Google Pub/Sub 的 RetryPolicy 中配置的指数退避如何工作？

Question

最近发布的cloud.google.com/go/pubsub库（在 v1.5.0 中，参见https://github.com/googleapis/google-cloud-go/releases/tag/pubsub%2Fv1.5.0 ）支持新的RetryPolicy服务器端功能。 当前的文档（ https://godoc.org/cloud.google.com/go/pubsub#RetryPolicy ）为

我读过 Wikipedia 文章，虽然它描述了离散时间的指数退避，但我看不出这篇文章与MinimumBackoff和MaximumBackoff参数的具体关系。 For guidance on this, I referred to the documentation for github.com/cenkalti/backoff , https://pkg.go.dev/github.com/cenkalti/backoff/v4?tab=doc#ExponentialBackOff . 该库将ExponentialBackoff定义为

type ExponentialBackOff struct {
    InitialInterval     time.Duration
    RandomizationFactor float64
    Multiplier          float64
    MaxInterval         time.Duration
    // After MaxElapsedTime the ExponentialBackOff returns Stop.
    // It never stops if MaxElapsedTime == 0.
    MaxElapsedTime time.Duration
    Stop           time.Duration
    Clock          Clock
    // contains filtered or unexported fields
}

其中每个随机间隔计算为

randomized interval =
    RetryInterval * (random value in range [1 - RandomizationFactor, 1 + RandomizationFactor])

其中RetryInterval是当前的重试间隔，据我所知，它从MaxInterval的值开始，并以InitialInterval为上限。

我是否正确理解MinimumBackoff和MaximumBackoff对应于github.com/cenkalti/backoff中的InitialInterval和MaxInterval ？ 也就是说， MinimumBackoff是初始等待时间， MaximumBackoff是重试之间允许的最大时间量？

为了测试我的理论，我编写了以下简化程序：

package main

import (
    "context"
    "flag"
    "fmt"
    "log"
    "os"
    "time"

    "cloud.google.com/go/pubsub"
    "google.golang.org/grpc/codes"
    "google.golang.org/grpc/status"
)

var (
    projectID                      string
    minimumBackoff, maximumBackoff time.Duration
)

const (
    topicName             = "test-topic"
    subName               = "test-subscription"
    defaultMinimumBackoff = 10 * time.Second
    defaultMaximumBackoff = 10 * time.Minute
)

func main() {
    flag.StringVar(&projectID, "projectID", "my-project", "Google Project ID")
    flag.DurationVar(&minimumBackoff, "minimumBackoff", 5*time.Second, "minimum backoff")
    flag.DurationVar(&maximumBackoff, "maximumBackoff", 60*time.Second, "maximum backoff")
    flag.Parse()
    log.Printf("Running with minumum backoff %v and maximum backoff %v...", minimumBackoff, maximumBackoff)

    retryPolicy := &pubsub.RetryPolicy{MinimumBackoff: minimumBackoff, MaximumBackoff: maximumBackoff}

    client, err := pubsub.NewClient(context.Background(), projectID)
    if err != nil {
        log.Fatalf("NewClient: %v", err)
    }

    topic, err := client.CreateTopic(context.Background(), topicName)
    if err != nil {
        log.Fatalf("CreateTopic: %v", err)
    }
    log.Printf("Created topic %q", topicName)
    defer func() {
        topic.Stop()
        if err := topic.Delete(context.Background()); err != nil {
            log.Fatalf("Delete topic: %v", err)
        }
        log.Printf("Deleted topic %s", topicName)
    }()

    sub, err := client.CreateSubscription(context.Background(), subName, pubsub.SubscriptionConfig{
        Topic:       topic,
        RetryPolicy: retryPolicy,
    })
    if err != nil {
        log.Fatalf("CreateSubscription: %v", err)
    }
    log.Printf("Created subscription %q", subName)
    defer func() {
        if err := sub.Delete(context.Background()); err != nil {
            log.Fatalf("Delete subscription: %v", err)
        }
        log.Printf("Deleted subscription %q", subName)
    }()

    go func() {
        sub.Receive(context.Background(), func(ctx context.Context, msg *pubsub.Message) {
            log.Printf("Nacking message: %s", msg.Data)
            msg.Nack()
        })
    }()

    topic.Publish(context.Background(), &pubsub.Message{Data: []byte("Hello, world!")})
    log.Println("Published message")
    time.Sleep(60 * time.Second)
}

如果我分别使用 5s 和 60s 的 flag-default MinimumBackoff和MaximumBackoff运行它，我会得到以下 output：

> go run main.go
2020/07/29 18:49:32 Running with minumum backoff 5s and maximum backoff 1m0s...
2020/07/29 18:49:33 Created topic "test-topic"
2020/07/29 18:49:34 Created subscription "test-subscription"
2020/07/29 18:49:34 Published message
2020/07/29 18:49:36 Nacking message: Hello, world!
2020/07/29 18:49:45 Nacking message: Hello, world!
2020/07/29 18:49:56 Nacking message: Hello, world!
2020/07/29 18:50:06 Nacking message: Hello, world!
2020/07/29 18:50:17 Nacking message: Hello, world!
2020/07/29 18:50:30 Nacking message: Hello, world!
2020/07/29 18:50:35 Deleted subscription "test-subscription"
2020/07/29 18:50:35 Deleted topic test-topic

而如果我分别以 1s 和 2s 的MinimumBackoff和MaximumBackoff运行它，我得到

> go run main.go --minimumBackoff=1s --maximumBackoff=2s
2020/07/29 18:50:42 Running with minumum backoff 1s and maximum backoff 2s...
2020/07/29 18:51:11 Created topic "test-topic"
2020/07/29 18:51:12 Created subscription "test-subscription"
2020/07/29 18:51:12 Published message
2020/07/29 18:51:15 Nacking message: Hello, world!
2020/07/29 18:51:18 Nacking message: Hello, world!
2020/07/29 18:51:21 Nacking message: Hello, world!
2020/07/29 18:51:25 Nacking message: Hello, world!
2020/07/29 18:51:28 Nacking message: Hello, world!
2020/07/29 18:51:31 Nacking message: Hello, world!
2020/07/29 18:51:35 Nacking message: Hello, world!
2020/07/29 18:51:38 Nacking message: Hello, world!
2020/07/29 18:51:40 Nacking message: Hello, world!
2020/07/29 18:51:44 Nacking message: Hello, world!
2020/07/29 18:51:47 Nacking message: Hello, world!
2020/07/29 18:51:50 Nacking message: Hello, world!
2020/07/29 18:51:52 Nacking message: Hello, world!
2020/07/29 18:51:54 Nacking message: Hello, world!
2020/07/29 18:51:57 Nacking message: Hello, world!
2020/07/29 18:52:00 Nacking message: Hello, world!
2020/07/29 18:52:03 Nacking message: Hello, world!
2020/07/29 18:52:06 Nacking message: Hello, world!
2020/07/29 18:52:09 Nacking message: Hello, world!
2020/07/29 18:52:12 Nacking message: Hello, world!
2020/07/29 18:52:13 Deleted subscription "test-subscription"
2020/07/29 18:52:13 Deleted topic test-topic

在后一个示例中，nack 之间的时间似乎非常一致~3s，这大概代表了在 2s 的MaximumBackoff中做到这一点的“最大努力”？ 我仍然不清楚的是是否有任何随机化，是否有乘数（从第一个例子来看，重试之间的时间似乎不是每次都加倍），以及是否有等价的的MaxElapsedTime超过了没有更多的重试？

Answer 1

最小退避和最大退避的重试策略字段类似于上面示例中的 InitialInterval 和 MaxInterval。 Cloud Pub/Sub 使用您提到的类似公式来计算指数延迟。 这也包括随机化。

在 MaxInterval 之后，每次后续重试都会增加 MaxInterval 延迟。 如果您想在尝试一定次数后停止重试，我们建议使用死信队列。