prometheus -- 错开时间拉取(scrape jitterSeed)分析

prometheus拉取target的metrics时,都是错开时间分别拉取的,这样有以下好处:

  • 避免prometheus实例启动太多的goroutine同时拉取;
  • 多prometheus实例拉取同一个target时,避免同时拉取对target造成流量压力;

单prometheus实例内的不同target

单实例内的不同target,不是在某个时刻一起拉取,而是错开时间分别拉取。比如scrape_interval=30s,不同target的拉取时间如下:

  • job=prometheus:20秒、50秒拉取;
  • job=node-exporter:04秒、34秒拉取;
  • job=pvc-test:07秒、37秒拉取;
> select * from up where job='prometheus' order by time desc limit 5;
name: up
time                     __name__ __type__ instance       job        value
----                     -------- -------- --------       ---        -----
2021-09-24T05:52:20.665Z up       gauge    localhost:9090 prometheus 1
2021-09-24T05:51:50.665Z up       gauge    localhost:9090 prometheus 1
2021-09-24T05:51:20.665Z up       gauge    localhost:9090 prometheus 1
2021-09-24T05:50:50.665Z up       gauge    localhost:9090 prometheus 1
2021-09-24T05:50:20.665Z up       gauge    localhost:9090 prometheus 1
>
> select * from up where job='node-exporter' order by time desc limit 5;
name: up
time                     __name__ __type__ instance        job           value
----                     -------- -------- --------        ---           -----
2021-09-24T05:52:34.184Z up       gauge    10.21.1.74:9100 node-exporter 0
2021-09-24T05:52:04.183Z up       gauge    10.21.1.74:9100 node-exporter 0
2021-09-24T05:51:34.183Z up       gauge    10.21.1.74:9100 node-exporter 0
2021-09-24T05:51:04.183Z up       gauge    10.21.1.74:9100 node-exporter 0
2021-09-24T05:50:34.183Z up       gauge    10.21.1.74:9100 node-exporter 0
>
> select * from up where job='pvc-test' order by time desc limit 5;
name: up
time                     __name__ __type__ instance       job      value
----                     -------- -------- --------       ---      -----
2021-09-24T05:53:07.315Z up       gauge    127.0.0.1:9999 pvc-test 0
2021-09-24T05:52:37.315Z up       gauge    127.0.0.1:9999 pvc-test 0
2021-09-24T05:52:07.315Z up       gauge    127.0.0.1:9999 pvc-test 0
2021-09-24T05:51:37.315Z up       gauge    127.0.0.1:9999 pvc-test 0
2021-09-24T05:51:07.315Z up       gauge    127.0.0.1:9999 pvc-test 0

多prometheus实例

在HA场景下,不同的prometheus实例可能拉取相同的target。
多个prometheus实例同时启动时,开始拉取的时间也要随机错开,否则会出现多个实例同时拉取同一个target的现象。

实现机制

通过全局的scrape jitterSeed实现:

  • 不同实例:jitterSeed是跟特定prometheus实例相关的uint64哈希值,不同实例的jitterSeed不同;
  • 单实例不同target: 使用target哈希值与jitterSeed做异或操作,得到target拉取的启动时间,不同的target启动时间不同;

prometheus -- 错开时间拉取(scrape jitterSeed)分析_第1张图片

prometheus实例的scrape jitterSeed的计算

// unit64
jitterSeed=hash({hostname}+{extertalLabels})

不同实例的hostname/externalLabels一般不同,故计算出来的jitterSeed值也不同。

// scrape/manager.go
// setJitterSeed calculates a global jitterSeed per server relying on extra label set.
func (m *Manager) setJitterSeed(labels labels.Labels) error {
    h := fnv.New64a()
    hostname, err := getFqdn()
    if err != nil {
        return err
    }
    if _, err := fmt.Fprintf(h, "%s%s", hostname, labels.String()); err != nil {
        return err
    }
    m.jitterSeed = h.Sum64()
    return nil
}

prometheus实例内target拉取启动时间的计算
target被拉取前,先随机等待一个时间,然后再用interval固定间隔循环拉取:

// scrape/scrape.go
func (sl *scrapeLoop) run(interval, timeout time.Duration, errc chan<- error) {
    //等待一段时间再开始
    select {
    case <-time.After(sl.scraper.offset(interval, sl.jitterSeed)):
        // Continue after a scraping offset.
    case <-sl.ctx.Done():
        close(sl.stopped)
        return
    }
    // interval固定间隔拉取
    ticker := time.NewTicker(interval)
    defer ticker.Stop()
    
    for {
        // HTTP拉取
        last = sl.scrapeAndReport(interval, timeout, last, errc)
        select {
        ......
        case <-ticker.C:
        }
    }
}   

不同target有不同的等待时间,这样就可以将不同target的开始时间错开。

//target等待时间的计算方法
offset=((hash(target) ^ jitterSeed) % interval)

代码实现:

// scrape/target.go
// offset returns the time until the next scrape cycle for the target.
// It includes the global server jitterSeed for scrapes from multiple Prometheus to try to be at different times.
func (t *Target) offset(interval time.Duration, jitterSeed uint64) time.Duration {
    now := time.Now().UnixNano()

    // Base is a pinned to absolute time, no matter how often offset is called.
    var (
        base   = int64(interval) - now%int64(interval)
        offset = (t.hash() ^ jitterSeed) % uint64(interval)
        next   = base + int64(offset)
    )

    if next > int64(interval) {
        next -= int64(interval)
    }
    return time.Duration(next)
}

你可能感兴趣的:(prometheus)