Passed -- 44.1 5 1812 370976 16
|
||
=== RUN TestPersist32C
|
||
... Passed -- 2.1 3 34 8370 4
|
||
=== RUN TestFigure82C
|
||
... Passed -- 34.7 5 308 60588 12
|
||
=== RUN TestUnreliableAgree2C
|
||
... Passed -- 8.5 5 308 96421 251
|
||
=== RUN TestFigure8Unreliable2C
|
||
... Passed -- 30.8 5 2216 12947486 895
|
||
=== RUN TestReliableChurn2C
|
||
... Passed -- 16.3 5 580 424371 315
|
||
=== RUN TestUnreliableChurn2C
|
||
... Passed -- 16.2 5 268 133185 64
|
||
PASS
|
||
```
|
||
|
||
### 8.4 2D: Snapshots
|
||
|
||
```
|
||
=== RUN TestSnapshotBasic2D
|
||
... Passed -- 7.7 3 136 48058 251
|
||
=== RUN TestSnapshotInstall2D
|
||
... Passed -- 66.2 3 1358 359800 321
|
||
=== RUN TestSnapshotInstallUnreliable2D
|
||
... Passed -- 88.0 3 1808 431965 341
|
||
=== RUN TestSnapshotInstallCrash2D
|
||
... Passed -- 37.9 3 658 183093 355
|
||
=== RUN TestSnapshotInstallUnCrash2D
|
||
... Passed -- 57.5 3 782 207519 388
|
||
PASS
|
||
```
|
||
|
||
---
|
||
|
||
## 9. 总结与收获
|
||
|
||
### 9.1 实现要点
|
||
|
||
1. **任期是核心**：所有操作都要正确处理任期变化
|
||
2. **随机化超时**：避免选举冲突
|
||
3. **快速回退**：优化日志同步效率
|
||
4. **正确持久化**：崩溃恢复的关键
|
||
5. **快照索引**：注意偏移量计算
|
||
|
||
### 9.2 调试经验
|
||
|
||
- 使用 `-race` 检测数据竞争
|
||
- 仔细阅读论文 Figure 2
|
||
- 处理边界条件（空日志、快照边界等）
|
||
- 测试中的超时时间很宽松，实际实现要更高效
|
||
|
||
### 9.3 后续工作
|
||
|
||
Lab3 将基于此 Raft 实现构建容错的 KV 服务。
|
||
|
||
---
|
||
|
||
> 💡 Raft 是分布式系统的基石，理解它的实现对于后续的分布式存储、分布式协调等系统至关重要。 |