- v0.31: AlertContact 重构(channel_type下拉+webhook_url替代三分立字段) - v0.32: Agent ProcessTop 增强(cmdline/user/cpu_pct/mem_pct) - v0.33: 端口监控 Tab(监听端口自动发现+down联动告警) - v0.34: 健康探针(Agent本地执行HTTP/TCP/Exec→结果随心跳上报) - Agent新增: ContextSwitch/FD/HealthProbe采集器 - 后端新增: health_probe handler + Agent model扩展 - HostMonitor新增Tab: 健康探针/端口监控
346 lines
9.1 KiB
Go
346 lines
9.1 KiB
Go
// Package main 是 AIOps Agent 的入口。
|
||
// Agent 定时采集 CPU、内存、磁盘、端口、进程 TOP、Swap、FD、上下文切换等指标,
|
||
// 并通过 HTTP 上报至 AIOps 后端。
|
||
// v2.1: 共享 HTTP client + 独立探针拉取 + 配置结构体
|
||
//
|
||
// 优化要点(1000 台服务器规模):
|
||
// - 启动随机 Jitter 避免惊群效应
|
||
// - 连接池复用 HTTP Keep-Alive
|
||
// - 进程 TOP 限制为 3(可配置)
|
||
// - Agent 自检心跳(序列号丢包检测)
|
||
// - 端口变化率采集(增量检测,减少 payload)
|
||
//
|
||
// 署名 Logan.Li
|
||
package main
|
||
|
||
import (
|
||
"context"
|
||
"flag"
|
||
"log"
|
||
"math/rand"
|
||
"net"
|
||
"os"
|
||
"os/signal"
|
||
"runtime"
|
||
"syscall"
|
||
"time"
|
||
|
||
agentclient "ezops/agent/client"
|
||
"ezops/agent/collector"
|
||
"ezops/agent/model"
|
||
"ezops/agent/reporter"
|
||
"ezops/agent/task"
|
||
)
|
||
|
||
var (
|
||
version = "v2.1.0"
|
||
buildTime = "unknown"
|
||
)
|
||
|
||
// Config Agent 运行配置(v2.1 从 flag 集中解析)
|
||
type Config struct {
|
||
Server string
|
||
Token string
|
||
Hostname string
|
||
Interval int
|
||
TopN int
|
||
}
|
||
|
||
func main() {
|
||
cfg := Config{
|
||
Server: *flag.String("server", "http://localhost:8080", "AIOps 后端上报地址"),
|
||
Token: *flag.String("token", "", "Bearer 鉴权令牌"),
|
||
Hostname: *flag.String("hostname", "", "上报主机名,默认使用系统主机名"),
|
||
Interval: *flag.Int("interval", 10, "采集与上报间隔,单位:秒"),
|
||
TopN: *flag.Int("top-n", 3, "进程 TOP N 数量"),
|
||
}
|
||
flag.Parse()
|
||
|
||
logger := log.New(os.Stdout, "[aiops-agent] ", log.LstdFlags|log.Lmicroseconds)
|
||
|
||
if cfg.Hostname == "" {
|
||
if h, err := os.Hostname(); err == nil {
|
||
cfg.Hostname = h
|
||
} else {
|
||
cfg.Hostname = "unknown"
|
||
}
|
||
}
|
||
|
||
if cfg.Server == "" {
|
||
logger.Fatal("参数 -server 不能为空")
|
||
}
|
||
if cfg.Interval < 5 {
|
||
logger.Fatalf("参数 -interval 必须 >= 5,当前为 %d(1000 台规模推荐 30~60s)", cfg.Interval)
|
||
}
|
||
|
||
logger.Printf("AIOps Agent v%s (built %s) go=%s", version, buildTime, runtime.Version())
|
||
logger.Printf("配置: server=%s interval=%ds top-n=%d hostname=%q", cfg.Server, cfg.Interval, cfg.TopN, cfg.Hostname)
|
||
|
||
// ═══ 共享 HTTP 客户端(Reporter + TaskClient 共用)═══
|
||
httpCli := agentclient.New(cfg.Server, cfg.Token, logger)
|
||
|
||
// ═══ 初始化采集器 ═══
|
||
cpuCollector := collector.NewCPUCollector()
|
||
memCollector := collector.NewMemCollector()
|
||
diskCollector := collector.NewDiskCollector()
|
||
portCollector := collector.NewPortCollector()
|
||
netCollector := collector.NewNetworkCollector()
|
||
diskIOCollector := collector.NewDiskIOCollector()
|
||
procTopCollector := collector.NewProcessTopCollector()
|
||
swapCollector := collector.NewSwapCollector()
|
||
healthProbeCollector := collector.NewHealthProbeCollector()
|
||
fdCollector := collector.NewFDCollector()
|
||
csCollector := collector.NewContextSwitchCollector()
|
||
rep := reporter.New(httpCli, logger)
|
||
|
||
// 系统静态信息(仅启动时一次)
|
||
sysInfo := collector.CollectSystem()
|
||
logger.Printf("系统信息: OS=%q Kernel=%q", sysInfo.OSVersion, sysInfo.Kernel)
|
||
|
||
// 探测本机 IP
|
||
localIP := detectLocalIP()
|
||
logger.Printf("探测本机 IP: %s", localIP)
|
||
|
||
sysReportCount := 0
|
||
|
||
// 任务执行器 + 任务客户端
|
||
taskExecutor := task.New(logger)
|
||
taskClient := task.NewClient(httpCli, cfg.Hostname, logger)
|
||
|
||
// ═══ 启动随机 Jitter ═══
|
||
startupJitter := time.Duration(rand.Intn(cfg.Interval)) * time.Second
|
||
logger.Printf("启动 Jitter: %v(防止惊群效应)", startupJitter)
|
||
|
||
ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
|
||
defer stop()
|
||
|
||
var reportSeq uint64
|
||
var probeConfigs []model.ProbeConfig
|
||
|
||
// ═══ collectAndReport 采集 + 上报 ═══
|
||
collectAndReport := func() {
|
||
report := &model.Report{
|
||
Hostname: cfg.Hostname,
|
||
IP: localIP,
|
||
Timestamp: time.Now().UTC().Format(time.RFC3339),
|
||
AgentVersion: version,
|
||
}
|
||
|
||
if sysReportCount < 3 && (sysInfo.OSVersion != "" || sysInfo.Kernel != "") {
|
||
report.System = &model.SystemStat{
|
||
OSVersion: sysInfo.OSVersion,
|
||
Kernel: sysInfo.Kernel,
|
||
Uptime: sysInfo.UptimeSeconds,
|
||
}
|
||
sysReportCount++
|
||
}
|
||
|
||
// 系统指标采集
|
||
if stat, err := cpuCollector.Collect(); err != nil {
|
||
logger.Printf("CPU 采集异常: %v", err)
|
||
} else {
|
||
report.CPU = stat
|
||
}
|
||
if stat, err := memCollector.Collect(); err != nil {
|
||
logger.Printf("内存采集异常: %v", err)
|
||
} else {
|
||
report.Memory = stat
|
||
}
|
||
if swap, err := swapCollector.Collect(); err != nil {
|
||
logger.Printf("Swap 采集异常: %v", err)
|
||
} else if swap != nil {
|
||
report.Swap = swap
|
||
}
|
||
if disks, err := diskCollector.Collect(); err != nil {
|
||
logger.Printf("磁盘采集异常: %v", err)
|
||
} else {
|
||
report.Disks = disks
|
||
}
|
||
if ports, err := portCollector.Collect(); err != nil {
|
||
logger.Printf("端口采集异常: %v", err)
|
||
} else {
|
||
report.Ports = ports
|
||
}
|
||
if netStat, err := netCollector.Collect(); err != nil {
|
||
logger.Printf("网络采集异常: %v", err)
|
||
} else {
|
||
report.Network = netStat
|
||
}
|
||
if dioStat, err := diskIOCollector.Collect(); err != nil {
|
||
logger.Printf("磁盘IO采集异常: %v", err)
|
||
} else {
|
||
report.DiskIO = dioStat
|
||
}
|
||
if procs, err := procTopCollector.Collect(cfg.TopN); err != nil {
|
||
logger.Printf("进程TOP采集异常: %v", err)
|
||
} else {
|
||
report.ProcessTop = procs
|
||
}
|
||
if fdStat, err := fdCollector.Collect(); err != nil {
|
||
logger.Printf("FD采集异常: %v", err)
|
||
} else {
|
||
report.FD = fdStat
|
||
}
|
||
if csStat, err := csCollector.Collect(); err != nil {
|
||
// 静默跳过
|
||
} else {
|
||
report.ContextSwitches = csStat
|
||
}
|
||
|
||
// v2.1 健康探针(使用缓存配置)
|
||
if len(probeConfigs) > 0 {
|
||
report.ProbeResults = healthProbeCollector.Execute(ctx, probeConfigs)
|
||
}
|
||
|
||
// Agent 自检心跳
|
||
reportSeq++
|
||
var memStats runtime.MemStats
|
||
runtime.ReadMemStats(&memStats)
|
||
report.Heartbeat = model.AgentHeartbeat{
|
||
Version: version,
|
||
UptimeSec: sysInfo.UptimeSeconds,
|
||
SelfCPU: 0,
|
||
SelfRSSMB: float64(memStats.Alloc) / 1024.0 / 1024.0,
|
||
ReportSeq: reportSeq,
|
||
}
|
||
|
||
// 上报指标(纯上报,不解析响应)
|
||
if err := rep.Send(ctx, report); err != nil {
|
||
logger.Printf("上报失败: %v", err)
|
||
return
|
||
}
|
||
|
||
probeCount := len(report.ProbeResults)
|
||
logger.Printf("上报成功: cpu=%.1f%% mem=%.1f%% swap=%.1f%% disks=%d ports=%d procs=%d fd=%d/%d cs=%.0f/s probes=%d",
|
||
report.CPU.UsagePercent,
|
||
report.Memory.UsagePercent,
|
||
func() float64 {
|
||
if report.Swap != nil {
|
||
return report.Swap.UsagePercent
|
||
}
|
||
return 0
|
||
}(),
|
||
len(report.Disks), len(report.Ports), len(report.ProcessTop),
|
||
func() int {
|
||
if report.FD != nil {
|
||
return report.FD.Allocated
|
||
}
|
||
return 0
|
||
}(),
|
||
func() int64 {
|
||
if report.FD != nil {
|
||
return report.FD.Max
|
||
}
|
||
return 0
|
||
}(),
|
||
func() float64 {
|
||
if report.ContextSwitches != nil {
|
||
return report.ContextSwitches.CTXPerSec
|
||
}
|
||
return 0
|
||
}(),
|
||
probeCount,
|
||
)
|
||
}
|
||
|
||
// ═══ pollAndExecute 任务轮询 ═══
|
||
pollAndExecute := func() {
|
||
tasks, err := taskClient.Poll(ctx)
|
||
if err != nil {
|
||
if ctx.Err() != nil {
|
||
return
|
||
}
|
||
logger.Printf("任务轮询失败: %v", err)
|
||
return
|
||
}
|
||
if len(tasks) > 0 {
|
||
logger.Printf("拉取到 %d 个待执行任务", len(tasks))
|
||
for _, t := range tasks {
|
||
if ctx.Err() != nil {
|
||
return
|
||
}
|
||
result := taskExecutor.ExecuteTask(ctx, t)
|
||
if err := taskClient.Submit(ctx, result); err != nil {
|
||
logger.Printf("任务 %s 结果回传失败: %v", t.TaskID, err)
|
||
} else {
|
||
logger.Printf("任务 %s 结果回传成功 status=%s", t.TaskID, result.Status)
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
// ═══ pullProbes 探针配置拉取(v2.1 独立于上报)═══
|
||
pullProbes := func() {
|
||
configs, err := taskClient.PollProbes(ctx)
|
||
if err != nil {
|
||
if ctx.Err() != nil {
|
||
return
|
||
}
|
||
logger.Printf("探针配置拉取失败: %v", err)
|
||
return
|
||
}
|
||
if len(configs) > 0 {
|
||
probeConfigs = configs
|
||
}
|
||
}
|
||
|
||
// ═══ 任务轮询 + 探针拉取协程(5s 间隔)═══
|
||
const pollInterval = 5 * time.Second
|
||
go func() {
|
||
pollAndExecute()
|
||
pullProbes()
|
||
ticker := time.NewTicker(pollInterval)
|
||
defer ticker.Stop()
|
||
for {
|
||
select {
|
||
case <-ticker.C:
|
||
pollAndExecute()
|
||
pullProbes()
|
||
case <-ctx.Done():
|
||
return
|
||
}
|
||
}
|
||
}()
|
||
|
||
// ═══ 主循环:采集 + 上报 ═══
|
||
time.Sleep(startupJitter)
|
||
collectAndReport()
|
||
ticker := time.NewTicker(time.Duration(cfg.Interval) * time.Second)
|
||
defer ticker.Stop()
|
||
for {
|
||
select {
|
||
case <-ticker.C:
|
||
collectAndReport()
|
||
case <-ctx.Done():
|
||
logger.Println("收到退出信号,Agent 正常关闭")
|
||
return
|
||
}
|
||
}
|
||
}
|
||
|
||
// detectLocalIP 探测本机主 IP
|
||
func detectLocalIP() string {
|
||
conn, err := net.Dial("udp", "8.8.8.8:53")
|
||
if err == nil {
|
||
defer conn.Close()
|
||
if addr, ok := conn.LocalAddr().(*net.UDPAddr); ok && addr.IP != nil {
|
||
return addr.IP.String()
|
||
}
|
||
}
|
||
ifaces, err := net.Interfaces()
|
||
if err == nil {
|
||
for _, iface := range ifaces {
|
||
if iface.Flags&net.FlagLoopback != 0 || iface.Flags&net.FlagUp == 0 {
|
||
continue
|
||
}
|
||
addrs, _ := iface.Addrs()
|
||
for _, addr := range addrs {
|
||
if ipnet, ok := addr.(*net.IPNet); ok && ipnet.IP.To4() != nil {
|
||
return ipnet.IP.String()
|
||
}
|
||
}
|
||
}
|
||
}
|
||
return "unknown"
|
||
}
|