Building high performance push notification server in Go

Transcript

1. Tatsuhiko Kubo@cubicdaiya
   buildercon tokyo 2017/08/04
   Building high performance
   push notification server in Go
   

   View Slide

2. @cubicdaiya / Tatsuhiko Kubo
   Principal Engineer, SRE @ Mercari, Inc.
   

   View Slide

3. View Slide

4. • How to send push notification to iOS or Android
   device
   • Past and present push infrastructure @
   • Why did we develop push notification server?
   • Gaurun ~General push notificatoin server in Go~
   • Features
   • Artchitecture and internals
   Agenda
   

   View Slide

5. Push notification to iPhone
   

   View Slide

6. • Notification is not pushed to smartphone directly
   • Notification payload is sent to push notification
   service such as
   • APNs, GCM/FCM, Amazon SNS, etc…
   • Only APNs and GCM/FCM are targets in this talk
   Push notification to iOS or Android device
   

   View Slide

7. • APNs
   • Via HTTPS
   • TLS certificate and key or JWT is required
   • GCM/FCM
   • Via HTTPS
   • Server key is required
   Communicating with APNs and GCM/FCM
   

   View Slide

8. • APNs Binary Provider API (Legacy)
   • Binary protocol on TLS
   • APNs Provider API
   • HTTP/2
   • Payload is JSON
   APNs (Apple Push Notification Service)
   

   View Slide

9. • GCM (Google Cloud Messaging)
   • FCM (Firebase Cloud Messaging)
   • Google says in https://developers.google.com/cloud-messaging/
   GCM / FCM
   Firebase Cloud Messaging (FCM) is the new version of GCM.
   It inherits the reliable and scalable GCM infrastructure, plus
   new features! See the FAQ to learn more. If you are integrating
   messaging in a new app, start with FCM. GCM users are
   strongly recommended to upgrade to FCM, in order to benefit
   from new FCM features today and in the future.
   

   View Slide

10. • High network latency
    • APNs and GCM/FCM endpoint is far
    • It takes between tens and hundreds millseconds
    to push
    • Connection handling
    • Keep-alive as possible
    • Frequent connect / close is bad
    Communicating with APNs and GCM/FCM
    

    View Slide

11. Push infrastructure @
    

    View Slide

12. Overview
    APNs
    HTTP/2
    GCM
    HTTP/2
    HTTP/1.1
    HTTP/1.1
    HTTP/1.1
    HTTP/1.1
    batch
    app
    app
    nginx
    Gaurun
    

    View Slide

13. • Push notification asynchronously kicked in-app
    events
    • comment, purchase, like, etc…
    • Push notification to many customers within 1~2
    hours on some campaign and event
    • Target number is over tens of millions
    Push infrastructure requirements @
    High concurrency and low latency are required！
    

    View Slide

14. • Push notification kicked in-app events such as
    comment, purchase, like, etc…
    • All logics were implemented in Mercari API
    • Mercari API is written in PHP (mod_php)
    • Push notification to many customers in large-
    scale campaign and event
    • PHP / Ruby script & Amazon SNS
    Past push infrastructure @
    

    View Slide

15. • Slow API response
    • Push was not kicked asynchronously in-app events
    • High network Latency
    • PHP processes frequently connected/closed APNs
    and GCM
    • Low throughput
    • It took a very long time to push notification to many
    users (more than a few hours)
    Problem of past push infrastructure @
    

    View Slide

16. 3 years ago
    

    View Slide

17. • Push runs synchronously!
    • API response had high network latency
    when in-app event is kicked
    • Response was returned to client after
    pushing notification to APNs or GCM
    3 years ago
    

    View Slide

18. 2 years ago
    

    View Slide

19. • Push runs asynchronously!
    • Job queue and worker were introduced
    • Q4M and php-parallel-prefork
    • Latency in API response was significantly
    reduced when in-app event is kicked
    • Throughput was significantly improved
    2 years ago
    

    View Slide

20. • Job queue: Q4M
    • Message queue for MySQL
    • https://q4m.github.io/
    • Job worker: php-parallel-prefork
    • Simple prefork server framework by PHP
    • https://github.com/travail/php-Parallel-Prefork
    Job queue and worker @
    

    View Slide

21. • Throughput is not enough
    • Preforking worker system based PHP is not
    fast and scalable for push notification
    • PHP is not good at concurrent processing
    • push notification processing requires high
    concurrency for achiving low network latency
    • APNs and GCM/FCM endpoints are far
    Problem still here
    

    View Slide

22. • Low latency
    • Push notification asynchronously kicked in-app events
    • comment, purchase, like, etc…
    • Push notification to many customers within 1~2
    hours in some campaign and event
    • Target number is over tens of millions
    Push infrastructure requirements @
    High concurrency and low latency are required！
    

    View Slide

23. Now
    APNs
    HTTP/2
    GCM
    HTTP/2
    HTTP/1.1
    HTTP/1.1
    HTTP/1.1
    HTTP/1.1
    batch
    app
    app
    nginx Gaurun
    

    View Slide

24. • Push infrastructure @ is built by
    • nginx: HTTP load balancer
    • Gaurun: HTTP/2 proxy for APNs and GCM/FCM
    Now
    

    View Slide

25. Gaurun
    

    View Slide

26. • Push notification server for APNs and GCM/FCM
    written in Go
    • https://github.com/mercari/gaurun
    • JSON based API via HTTP
    • Queueing & Pushing notifications to APNS
    and GCM/FCM asynchronously
    • Monitoring
    Gaurun
    

    View Slide

27. Send push notification to iPhone by Gaurun
    $ gaurun -c /etc/gaurun/gaurun.toml -p 1056 &
    $ curl \
    -X POST \
    -H "Content-Type: application/json" \
    “http://127.0.0.1:1056/push” \
    -d '{"notifications": [ {“token":["token-string"],"platform":
    1,"message":"Hello, iOS"} ] }'
    

    View Slide

28. • POST /push
    • Proxy push-notification requests to APNs and GCM/FCM
    • Response to client immediately and push notification asynchronously
    • GET /stat/app
    • Return operational stats by JSON
    • e.g. channel-usage, push-success/error number
    • GET /stat/go
    • Return Go stats by JSON
    • e.g. number of goroutine, memory usage in Go runtime
    • PUT /config/pushers
    • configure push-throughput dynamically
    Gaurun HTTP API
    

    View Slide

29. Why is Gaurun written in Go
    • High performance HTTP server
    • Go provides net/http package.
    • High concurrency
    • Go can handle too many goroutines
    simultaneously
    

    View Slide

30. Go provides net/http package. We can get enough performance
    by only this for introducing proxy server in Go.
    package main
    import (
    “fmt”
    “net/http”
    )
    func handler(w http.ResponseWriter, r *http.Request) {
    w.Header().Set(“Content-Type”, “text/plain”)
    fmt.Fprintf(w, “Hello, World!\n”)
    }
    func main() {
    http.HandleFunc(“/“, handler)
    http.ListenAndServe(“:8080”, nil)
    }
    

    View Slide

31. $ ab \
    -k \
    -c 100 \
    -n 100000 \
    "http://127.0.0.1:8080/" 2>&1 | \
    grep “Requests per second:”
    Requests per second: 56127.40 [#/sec] (mean)
    Simple benchmark on my MacBook Pro
    

    View Slide

32. Gaurun Internals
    

    View Slide

33. Gaurun internals
    • Gaurun has 3 components
    • HTTP API server
    • Proxy for APNs and GCM/FCM
    • Message queue and multiple workers
    • Based goroutine and channel
    

    View Slide

34. Push process in Gaurun
    

    View Slide

35. Job queue and workers by channel and goroutine
    ・channel is available as in-memory queue
    // channel based queue
    QueueNotification chan RequestGaurunNotification
    ・Start workers and initialize queue
    func StartPushWorkers(workerNum, queueNum int64) {
    QueueNotification = make(chan RequestGaurunNotification, queueNum)
    for i := int64(0); i < workerNum; i++ {
    go pushNotificationWorker()
    }
    }
    

    View Slide

36. Each worker has pusher pool
    • worker and pusher are goroutine
    • worker
    • Dequeue push-job from channel
    • Run pusher function by goroutine
    • pusher
    • Push notification to APNs and GCM/FCM
    

    View Slide

37. Pusher management by atomic package
    atomic.AddInt64(&pusherCount, 1)
    if atomic.LoadInt64(pusherCount) < pusherMax {
    go PusherFunc()
    } else {
    …
    func PusherFunc() {
    err := push()
    // error handling
    atomic.AddInt64(pusherCount, -1)
    }
    each worker know only number of active pusher.
    worker
    pusher
    

    View Slide

38. Push process in Gaurun
    

    View Slide

39. Connection handling for APNs and GCM/FCM
    • Gaurun uses http.Client in net/http package
    • http.Client reuses connection transparently
    • Behavior is configurable by http.Transport
    • MaxIdleConns
    • MaxIdleConnsPerHost
    • IdleConnTimeout
    • Gaurun provides parameters for configuring them
    

    View Slide

40. HTTP proxy server has various timeouts
    • connection timeout
    • read request timeout
    • write response timeout
    • keepalive timeout
    • Proxy is client also
    • proxy connection timeout
    • proxy write request timeout
    • proxy read response timeout
    • proxy keepalive timeout
    • etc…
    

    View Slide

41. Timeouts in net and net/http
    • net.Dial
    • Timeout
    • http.Transport
    • TLSHandshakeTimeout, IdleConnTimeout, ResponseHeaderTimeout,
    ExpectContinueTimeout
    • http.Client
    • TImeout
    • http.Server
    • ReadTimeout, ReadHeaderTimeout, WriteTimeout, IdleTimeout
    • Gopher should read this article
    • https://blog.cloudflare.com/the-complete-guide-to-golang-net-http-timeouts/
    

    View Slide

42. Monitoring Gaurun
    

    View Slide

43. Monitoring Gaurun status
    ・GET /stat/app
    {
    "queue_max": 12000000,
    "queue_usage": 515497,
    "pusher_max": 1152,
    "pusher_count": 640,
    "ios": {
    "push_success": 2465897,
    "push_error": 5704
    },
    "android": {
    "push_success": 1416295,
    "push_error": 4118
    }
    }
    

    View Slide

44. Monitoring Go status
    ・GET /stat/go
    $ curl -s http://127.0.0.1:1056/stat/go | jq ‘.goroutine_num’
    2326
    $ curl -s http://127.0.0.1:1056/stat/go | jq ‘.heap_objects’
    27428
    $ curl -s http://127.0.0.1:1056/stat/go | jq ‘.gc_num’
    44695
    $ …
    

    View Slide

45. ・number of goroutine
    ・memory allocation
    

    View Slide

46. push success rate
    

    View Slide

47. Threshold alerting with Mackerel
    

    View Slide

48. High Performance Gaurun
    

    View Slide

49. • Gaurun's behavior is configurable by TOML
    • Parameter tuning is required for high
    performant
    • The default configuration is very
    conservative and not high performant
    High Performance Gaurun
    

    View Slide

50. Configuration in TOML
    [core]
    port = “1056”
    workers = 8
    queues = 4192
    pusher_max = 32
    [android]
    apikey = “…”
    enabled = true
    keepalive_conns = 32
    [ios]
    pem_cert_path = “/path/to/cert.pem”
    pem_key_path = “/path/to/key.pem”
    enabled = true
    sandbox = false
    topic = “…”
    keepalive_conns = 32
    

    View Slide

51. • core.workers
    • number of goroutine dequeues push notification
    from channel-based queue
    • core.queues
    • size of channel based queue for push notification
    • core.pusher_max
    • number of goroutine per worker pushes notification
    to APNs and CGM/FCM
    Parameter tuning
    

    View Slide

52. • (ios|android).timeout
    • timeout for pushing notification to APNs or GCM/FCM
    • (ios|android).keepalive_conns
    • number of idling connection to APNs or GCM/FCM
    • (ios|android).keepalive_timeout
    • time for continuing keep-alive connection to APNs or GCM/
    FCM
    • (ios|android).retry_max
    • maximum retry count for pushing notification to APNs or GCM/FCM
    Parameter tuning
    

    View Slide

53. • Increase simulatanous number of push
    notification
    • core.workers x core.pusher_max
    • Increase core.queues
    • If channel is full, number of goroutine
    grows and Gaurun slows down.
    • Increase (ios|android).keepalive_conns
    Performance tuning
    But too large number is not good!
    

    View Slide

54. Configure push-throughput dynamically
    curl \
    -XPUT \
    “http://127.0.0.1:1056/config/pushers?max=32"
    configure core.pusher_max
    via HTTP dynamically
    Does not give too large number!
    

    View Slide

55. • POST /push can accepts multiple push
    notifications in single request
    • Limited by core.notification_max
    • default value is 100
    • Example request payload ->
    Bulk enqueue
    {
    "notifications" : [
    {
    "token" : ["xxx"],
    "platform" : 1,
    "message" : "Hello, iOS!"
    },
    {
    "token" : ["yyy"],
    "platform" : 2,
    "message" : "Hello, Android!"
    }
    ]
    }
    

    View Slide

56. • Device token is sometimes invalidated
    • Let’s remove invalidated token in database
    periodically
    • If number of invalidated device token is reduced,
    the time it takes to push notification will be
    shortened
    • We can know whether device token is invalidated
    by response from APNs and GCM/FCM
    Device token screening
    

    View Slide

57. Daily device token screening @
    S3 Batch
    Gaurun
    Output JSON log Upload Download
    MySQL
    Issue DELETE
    Parse JSON log
    

    View Slide

58. • Push notification has high network latency
    • High concurrency is required
    • Go is good choice for push notification
    server. Because,
    • Go provides useful net/http package
    • Go can handle too many goroutines
    simultaneously
    Conclusion
    

    View Slide

59. • Gaurun
    • https://github.com/mercari/gaurun
    • nginxとGoでつくるメルカリのプッシュ通知システム
    • http://tech.mercari.com/entry/2015/08/11/172206
    • ハイパフォーマンスGaurun
    〜メルカリの⼤大規模プッシュ配信を⽀支えるミドルウェ
    ア〜
    • http://tech.mercari.com/entry/2016/11/08/170343
    References
    

    View Slide