站点图标 Linux-技术共享

基于Prometheus + Grafana的服务性能,VPS质量监控实践

Linux网站管理员

服务的质量离不开一个监控和反馈体系，如果没有一个监控反馈体系，我们并不能知道线上的服务出现了什么，正在发生什么、可能存在什么样的问题。

这时候如果有一个可视化的表盘，可以让你直观清晰的感知整个系统的运行状况，那么你就能根据这些反馈做出调整，进入一个【行动->反馈->行动..】的过程之中。渐渐的，整个系统的服务质量在这些不断的调整中就大大提高了，极为重要的一点是我们有必要对自己开发的服务质量负责。

搜索接口性能从一开始的混乱无序，到后续的逐渐稳定。

怎么做？

基于Prometheus，我们可以按时间间隔定义采集任务，用拉的方式去采集各个服务应用上的性能指标。基于Grafana我们可以对这些采集到的指标进行可视化展示，查询分析，此外Grafana还支持报警规则，再通过Webhook、我们能把这些报警推送到飞书或者叮叮群里，及时响应异常情况。

指标类型

Prometheus主要支持三种类型一种分别是Counter，Histogram，Gauge，Summary

  • Counter主要用来记录一个总数，比如说api请求总次数

  • Gauge主要记录可上可下的值，比如CPU使用率

  • Summary可以用来记录接口的响应时间

  • Histogram和Summary可以做差不多同样的事，就是多了一个bucket的概念，响应时间按范围分段统计，具体差异详见

接口准备

Prometheus主要是通过定时抓取目标{target}/metrics 接口来采集数据的，所以我们需要提供一个接口来暴露这些数据

Nodejs

以express为例，安装prom-clien response-time依赖

require('prom-client').collectDefaultMetrics();

记录一些默认的指标比如说当前nodejs进程内存 cpu gc http request等各种系统指标。

var Register = require('prom-client').register;  
var Counter = require('prom-client').Counter;  
var Histogram = require('prom-client').Histogram;  
var Summary = require('prom-client').Summary;  
var ResponseTime = require('response-time');  
// var Logger = require('logger').createLogger();

/**
 * A Prometheus counter that counts the invocations of the different HTTP verbs
 * e.g. a GET and a POST call will be counted as 2 different calls
 */
var numOfRequests = new Counter({  
    name: 'numOfRequests',
    help: 'Number of requests made',
    labelNames: ['method']
});

/**
 * A Prometheus counter that counts the invocations with different paths
 * e.g. /foo and /bar will be counted as 2 different paths
 */
var pathsTaken = new Counter({  
    name: 'pathsTaken',
    help: 'Paths taken in the app',
    labelNames: ['path']
});

/**
 * A Prometheus summary to record the HTTP method, path, response code and response time
 */
var responses = new Summary({  
    name: 'responses',
    help: 'Response time in millis',
    labelNames: ['method', 'path', 'status']
});

/**
 * This funtion will start the collection of metrics and should be called from within in the main js file
 */
module.exports.startCollection = function () {  
    // Logger.info(Logger.LOG_INFO, 'Starting the collection of metrics, the metrics are available on /metrics');
    require('prom-client').collectDefaultMetrics();
};

/**
 * This function increments the counters that are executed on the request side of an invocation
 * Currently it increments the counters for numOfPaths and pathsTaken
 */
module.exports.requestCounters = function (req, res, next) {  
    if (req.path != '/metrics') {
        numOfRequests.inc({ method: req.method });
        pathsTaken.inc({ path: req.path });
    }
    next();
}

/**
 * This function increments the counters that are executed on the response side of an invocation
 * Currently it updates the responses summary
 */
module.exports.responseCounters = ResponseTime(function (req, res, time) {  
    if(req.url != '/metrics') {
        responses.labels(req.method, req.url, res.statusCode).observe(time);
    }
})

/**
 * In order to have Prometheus get the data from this app a specific URL is registered
 */
module.exports.injectMetricsRoute = function (App) {  
    App.get('/metrics', (req, res) => {
        res.set('Content-Type', Register.contentType);
        res.end(Register.metrics());
    });
};

记录不同route的请求次数numOfRequests ，和响应时间responses

app.use(Prometheus.requestCounters);  
app.use(Prometheus.responseCounters);

/**
 * Enable metrics endpoint
 */
Prometheus.injectMetricsRoute(app);

/**
 * Enable collection of default metrics
 */
Prometheus.startCollection();

中间件加上，注入/metrics接口

PHP

php我们要用到php-apc 或者redis也可以 因为php自己是被web server执行完了就内存释放了，计数统计还得依赖redis或者apc这样的

php7得安装php-apcu还得安装apcu_bc, php7可以用 endclothing/prometheus_client_php, 5.x可以用 jimdo/prometheus_client_php

以laravel为例

我们新加一个中间件Metrics.php

namespace App\Http\Middleware;

use Closure;
use Illuminate\Support\Facades\Auth;
use Illuminate\Http\Request;
use Prometheus\CollectorRegistry;

class Metrics
{

    public function __construct()
    {   
        if(function_exists('apcu_add')) {
            $adapter = new \Prometheus\Storage\APC();
        } else {
            $adapter = new \Prometheus\Storage\InMemory();
        }
        
        $registry = new CollectorRegistry($adapter);
        $this->registry = $registry;
        $this->initRouteMetrics();
    }

    public function initRouteMetrics()
    {
        $namespace = "api";
        $buckets = [
            100.0,
            200.0,
            300.0,
            600.0,
            1000.0,
            1500.0,
            2500.0,
            5000.0,
            8000.0,
            15000.0,
            30000.0,
            50000.0,
            80000.0
        ];
        // $buckets = null;
        $labelNames = $this->getRequestCounterLabelNames();

        $name = 'request_duration_milliseconds';
        $help = 'duration of http_requests';
        $this->requestDurationHistogram = $this->registry->getOrRegisterHistogram(
            $namespace, $name, $help, $labelNames, $buckets
        );
    }


    protected function getRequestCounterLabelNames()
    {
        return [
            'route', 'method', 'status_code',
        ];
    }

    /**
     * Handle an incoming request.
     *
     * @param  \Illuminate\Http\Request  $request
     * @param  \Closure  $next
     * @param  string|null  $guard
     * @return mixed
     */
    public function handle(Request $request, Closure $next)
    {

        $start = $_SERVER['REQUEST_TIME_FLOAT'];
        $this->request = $request;

        /** @var \Illuminate\Http\Response $response */
        $response = $next($request);

        $route = $request->route();

        if($route) {
            $route_name = $route->uri() ?: 'unnamed';
            $method = $request->getMethod();
            $status = $response->getStatusCode();
            $duration = microtime(true) - $start;
            $duration_milliseconds = $duration * 1000.0;
            $this->countRequest($route_name, $method, $status, $duration_milliseconds);
        }

        return $response;
    }

    public function countRequest($route, $method, $statusCode, $duration_milliseconds)
    {
        $labelValues = [(string)$route, (string)$method, (string) $statusCode];
        $this->requestDurationHistogram->observe($duration_milliseconds, $labelValues);
    }
}

按route统计api的响应时间

Prometheus数据采集

docker-compose.yml

version: "2.2"
services:
  tapi:
    logging:
      driver: "json-file"
      options:
          max-size: "50m"
    image: prom/prometheus
    restart: always
    volumes:
      - /prometheus_data:/prometheus_data:rw
      - ./prometheus.yml:/etc/prometheus/prometheus.yml
      - ./targets.json:/etc/prometheus/targets.json
    command:
      - '--storage.tsdb.path=/prometheus_data'
      - '--config.file=/etc/prometheus/prometheus.yml'
      - '--storage.tsdb.retention=10d'
    ports:
      - "9090:9090"

挂载一个prometheus_data外部目录来对数据持久化，config设置保留10天的数据

prometheus.yml

global:
  scrape_interval:     15s # Set the scrape interval to every 15 seconds. Default is every 1 minute.
  evaluation_interval: 15s # Evaluate rules every 15 seconds. The default is every 1 minute.

scrape_configs:
  - job_name: "node"
    file_sd_configs:
    - refresh_interval: 1m
      files: 
        - "/etc/prometheus/targets.json"
  - job_name: 'prometheus'
    static_configs:
      - targets: ['localhost:9090']
  - job_name: web
    static_configs:
      - targets: ['192.168.1.22:8096', '192.168.1.89:8096']

file_sd会定时动态读取targets.json在不重启的情况下增加新的job

targets.json

[
    {
      "targets": [ "192.168.1.86:8029" ],
      "labels": {
        "env": "prod",
        "job": "api"
      }
    },
    {
      "targets": [ "192.168.1.86:8030"],
      "labels": {
        "env": "prod",
        "job": "api"
      }
    }
  ]

数据指标可视化

启动Grafana

docker run -d --name=grafana -p 3000:3000 grafana/grafana
配置数据源

Configuration > Add data source 选择Prometheus，URL=http://localhost:9090

配置Dashboard

增加一个特定接口的性能监控

5分钟内请求总时间除以总次数得到api响应时间均值

rate(api_request_duration_milliseconds_sum{route="g/search", method="POST"}[5m]) 
/ 
rate(api_request_duration_milliseconds_count{route="g/search", method="POST"}[5m])
发现慢接口

topk(20, 
  sum by (route) (
    rate(api_request_duration_milliseconds_sum{status_code!="500", method!="OPTIONS"}[2m]) 
    / 
    rate(api_request_duration_milliseconds_count{status_code!="500", method!="OPTIONS"}[2m])
  )
)

sum byroute 的响应时间 取Top20，得到响应时间慢的API

topk(10, sum by (route) (rate(api_request_duration_milliseconds_count{status_code!="500"}[5m])))

稍微调整下可以拿到当前时间请求频繁的接口

报警

设置Alerting > Notification channels

由于feishu的webhook传参格式不一样，我们需要格式化一下信息格式，做个转发

所以是 grafana > webhook proxy > feishu

const express = require('express')
const bodyParser = require('body-parser')
const app = express()
const fetch = require('node-fetch');

app.use(bodyParser.json());
app.use(bodyParser.urlencoded({extended: false}));

app.post('/webhook/feishu/bot1', async (req, res, next) => {
    const body = req.body;
    const matchs = body.evalMatches.length ? JSON.stringify(body.evalMatches, null, 2) : '';
    const formatted = {
        title: body.title,
        text: `${body.message}\n\nruleName:${body.ruleName}\nstate:${body.state}\n\n${matchs}\n`
    };
    const proxyRes = await fetch('https://open.feishu.cn/open-apis/bot/hook/xxxx', {
        method: 'post',
        body:    JSON.stringify(formatted),
        headers: { 'Content-Type': 'application/json' }
    });
    console.log(await proxyRes.json());
    res.send('hello');
});

app.listen(3002);
规则设置

设置一个当搜索接口响应时间高于某个值的报警通知

收到报警提醒

退出移动版