filebeat源码分析服务启动

在开始源码分析之前先说一下filebeat是什么？beats是知名的ELK日志分析套件的一部分。它的前身是logstash-forwarder，用于收集日志并转发给后端（logstash、elasticsearch、redis、kafka等等）。filebeat是beats项目中的一种beats，负责收集日志文件的新增内容。当前的代码分支是最新的6.x的代码。
先看我们服务启动配置文件的一个例子，这个是filebeat采集k8s的日志的一个例子：

filebeat.prospectors:
- type: log
  paths:
     - /var/lib/docker/containers/*/*-json.log
     - /var/log/filelog/containers/*/*/*/*.log

processors:
- add_docker_metadata:
    host: "unix:///var/run/docker.sock"
- add_fields:
    fields:
      log: '{message}'
- decode_json_fields:
    when:
       regexp:
         message: "{*}"
    fields: ["message"]
    overwrite_keys: true
    target: ""
- drop_fields:
    fields: ["docker.container.labels.annotation.io.kubernetes.container.terminationMessagePath", "docker.container.labels.annotation.io.kubernetes.container.hash", "docker.container.labels.annotation.io.kubernetes.container.terminationMessagePolicy", "docker.container.labels.annotation.io.kubernetes.pod.terminationGracePeriod", "beat.version", "docker.container.labels.annotation.io.kubernetes.container.ports", "docker.container.labels.io.kubernetes.container.terminationMessagePath", "docker.container.labels.io.kubernetes.container.restartCount", "docker.container.labels.io.kubernetes.container.ports", "docker.container.labels.io.kubernetes.container.hash", "docker.container.labels.io.kubernetes.pod.terminationGracePeriod", "docker.container.labels.annotation.io.kubernetes.container.restartCount", "message"]
- parse_level:
    levels: ["fatal", "error", "warn", "info", "debug"]
    field: "log"

logging.level: info
setup.template.enabled: true
setup.template.name: "filebeat-%{+yyyy.MM.dd}"
setup.template.pattern: "filebeat-*"
setup.template.fields: "${path.config}/fields.yml"
setup.template.overwrite: true
setup.template.settings:
  index:
     analysis:
       analyzer:
         enncloud_analyzer:
           filter: ["standard", "lowercase", "stop"]
           char_filter: ["my_filter"]
           type: custom
           tokenizer: standard
       char_filter:
         my_filter:
           type: mapping
           mappings: ["-=>_"]

output:
  elasticsearch:
    hosts: ["paasdev.enncloud.cn:9200"]
    index: "filebeat-%{+yyyy.MM.dd}"

filebeat启动时候会加载这个配置文件。再看看总结的接口libbeat/beat/beat.go

type Beater interface {
    // The main event loop. This method should block until signalled to stop by an
    // invocation of the Stop() method.
    Run(b *Beat) error

    // Stop is invoked to signal that the Run method should finish its execution.
    // It will be invoked at most once.
    Stop()
}

这个是每个beat都需要实现的两个接口，当然filebeat也不例外，filebeat/beater/filebeat.go
这个里面是filebeat的具体实现，篇幅有限就，省略的粘贴一下

// Run allows the beater to be run as a beat.
func (fb *Filebeat) Run(b *beat.Beat) error {
    var err error
    config := fb.config

    if !fb.moduleRegistry.Empty() {
        err = fb.loadModulesPipelines(b)
        if err != nil {
            return err
        }
    }
    // Setup registrar to persist state
    registrar, err := registrar.New(config.RegistryFile, config.RegistryFlush, finishedLogger)
    if err != nil {
        logp.Err("Could not init registrar: %v", err)
        return err
    }

    err = b.Publisher.SetACKHandler(beat.PipelineACKHandler{
        ACKEvents: newEventACKer(registrarChannel).ackEvents,
    })
    if err != nil {
        logp.Err("Failed to install the registry with the publisher pipeline: %v", err)
        return err
    }


    crawler, err := crawler.New(
        channel.NewOutletFactory(outDone, b.Publisher, wgEvents).Create,
        config.Prospectors,
        b.Info.Version,
        fb.done,
        *once)
    if err != nil {
        logp.Err("Could not init crawler: %v", err)
        return err
    }


    err = registrar.Start()
    if err != nil {
        return fmt.Errorf("Could not start registrar: %v", err)
    }


    var pipelineLoaderFactory fileset.PipelineLoaderFactory
    if b.Config.Output.Name() == "elasticsearch" {
        pipelineLoaderFactory = newPipelineLoaderFactory(b.Config.Output.Config())
    } else {
        logp.Warn(pipelinesWarning)
    }

    err = crawler.Start(registrar, config.ConfigProspector, config.ConfigModules, pipelineLoaderFactory)
    if err != nil {
        crawler.Stop()
        return err
    }

    var adiscover *autodiscover.Autodiscover
    if fb.config.Autodiscover != nil {
        adapter := NewAutodiscoverAdapter(crawler.ProspectorsFactory, crawler.ModulesFactory)
        adiscover, err = autodiscover.NewAutodiscover("filebeat", adapter, config.Autodiscover)
        if err != nil {
            return err
        }
    }
    adiscover.Start()

    return nil
}

// Stop is called on exit to stop the crawling, spooling and registration processes.
func (fb *Filebeat) Stop() {
    logp.Info("Stopping filebeat")

    // Stop Filebeat
    close(fb.done)
}

上面的代码省略的介绍了start和stop的函数，stop天简单，就是一个关闭的总开关，就不说了。
详细分析一下这个start方法，它是整个filebeat最核心的地方。
filebeat支持采集特定程序的日志，譬如redis、nginx等，这些都是通过module支持的，所以在程序开始时候先确定elasticsearch里面有没有这些关联的pipeline、ingest，

    if !fb.moduleRegistry.Empty() {
        err = fb.loadModulesPipelines(b)
        if err != nil {
            return err
        }
    }

深入看看注册方法

func (fb *Filebeat) loadModulesPipelines(b *beat.Beat) error {
    if b.Config.Output.Name() != "elasticsearch" {
        logp.Warn(pipelinesWarning)
        return nil
    }

    // 这里注册一个回调的方法，每当和一个es建立连接的时候，都会重新和es确认这些pipeline
    callback := func(esClient *elasticsearch.Client) error {
        return fb.moduleRegistry.LoadPipelines(esClient)
    }
    elasticsearch.RegisterConnectCallback(callback)

    return nil
}

上面的代码主要是先和es确认一下pipeline，下面接着看启动，然后就创建registrar，registrar是啥呢？

registrar, err := registrar.New(config.RegistryFile, config.RegistryFlush, finishedLogger)
    if err != nil {
        logp.Err("Could not init registrar: %v", err)
        return err
    }

其实它是注册日志读取进度的，通过记录offset,下面就是我截取的一段registry文件。

{"source":"/var/lib/docker/containers/e892ad615535e877c8af5856bd27631937d050d00b4ca55554bec41e3391685f/e892ad615535e877c8af5856bd27631937d050d00b4ca55554bec41e3391685f-json.log","offset":0,"timestamp":"2017-11-29T17:01:28.645203497Z","ttl":-1,"type":"log","FileStateOS":{"inode":526963,"device":64769}}

这个json文件里面保存了容器和对应的offset，这样当filebeat重启过后则能继续工作。

然后创建crawler，这个是负责日志采集的。

crawler, err := crawler.New(
        channel.NewOutletFactory(outDone, b.Publisher, wgEvents).Create,
        config.Prospectors,
        b.Info.Version,
        fb.done,
        *once)

通过config.Prospectors，crawler就知道采集哪些目标。然后就通过

err = crawler.Start(registrar, config.ConfigProspector, config.ConfigModules, pipelineLoaderFactory)
    if err != nil {
        crawler.Stop()
        return err
    }

启动采集任务。下面看看具体启动任务地方，filebeat/crawler/crawler.go

for _, prospectorConfig := range c.prospectorConfigs {
        err := c.startProspector(prospectorConfig, r.GetStates())
        if err != nil {
            return err
        }
    }

就来到filebeat/prospector/prospector.go里面

func (p *Prospector) Run() {
    // Initial prospector run
    p.prospectorer.Run()

    // Shuts down after the first complete run of all prospectors
    if p.Once {
        return
    }

    for {
        select {
        case <-p.done:
            logp.Info("Prospector ticker stopped")
            return
        case <-time.After(p.config.ScanFrequency):
            logp.Debug("prospector", "Run prospector")
            p.prospectorer.Run()
        }
    }
}

这个prospectorer.Run是一个接口，可以支持从UDP/STDIN/LOG/REDIS/DOCKER里面直接获取日志，我们看一个log的filebeat/prospector/log/prospector.go

func (p *Prospector) Run() {
...
p.scan()
...
}

这里如果发现文件需要被采集，则创建采集任务

if lastState.IsEmpty() {
    logp.Debug("prospector", "Start harvester for new file: %s", newState.Source)
    err := p.startHarvester(newState, 0)
    if err != nil {
        logp.Err("Harvester could not be started on new file: %s, Err: %s", newState.Source, err)
        }
    } else {
        p.harvestExistingFile(newState, lastState)
}

startHarvester启动日志采集，还是相同的套路，先创建createHarvester，然后启动harvesters.Start(h)。
这个里面通过for死循环里面执行

message, err := h.reader.Next()

这样分批读取。启动程序的内容先说到这里。还有很多细节后面逐一描述。