为什么需要一致性 hash?
在采用分片方式建立分布式缓存时,我们面临的第一个问题是如何决定存储数据的节点。最自然的方式是参考 hash 表的做法,假设集群中存在 n 个节点,我们用 node = hashCode(key) % n 来决定所属的节点。
普通 hash 算法解决了如何选择节点的问题,但在分布式系统中经常出现增加节点或某个节点宕机的情况。若节点数 n 发生变化, 大多数 key 根据 node = hashCode(key) % n 计算出的节点都会改变。这意味着若要在 n 变化后维持系统正常运转,需要将大多数数据在节点间进行重新分布。这个操作会消耗大量的时间和带宽等资源,这在生产环境下是不可接受的。
算法原理
一致性 hash 算法的目的是在节点数量 n 变化时, 使尽可能少的 key 需要进行节点间重新分布。一致性 hash 算法将数据 key 和服务器地址 addr 散列到 2^32 的空间中。
我们将 2^32 个整数首尾相连形成一个环,首先计算服务器地址 addr 的 hash 值放置在环上。然后计算 key 的 hash 值放置在环上,顺时针查找,将数据放在找到的的第一个节点上。
在增加或删除节点时只有该节点附近的数据需要重新分布,从而解决了上述问题。
如果服务器节点较少则比较容易出现数据分布不均匀的问题,一般来说环上的节点越多数据分布越均匀。我们不需要真的增加一台服务器,只需要将实际的服务器节点映射为几个虚拟节点放在环上即可。
参考:https://www.cnblogs.com/Finley/p/14038398.html
lib/consistenthash/consistenthash.go
type HashFunc func(data []byte) uint32
type NodeMap struct {
hashFunc HashFunc
nodeHashs []int
nodehashMap map[int]string
}
func NewNodeMap(fn HashFunc) *NodeMap {
m := &NodeMap{
hashFunc: fn,
nodehashMap: make(map[int]string),
}
if m.hashFunc == nil {
m.hashFunc = crc32.ChecksumIEEE
}
return m
}
func (m *NodeMap) IsEmpty() bool {
return len(m.nodeHashs) == 0
}
func (m *NodeMap) AddNode(keys ...string) {
for _, key := range keys {
if key == "" {
continue
}
hash := int(m.hashFunc([]byte(key)))
m.nodeHashs = append(m.nodeHashs, hash)
m.nodehashMap[hash] = key
}
sort.Ints(m.nodeHashs)
}
func (m *NodeMap) PickNode(key string) string {
if m.IsEmpty() {
return ""
}
hash := int(m.hashFunc([]byte(key)))
idx := sort.Search(len(m.nodeHashs), func(i int) bool {
return m.nodeHashs[i] >= hash
})
if idx == len(m.nodeHashs) {
idx = 0
}
return m.nodehashMap[m.nodeHashs[idx]]
}
HashFunc:hash函数定义,Go的hash函数就是这样定义的
NodeMap:存储所有节点和节点的hash
AddNode:添加节点到一致性哈希中
PickNode:选择节点。使用二分查找,如果hash比nodeHashs中最大的hash还要大,idx=0
database/standalone_database.go
type StandaloneDatabase struct {
dbSet []*DB
aofHandler *aof.AofHandler
}
func NewStandaloneDatabase() *StandaloneDatabase {
......
}
把database/database改名为database/standalone_database,再增加一个cluster_database用于对key的路由
resp/client/client.go
// Client is a pipeline mode redis client
type Client struct {
conn net.Conn
pendingReqs chan *request // wait to send
waitingReqs chan *request // waiting response
ticker *time.Ticker
addr string
working *sync.WaitGroup // its counter presents unfinished requests(pending and waiting)
}
// request is a message sends to redis server
type request struct {
id uint64
args [][]byte
reply resp.Reply
heartbeat bool
waiting *wait.Wait
err error
}
const (
chanSize = 256
maxWait = 3 * time.Second
)
// MakeClient creates a new client
func MakeClient(addr string) (*Client, error) {
conn, err := net.Dial("tcp", addr)
if err != nil {
return nil, err
}
return &Client{
addr: addr,
conn: conn,
pendingReqs: make(chan *request, chanSize),
waitingReqs: make(chan *request, chanSize),
working: &sync.WaitGroup{},
}, nil
}
// Start starts asynchronous goroutines
func (client *Client) Start() {
client.ticker = time.NewTicker(10 * time.Second)
go client.handleWrite()
go func() {
err := client.handleRead()
if err != nil {
logger.Error(err)
}
}()
go client.heartbeat()
}
// Close stops asynchronous goroutines and close connection
func (client *Client) Close() {
client.ticker.Stop()
// stop new request
close(client.pendingReqs)
// wait stop process
client.working.Wait()
// clean
_ = client.conn.Close()
close(client.waitingReqs)
}
func (client *Client) handleConnectionError(err error) error {
err1 := client.conn.Close()
if err1 != nil {
if opErr, ok := err1.(*net.OpError); ok {
if opErr.Err.Error() != "use of closed network connection" {
return err1
}
} else {
return err1
}
}
conn, err1 := net.Dial("tcp", client.addr)
if err1 != nil {
logger.Error(err1)
return err1
}
client.conn = conn
go func() {
_ = client.handleRead()
}()
return nil
}
func (client *Client) heartbeat() {
for range client.ticker.C {
client.doHeartbeat()
}
}
func (client *Client) handleWrite() {
for req := range client.pendingReqs {
client.doRequest(req)
}
}
// Send sends a request to redis server
func (client *Client) Send(args [][]byte) resp.Reply {
request := &request{
args: args,
heartbeat: false,
waiting: &wait.Wait{},
}
request.waiting.Add(1)
client.working.Add(1)
defer client.working.Done()
client.pendingReqs <- request
timeout := request.waiting.WaitWithTimeout(maxWait)
if timeout {
return reply.MakeErrReply("server time out")
}
if request.err != nil {
return reply.MakeErrReply("request failed")
}
return request.reply
}
func (client *Client) doHeartbeat() {
request := &request{
args: [][]byte{[]byte("PING")},
heartbeat: true,
waiting: &wait.Wait{},
}
request.waiting.Add(1)
client.working.Add(1)
defer client.working.Done()
client.pendingReqs <- request
request.waiting.WaitWithTimeout(maxWait)
}
func (client *Client) doRequest(req *request) {
if req == nil || len(req.args) == 0 {
return
}
re := reply.MakeMultiBulkReply(req.args)
bytes := re.ToBytes()
_, err := client.conn.Write(bytes)
i := 0
for err != nil && i < 3 {
err = client.handleConnectionError(err)
if err == nil {
_, err = client.conn.Write(bytes)
}
i++
}
if err == nil {
client.waitingReqs <- req
} else {
req.err = err
req.waiting.Done()
}
}
func (client *Client) finishRequest(reply resp.Reply) {
defer func() {
if err := recover(); err != nil {
debug.PrintStack()
logger.Error(err)
}
}()
request := <-client.waitingReqs
if request == nil {
return
}
request.reply = reply
if request.waiting != nil {
request.waiting.Done()
}
}
func (client *Client) handleRead() error {
ch := parser.ParseStream(client.conn)
for payload := range ch {
if payload.Err != nil {
client.finishRequest(reply.MakeErrReply(payload.Err.Error()))
continue
}
client.finishRequest(payload.Data)
}
return nil
}
client:Redis客户端,具体看:https://www.cnblogs.com/Finley/p/14028402.html
go.mod
require github.com/jolestar/go-commons-pool/v2 v2.1.2
key的转发需要当前节点存储其他节点的连接,互相作为客户端,使用连接池将其他连接池化
cluster/client_pool.go
type connectionFactory struct {
Peer string // 连接地址
}
func (f *connectionFactory) MakeObject(ctx context.Context) (*pool.PooledObject, error) {
c, err := client.MakeClient(f.Peer)
if err != nil {
return nil, err
}
c.Start()
return pool.NewPooledObject(c), nil
}
func (f *connectionFactory) DestroyObject(ctx context.Context, object *pool.PooledObject) error {
c, ok := object.Object.(*client.Client)
if !ok {
return errors.New("type mismatch")
}
c.Close()
return nil
}
func (f *connectionFactory) ValidateObject(ctx context.Context, object *pool.PooledObject) bool {
// do validate
return true
}
func (f *connectionFactory) ActivateObject(ctx context.Context, object *pool.PooledObject) error {
// do activate
return nil
}
func (f *connectionFactory) PassivateObject(ctx context.Context, object *pool.PooledObject) error {
// do passivate
return nil
}
client_pool:使用连接池的NewObjectPoolWithDefaultConfig创建连接,需要实现PooledObjectFactory接口
redis.conf
self 127.0.0.1:6379
peers 127.0.0.1:6380
配置中写自己和其他节点的地址
cluster/cluster_database.go
type clusterDatabase struct {
self string
nodes []string
peerPicker *consistenthash.NodeMap
peerConnection map[string]*pool.ObjectPool
db databaseface.Database
}
func MakeClusterDatabase() *clusterDatabase {
cluster := &clusterDatabase{
self: config.Properties.Self,
db: database.NewStandaloneDatabase(),
peerPicker: consistenthash.NewNodeMap(nil),
peerConnection: make(map[string]*pool.ObjectPool),
}
nodes := make([]string, 0, len(config.Properties.Peers)+1)
for _, peer := range config.Properties.Peers {
nodes = append(nodes, peer)
}
nodes = append(nodes, config.Properties.Self)
cluster.peerPicker.AddNode(nodes...)
ctx := context.Background()
for _, peer := range config.Properties.Peers {
cluster.peerConnection[peer] = pool.NewObjectPoolWithDefaultConfig(ctx, &connectionFactory{
Peer: peer,
})
}
cluster.nodes = nodes
return cluster
}
func (cluster *clusterDatabase) Close() {
cluster.db.Close()
}
func (cluster *ClusterDatabase) AfterClientClose(c resp.Connection) {
cluster.db.AfterClientClose(c)
}
type CmdFunc func(cluster *clusterDatabase, c resp.Connection, cmdAndArgs [][]byte) resp.Reply
cluster_database用于对key的路由
clusterDatabase:
nodes:所有节点
peerPicker :节点的添加和选择
peerConnection:Map
db:单机database
CmdFunc:表示Redis的指令类型
cluster/com.go
func (cluster *clusterDatabase) getPeerClient(peer string) (*client.Client, error) {
factory, ok := cluster.peerConnection[peer]
if !ok {
return nil, errors.New("connection factory not found")
}
raw, err := factory.BorrowObject(context.Background())
if err != nil {
return nil, err
}
conn, ok := raw.(*client.Client)
if !ok {
return nil, errors.New("connection factory make wrong type")
}
return conn, nil
}
func (cluster *clusterDatabase) returnPeerClient(peer string, peerClient *client.Client) error {
connectionFactory, ok := cluster.peerConnection[peer]
if !ok {
return errors.New("connection factory not found")
}
return connectionFactory.ReturnObject(context.Background(), peerClient)
}
func (cluster *clusterDatabase) relay(peer string, c resp.Connection, args [][]byte) resp.Reply {
if peer == cluster.self {
return cluster.db.Exec(c, args)
}
peerClient, err := cluster.getPeerClient(peer)
if err != nil {
return reply.MakeErrReply(err.Error())
}
defer func() {
_ = cluster.returnPeerClient(peer, peerClient)
}()
peerClient.Send(utils.ToCmdLine("SELECT", strconv.Itoa(c.GetDBIndex())))
return peerClient.Send(args)
}
func (cluster *clusterDatabase) broadcast(c resp.Connection, args [][]byte) map[string]resp.Reply {
result := make(map[string]resp.Reply)
for _, node := range cluster.nodes {
relay := cluster.relay(node, c, args)
result[node] = relay
}
return result
}
communication:与其他节点通信。执行模式有本地(自己执行),转发(别人执行),群发(所有节点执行)
getPeerClient :从连接池拿一个连接
returnPeerClient :归还连接
relay :转发指令给其他客户端,发送指令之前需要先发一下选择的db
broadcast :指令广播给所有节点
cluster/router.go
func makeRouter() map[string]CmdFunc {
routerMap := make(map[string]CmdFunc)
routerMap["ping"] = ping
routerMap["del"] = Del
routerMap["exists"] = defaultFunc
routerMap["type"] = defaultFunc
routerMap["rename"] = Rename
routerMap["renamenx"] = Rename
routerMap["set"] = defaultFunc
routerMap["setnx"] = defaultFunc
routerMap["get"] = defaultFunc
routerMap["getset"] = defaultFunc
routerMap["flushdb"] = FlushDB
routerMap["select"] = execSelect
return routerMap
}
func defaultFunc(cluster *clusterDatabase, c resp.Connection, args [][]byte) resp.Reply {
key := string(args[1])
peer := cluster.peerPicker.PickNode(key)
return cluster.relay(peer, c, args)
}
defaultFunc:转发指令的默认实现
cluster/ping.go
func ping(cluster *clusterDatabase, c resp.Connection, cmdAndArgs [][]byte) resp.Reply {
return cluster.db.Exec(c, cmdAndArgs)
}
cluster/rename.go
func Rename(cluster *clusterDatabase, c resp.Connection, args [][]byte) resp.Reply {
if len(args) != 3 {
return reply.MakeErrReply("ERR wrong number of arguments for 'rename' command")
}
src := string(args[1])
dest := string(args[2])
srcPeer := cluster.peerPicker.PickNode(src)
destPeer := cluster.peerPicker.PickNode(dest)
if srcPeer != destPeer {
return reply.MakeErrReply("ERR rename must within one slot in cluster mode")
}
return cluster.relay(srcPeer, c, args)
}
Rename:修改key的name,两个key的hash必须在同一个节点中
cluster/keys.go
func FlushDB(cluster *clusterDatabase, c resp.Connection, args [][]byte) resp.Reply {
replies := cluster.broadcast(c, args)
var errReply reply.ErrorReply
for _, v := range replies {
if reply.IsErrorReply(v) {
errReply = v.(reply.ErrorReply)
break
}
}
if errReply == nil {
return &reply.OkReply{}
}
return reply.MakeErrReply("error occurs: " + errReply.Error())
}
cluster/del.go
func Del(cluster *clusterDatabase, c resp.Connection, args [][]byte) resp.Reply {
replies := cluster.broadcast(c, args)
var errReply reply.ErrorReply
var deleted int64 = 0
for _, v := range replies {
if reply.IsErrorReply(v) {
errReply = v.(reply.ErrorReply)
break
}
intReply, ok := v.(*reply.IntReply)
if !ok {
errReply = reply.MakeErrReply("error")
}
deleted += intReply.Code
}
if errReply == nil {
return reply.MakeIntReply(deleted)
}
return reply.MakeErrReply("error occurs: " + errReply.Error())
}
cluster/select.go
func execSelect(cluster *clusterDatabase, c resp.Connection, cmdAndArgs [][]byte) resp.Reply {
return cluster.db.Exec(c, cmdAndArgs)
}
cluster/cluster_database.go
var router = makeRouter()
func (cluster *clusterDatabase) Exec(c resp.Connection, cmdLine [][]byte) (result resp.Reply) {
defer func() {
if err := recover(); err != nil {
logger.Warn(fmt.Sprintf("error occurs: %v\n%s", err, string(debug.Stack())))
result = &reply.UnknownErrReply{}
}
}()
cmdName := strings.ToLower(string(cmdLine[0]))
cmdFunc, ok := router[cmdName]
if !ok {
return reply.MakeErrReply("ERR unknown command '" + cmdName + "', or not supported in cluster mode")
}
result = cmdFunc(cluster, c, cmdLine)
return
}
resp/handler/handler.go
func MakeHandler() *RespHandler {
var db databaseface.Database
if config.Properties.Self != "" && len(config.Properties.Peers) > 0 {
db = cluster.MakeClusterDatabase()
} else {
db = database.NewStandaloneDatabase()
}
return &RespHandler{
db: db,
}
}
MakeHandler:判断是单机还是集群
先go build,打开项目文件夹找到exe文件,把exe文件和redis.conf放到一个文件夹里,redis.conf改成如下,然后启动exe文件。再回到GoLand启动第二个节点6379。
bind 0.0.0.0
port 6380
appendonly yes
appendfilename appendonly.aof
self 127.0.0.1:6380
peers 127.0.0.1:6379