func StartSocket() {
if !g.Config().Socket.Enabled {
return
}
addr := g.Config().Socket.Listen
tcpAddr, err := net.ResolveTCPAddr("tcp", addr)
if err != nil {
log.Fatalf("net.ResolveTCPAddr fail: %s", err)
}
listener, err := net.ListenTCP("tcp", tcpAddr)
if err != nil {
log.Fatalf("listen %s fail: %s", addr, err)
} else {
log.Println("socket listening", addr)
}
defer listener.Close()
for {
conn, err := listener.Accept()
if err != nil {
log.Println("listener.Accept occur error:", err)
continue
}
go socketTelnetHandle(conn)
}
}
func newInfluxdbConnPool(name string, address string, connTimeout time.Duration, maxConns int, maxIdle int) *ConnPool {
pool := NewConnPool(name, address, maxConns, maxIdle)
pool.Username = g.Config().Influxdb.Username
pool.Password = g.Config().Influxdb.Password
pool.Database = g.Config().Influxdb.Database
pool.Precision = "s"
pool.New = func(connName string) (NConn, error) {
nconn := InfluxdbClient{
name: connName,
Address: pool.Address,
Username: pool.Username,
Password: pool.Password,
Database: pool.Database,
Precision: pool.Precision,
Timeout: connTimeout,
}
err := nconn.Connect()
if err != nil {
return nil, err
}
return nconn, nil
}
return pool
}
func StartRpc() {
if !g.Config().Rpc.Enabled {
return
}
addr := g.Config().Rpc.Listen
tcpAddr, err := net.ResolveTCPAddr("tcp", addr)
if err != nil {
log.Fatalf("net.ResolveTCPAddr fail: %s", err)
}
listener, err := net.ListenTCP("tcp", tcpAddr)
if err != nil {
log.Fatalf("listen %s fail: %s", addr, err)
} else {
log.Println("rpc listening", addr)
}
server := rpc.NewServer()
server.Register(new(Transfer))
for {
conn, err := listener.Accept()
if err != nil {
log.Println("listener.Accept occur error:", err)
continue
}
// go rpc.ServeConn(conn)
go server.ServeCodec(jsonrpc.NewServerCodec(conn))
}
}
// 将数据 打入 某个Graph的发送缓存队列, 具体是哪一个Graph 由一致性哈希 决定
// 如果正在数据迁移, 数据除了打到原有配置上 还要向新的配置上打一份(新老重叠时要去重,防止将一条数据向一台Graph上打两次)
func Push2GraphSendQueue(items []*cmodel.MetaData, migrating bool) {
cfg := g.Config().Graph
for _, item := range items {
graphItem, err := convert2GraphItem(item)
if err != nil {
log.Println("E:", err)
continue
}
pk := item.PK()
// statistics. 为了效率,放到了这里,因此只有graph是enbale时才能trace
proc.RecvDataTrace.Trace(pk, item)
proc.RecvDataFilter.Filter(pk, item.Value, item)
node, err := GraphNodeRing.GetNode(pk)
if err != nil {
log.Println("E:", err)
continue
}
cnode := cfg.Cluster2[node]
errCnt := 0
for _, addr := range cnode.Addrs {
Q := GraphQueues[node+addr]
if !Q.PushFront(graphItem) {
errCnt += 1
}
}
// statistics
if errCnt > 0 {
proc.SendToGraphDropCnt.Incr()
}
if migrating {
migratingNode, err := GraphMigratingNodeRing.GetNode(pk)
if err != nil {
log.Println("E:", err)
continue
}
if node != migratingNode { // 数据迁移时,进行新老去重
cnodem := cfg.ClusterMigrating2[migratingNode]
errCnt := 0
for _, addr := range cnodem.Addrs {
MQ := GraphMigratingQueues[migratingNode+addr]
if !MQ.PushFront(graphItem) {
errCnt += 1
}
}
// statistics
if errCnt > 0 {
proc.SendToGraphMigratingDropCnt.Incr()
}
}
}
}
}
func configCommonRoutes() {
http.HandleFunc("/health", func(w http.ResponseWriter, r *http.Request) {
w.Write([]byte("ok\n"))
})
http.HandleFunc("/version", func(w http.ResponseWriter, r *http.Request) {
w.Write([]byte(fmt.Sprintf("%s\n", g.VERSION)))
})
http.HandleFunc("/workdir", func(w http.ResponseWriter, r *http.Request) {
w.Write([]byte(fmt.Sprintf("%s\n", file.SelfDir())))
})
http.HandleFunc("/config", func(w http.ResponseWriter, r *http.Request) {
RenderDataJson(w, g.Config())
})
http.HandleFunc("/config/reload", func(w http.ResponseWriter, r *http.Request) {
if strings.HasPrefix(r.RemoteAddr, "127.0.0.1") {
g.ParseConfig(g.ConfigFile)
RenderDataJson(w, "ok")
} else {
RenderDataJson(w, "no privilege")
}
})
}
func initSendQueues() {
cfg := g.Config()
for node, _ := range cfg.Judge.Cluster {
Q := nlist.NewSafeListLimited(DefaultSendQueueMaxSize)
JudgeQueues[node] = Q
}
if cfg.Influxdb.Enabled {
for tnode, _ := range cfg.Influxdb.Cluster {
Q := nlist.NewSafeListLimited(DefaultSendQueueMaxSize)
InfluxdbQueues[tnode] = Q
}
}
for node, nitem := range cfg.Graph.Cluster2 {
for _, addr := range nitem.Addrs {
Q := nlist.NewSafeListLimited(DefaultSendQueueMaxSize)
GraphQueues[node+addr] = Q
}
}
if cfg.Graph.Migrating && cfg.Graph.ClusterMigrating != nil {
for node, cnode := range cfg.Graph.ClusterMigrating2 {
for _, addr := range cnode.Addrs {
Q := nlist.NewSafeListLimited(DefaultSendQueueMaxSize)
GraphMigratingQueues[node+addr] = Q
}
}
}
}
// Judge定时任务, 将 Judge发送缓存中的数据 通过rpc连接池 发送到Judge
func forward2JudgeTask(Q *list.SafeListLimited, node string, concurrent int) {
batch := g.Config().Judge.Batch // 一次发送,最多batch条数据
addr := g.Config().Judge.Cluster[node]
sema := nsema.NewSemaphore(concurrent)
for {
items := Q.PopBackBy(batch)
count := len(items)
if count == 0 {
time.Sleep(DefaultSendTaskSleepInterval)
continue
}
judgeItems := make([]*cmodel.JudgeItem, count)
for i := 0; i < count; i++ {
judgeItems[i] = items[i].(*cmodel.JudgeItem)
}
// 同步Call + 有限并发 进行发送
sema.Acquire()
go func(addr string, judgeItems []*cmodel.JudgeItem, count int) {
defer sema.Release()
resp := &cmodel.SimpleRpcResponse{}
var err error
sendOk := false
for i := 0; i < 3; i++ { //最多重试3次
err = JudgeConnPools.Call(addr, "Judge.Send", judgeItems, resp)
if err == nil {
sendOk = true
break
}
time.Sleep(time.Millisecond * 10)
}
// statistics
if !sendOk {
log.Printf("send judge %s:%s fail: %v", node, addr, err)
proc.SendToJudgeFailCnt.IncrBy(int64(count))
} else {
proc.SendToJudgeCnt.IncrBy(int64(count))
}
}(addr, judgeItems, count)
}
}
func initNodeRings() {
cfg := g.Config()
JudgeNodeRing = newConsistentHashNodesRing(cfg.Judge.Replicas, KeysOfMap(cfg.Judge.Cluster))
GraphNodeRing = newConsistentHashNodesRing(cfg.Graph.Replicas, KeysOfMap(cfg.Graph.Cluster))
if cfg.Graph.Migrating && cfg.Graph.ClusterMigrating != nil {
GraphMigratingNodeRing = newConsistentHashNodesRing(cfg.Graph.Replicas, KeysOfMap(cfg.Graph.ClusterMigrating))
}
}
func startHttpServer() {
if !g.Config().Http.Enabled {
return
}
addr := g.Config().Http.Listen
if addr == "" {
return
}
configCommonRoutes()
configProcHttpRoutes()
configDebugHttpRoutes()
configApiHttpRoutes()
s := &http.Server{
Addr: addr,
MaxHeaderBytes: 1 << 30,
}
log.Println("http.startHttpServer ok, listening", addr)
log.Fatalln(s.ListenAndServe())
}
// TODO 添加对发送任务的控制,比如stop等
func startSendTasks() {
cfg := g.Config()
// init semaphore
influxdbConcurrent := cfg.Influxdb.MaxIdle
judgeConcurrent := cfg.Judge.MaxIdle
graphConcurrent := cfg.Graph.MaxIdle
if influxdbConcurrent < 1 {
influxdbConcurrent = 1
}
if judgeConcurrent < 1 {
judgeConcurrent = 1
}
if graphConcurrent < 1 {
graphConcurrent = 1
}
// init send go-routines
if cfg.Influxdb.Enabled {
for node, _ := range cfg.Influxdb.Cluster {
queue := InfluxdbQueues[node]
go forward2InfluxdbTask(queue, node, influxdbConcurrent)
}
}
for node, _ := range cfg.Judge.Cluster {
queue := JudgeQueues[node]
go forward2JudgeTask(queue, node, judgeConcurrent)
}
for node, nitem := range cfg.Graph.Cluster2 {
for _, addr := range nitem.Addrs {
queue := GraphQueues[node+addr]
go forward2GraphTask(queue, node, addr, graphConcurrent)
}
}
if cfg.Graph.Migrating {
for node, cnodem := range cfg.Graph.ClusterMigrating2 {
for _, addr := range cnodem.Addrs {
queue := GraphMigratingQueues[node+addr]
go forward2GraphMigratingTask(queue, node, addr, graphConcurrent)
}
}
}
}
// Graph定时任务, 进行数据迁移时的 数据冗余发送
func forward2GraphMigratingTask(Q *list.SafeListLimited, node string, addr string, concurrent int) {
batch := g.Config().Graph.Batch // 一次发送,最多batch条数据
sema := nsema.NewSemaphore(concurrent)
for {
items := Q.PopBackBy(batch)
count := len(items)
if count == 0 {
time.Sleep(DefaultSendTaskSleepInterval)
continue
}
graphItems := make([]*cmodel.GraphItem, count)
for i := 0; i < count; i++ {
graphItems[i] = items[i].(*cmodel.GraphItem)
}
sema.Acquire()
go func(addr string, graphItems []*cmodel.GraphItem, count int) {
defer sema.Release()
resp := &cmodel.SimpleRpcResponse{}
var err error
sendOk := false
for i := 0; i < 3; i++ { //最多重试3次
err = GraphMigratingConnPools.Call(addr, "Graph.Send", graphItems, resp)
if err == nil {
sendOk = true
break
}
time.Sleep(time.Millisecond * 10) //发送失败了,睡10ms
}
// statistics
if !sendOk {
log.Printf("send to graph migrating %s:%s fail: %v", node, addr, err)
proc.SendToGraphMigratingFailCnt.IncrBy(int64(count))
} else {
proc.SendToGraphMigratingCnt.IncrBy(int64(count))
}
}(addr, graphItems, count)
}
}
func initConnPools() {
cfg := g.Config()
if cfg.Influxdb.Enabled {
influxdbInstances := nset.NewStringSet()
for _, instance := range cfg.Influxdb.Cluster {
influxdbInstances.Add(instance)
}
InfluxdbConnPools = cpool.CreateInfluxdbCliPools(cfg.Influxdb.MaxConns, cfg.Influxdb.MaxIdle,
cfg.Influxdb.ConnTimeout, cfg.Influxdb.CallTimeout, influxdbInstances.ToSlice())
}
judgeInstances := nset.NewStringSet()
for _, instance := range cfg.Judge.Cluster {
judgeInstances.Add(instance)
}
JudgeConnPools = cpool.CreateSafeRpcConnPools(cfg.Judge.MaxConns, cfg.Judge.MaxIdle,
cfg.Judge.ConnTimeout, cfg.Judge.CallTimeout, judgeInstances.ToSlice())
// graph
graphInstances := nset.NewSafeSet()
for _, nitem := range cfg.Graph.Cluster2 {
for _, addr := range nitem.Addrs {
graphInstances.Add(addr)
}
}
GraphConnPools = cpool.CreateSafeRpcConnPools(cfg.Graph.MaxConns, cfg.Graph.MaxIdle,
cfg.Graph.ConnTimeout, cfg.Graph.CallTimeout, graphInstances.ToSlice())
// graph migrating
if cfg.Graph.Migrating && cfg.Graph.ClusterMigrating != nil {
graphMigratingInstances := nset.NewSafeSet()
for _, cnode := range cfg.Graph.ClusterMigrating2 {
for _, addr := range cnode.Addrs {
graphMigratingInstances.Add(addr)
}
}
GraphMigratingConnPools = cpool.CreateSafeRpcConnPools(cfg.Graph.MaxConns, cfg.Graph.MaxIdle,
cfg.Graph.ConnTimeout, cfg.Graph.CallTimeout, graphMigratingInstances.ToSlice())
}
}
// Tsdb定时任务, 将 Tsdb发送缓存中的数据 通过api连接池 发送到Tsdb
// 单个Cluster配置多个influxdb地址，修改为并发发送。
func forward2InfluxdbTask(Q *list.SafeListLimited, node string, concurrent int) {
cfg := g.Config()
batch := cfg.Influxdb.Batch // 一次发送,最多batch条数据
sema := nsema.NewSemaphore(concurrent * len(cfg.Influxdb.Cluster2[node].Addrs))
retry := cfg.Influxdb.MaxRetry
for {
items := Q.PopBackBy(batch)
count := len(items)
if count == 0 {
time.Sleep(DefaultSendTaskSleepInterval)
continue
}
pts := make([]*client.Point, count)
for i := 0; i < count; i++ {
pts[i] = items[i].(*client.Point)
}
for _, addr := range cfg.Influxdb.Cluster2[node].Addrs {
sema.Acquire()
go coreSend2Influxdb(addr, sema, retry, pts)
}
}
}
// 将数据 打入所有的Tsdb的发送缓存队列, 相互备份
func Push2TsdbSendQueue(items []*cmodel.MetaData) {
removeMetrics := g.Config().Influxdb.RemoveMetrics
//log.Printf("Push2TsdbSendQueue")
for _, item := range items {
b, ok := removeMetrics[item.Metric]
//log.Printf ("select:%V,%V,%V", b, ok,item )
if b && ok {
continue
}
influxPoint := Convert2InfluxPoint(item)
errCnt := 0
for _, Q := range InfluxdbQueues {
if !Q.PushFront(influxPoint) {
errCnt += 1
}
}
// statistics
if errCnt > 0 {
proc.SendToInfluxdbDropCnt.Incr()
}
}
}
func socketTelnetHandle(conn net.Conn) {
defer conn.Close()
items := []*cmodel.MetaData{}
buf := bufio.NewReader(conn)
cfg := g.Config()
timeout := time.Duration(cfg.Socket.Timeout) * time.Second
for {
conn.SetReadDeadline(time.Now().Add(timeout))
line, err := buf.ReadString('\n')
if err != nil {
break
}
line = strings.Trim(line, "\n")
if line == "quit" {
break
}
if line == "" {
continue
}
t := strings.Fields(line)
if len(t) < 2 {
continue
}
cmd := t[0]
if cmd != "update" {
continue
}
item, err := convertLine2MetaData(t[1:])
if err != nil {
continue
}
items = append(items, item)
}
// statistics
proc.SocketRecvCnt.IncrBy(int64(len(items)))
proc.RecvCnt.IncrBy(int64(len(items)))
if cfg.Graph.Enabled {
sender.Push2GraphSendQueue(items, cfg.Graph.Migrating)
}
if cfg.Judge.Enabled {
sender.Push2JudgeSendQueue(items)
}
if cfg.Influxdb.Enabled {
sender.Push2TsdbSendQueue(items)
}
return
}
// process new metric values
func RecvMetricValues(args []*cmodel.MetricValue, reply *cmodel.TransferResponse, from string) error {
start := time.Now()
reply.Invalid = 0
items := []*cmodel.MetaData{}
for _, v := range args {
if v == nil {
reply.Invalid += 1
continue
}
// 历史遗留问题.
// 老版本agent上报的metric=kernel.hostname的数据,其取值为string类型,现在已经不支持了;所以,这里硬编码过滤掉
if v.Metric == "kernel.hostname" {
reply.Invalid += 1
continue
}
if v.Metric == "" || v.Endpoint == "" {
reply.Invalid += 1
continue
}
if v.Type != g.COUNTER && v.Type != g.GAUGE && v.Type != g.DERIVE {
reply.Invalid += 1
continue
}
if v.Value == "" {
reply.Invalid += 1
continue
}
if v.Step <= 0 {
reply.Invalid += 1
continue
}
if len(v.Metric)+len(v.Tags) > 510 {
reply.Invalid += 1
continue
}
// TODO 呵呵,这里需要再优雅一点
now := start.Unix()
if v.Timestamp <= 0 || v.Timestamp > now*2 {
v.Timestamp = now
}
fv := &cmodel.MetaData{
Metric: v.Metric,
Endpoint: v.Endpoint,
Timestamp: v.Timestamp,
Step: v.Step,
CounterType: v.Type,
Tags: cutils.DictedTagstring(v.Tags), //TODO tags键值对的个数,要做一下限制
}
valid := true
var vv float64
var err error
switch cv := v.Value.(type) {
case string:
vv, err = strconv.ParseFloat(cv, 64)
if err != nil {
valid = false
}
case float64:
vv = cv
case int64:
vv = float64(cv)
default:
valid = false
}
if !valid {
reply.Invalid += 1
continue
}
fv.Value = vv
items = append(items, fv)
}
// statistics
cnt := int64(len(items))
proc.RecvCnt.IncrBy(cnt)
if from == "rpc" {
proc.RpcRecvCnt.IncrBy(cnt)
} else if from == "http" {
proc.HttpRecvCnt.IncrBy(cnt)
}
cfg := g.Config()
if cfg.Graph.Enabled {
sender.Push2GraphSendQueue(items, cfg.Graph.Migrating)
}
if cfg.Influxdb.Enabled {
sender.Push2TsdbSendQueue(items)
}
if cfg.Judge.Enabled {
sender.Push2JudgeSendQueue(items)
}
reply.Message = "ok"
reply.Total = len(args)
reply.Latency = (time.Now().UnixNano() - start.UnixNano()) / 1000000
return nil
}
