Fast & Functional (Scala Matsuri, JP)

In this talk we are going to create a functional and blazingly fast microservice. We will use functional programming abstractions to quickly mix & match different HTTP libraries, state implementations and concurrency configurations. Each step will be followed by a performance analysis using different tools from JVM toolbox. This talk is for you if you want to see how cats IO monad, async-profiler, flame graphs and wrk are used together to create microseconds-fast Scala service for YouTube videos statistics.

Transcript

1. @miciek
   AST &
   UNCTIONAL
   F
   F
   Room A, 14:00
   In this talk we are going to create
   a functional and blazingly fast
   microservice in Scala. We will use
   f u n c t i o n a l p r o g r a m m i n g
   abstractions to quickly mix &
   match different HTTP libraries,
   state implementations and
   concurrency configurations. Each
   step will be followed by a
   performance analysis using
   different tools from JVM toolbox.
   

   View Slide

2. @miciek
   λ
   Michał Płachta
   www.michalplachta.com
   @miciek
   AST &
   UNCTIONAL
   F
   F
   

   View Slide

3. @miciek
   Hello, I’d like to hear a
   functional programming joke
   関数型プログラミングの冗談を聴いたことがある
   

   View Slide

4. @miciek
   OK. Why did the chicken
   cross the road?
   functional programming joke
   どうしてニワトリは道路路を渡るの？
   

   View Slide

5. @miciek
   cross the road?
   I don’t know. Why?
   分かりません。どうしてでしょう?
   

   View Slide

6. @miciek
   It’s very simple really.
   I don’t know. Why?
   それは本当にとても簡単です。
   

   View Slide

7. @miciek
   The best way to understand
   it comes from the idea of
   contravariant functor.
   It’s very simple really.
   理理解するための最良の⽅方法は、
   反変ファンクターの考えから来ています。
   

   View Slide

8. @miciek
   As you probably know, a functor is just a
   mapping between categories that respects the
   category structure; a homomorphism, if you
   will.
   contravariant functor.
   ファンクターは圏の構造を保った写像で、
   準同型写像とも⾔言われます
   

   View Slide

9. @miciek
   Now, some constructions that we want to
   express look like functors but in some sense
   “turn morphisms around”. We call these
   contravariant functors.
   will.
   さて、私たちが表現したいいくつかの構造はファンクタのように⾒見見えます
   が、ある意味では「変形を逆転させる」のです。私たちはこれらの反変ファ
   ンクタと呼びます。
   

   View Slide

10. @miciek
    As you have probably noticed, we can indeed define a
    contravariant functor as simply a covariant functor on opposite
    categories, invoking the obvious correspondence. Of course you
    have already noticed the obvious connection this has with the
    category theoretic generalization of topological presheaves — a
    presheaf on a topology is just a contravariant functor from the…
    私たちは反変ファンクターを単に逆圏の共変ファンクターとして定義しなお
    すことができます。もちろん、これが圏論的なトポロジカルなプリシーフの
    ⼀一般化に関連していることにすでにお分かりですね (以下延々
    

    View Slide

11. @miciek
    私は Haskell の冗談を聞きたいです。
    

    View Slide

12. @miciek
    AST &
    UNCTIONAL
    F
    F
    

    View Slide

13. @miciek
    UNCTIONAL
    F
    very modular
    plug & play
    using functional mechanics
    quickly replacing modules
    without affecting the logic
    simple real-world
    service in Scala
    モジュール性、プラグアンドプレイ
    

    View Slide

14. @miciek
    F
    1st meaning:
    PERFORMANCE
    AST
    Fast の第⼀一義は、実⾏行行速度
    

    View Slide

15. @miciek
    F
    2nd meaning:
    DEV SPEED
    AST
    very modular
    plug & play
    Fast の第⼆二義は、開発速度
    

    View Slide

16. @miciek
    In this talk…
    AST &
    UNCTIONAL
    F
    F
    modularity (the λ way!)
    plug & play
    simple real-world
    service in Scala
    

    View Slide

17. @miciek
    Influencer Stats
    インフルエンサーの統計情報
    

    View Slide

18. @miciek
    We need a marketing strategy
    for our product!
    Marketing
    マーケティング
    

    View Slide

19. @miciek
    Remember when…
    Old Ways Don’t Work
    …people watched TV
    みんながテレビを⾒見見ていた時代の古いやり⽅方は通⽤用しない
    

    View Slide

20. @miciek
    Influencers
    22 237 likes 2302 likes, 70 retweets
    201 232 views, 23987 likes 502 123 views, 2012 likes
    …and more
    今のインフルエンサー
    

    View Slide

21. @miciek
    Influencers items aggregation
    1 201 882 views, 423 987 likes
    …and more
    All tweets, youtube videos, etc for one product
    ⼀一つのプロダクトに関して全てのアイテムを集約
    

    View Slide

22. @miciek
    Influencer Stats service
    collection of videos
    how many total views
    how many total likes
    動画、PV、like 数など
    

    View Slide

23. @miciek
    Requirement #1
    Collections
    ability to
    add & update
    Collections of videos
    …
    COLLECTION 1
    …
    COLLECTION 2
    要件1: 動画のコレクションを作成、更更新できること
    

    View Slide

24. @miciek
    Requirement #2
    Fetch YouTube Stats
    ability to
    fetch individual video
    stats through
    YouTube API
    YouTube
    API
    20 views, 2 likes
    要件2: 個々の動画に関して統計を取得できること
    

    View Slide

25. @miciek
    Requirement #3
    Calculate Stats
    ability to
    calculate & return
    Collection statistics
    …
    1 201 882 views, 423 987 likes
    要件3: コレクション全体の統計情報を集計できること
    

    View Slide

26. @miciek
    Non-functional
    requirements
    ✓HTTP Web Server
    ✓HTTP Client
    ✓Collection Storage
    ✓Logging
    ⾮非機能要件
    

    View Slide

27. @miciek
    PURE Side-effect’y
    f
    input value output value
    f
    input value
    
    
    hard to keep separate
    hard to test
    easy to keep separate
    easy to test
    副作⽤用はテストしづらい
    

    View Slide

28. @miciek
    PURE
    ✓Save & Load Collections
    ✓Fetch YouTube Stats
    ✓HTTP Web Server
    ✓HTTP Client
    ✓Collection storage
    ✓Logging
    Side-effect’y
    ✓Calculate Stats
    f
    video #1 stats
    video #2 stats
    aggregated stats
    とは⾔言っても機能の多くは副作⽤用を伴う
    

    View Slide

29. @miciek
    PURE
    ✓Save & Load Collections
    ✓Fetch YouTube Stats
    ✓HTTP Web Server
    ✓HTTP Client
    ✓Collection storage
    ✓Logging
    Side-effect’y
    ✓Calculate Stats
    ULTIMATE GOAL
    move as much as we
    can to the PURE side
    できる限り多くの機能を pure 側へと移⾏行行させる
    

    View Slide

30. @miciek
    ULTIMATE GOAL
    move as much as we
    can to the PURE side
    できる限り多くの機能を pure 側へと移⾏行行させる
    

    View Slide

31. @miciek
    ULTIMATE GOAL
    move as much as we
    can to the PURE side
    to test in isolation
    to reason in isolation
    to easily swap components
    (because they are isolated)
    隔離された単体でテストしたり、論理理的に振る舞いを分析したり、
    コンポーネントを差し替えるのが容易易な状態になるのが究極の⽬目標
    

    View Slide

32. @miciek
    What happens next…
    intro
    code
    Java code
    performance
    tests
    t
    we are
    here
    

    View Slide

33. @miciek
    Modular programming
    using algebras
    代数を利利⽤用したモジュラープログラミング
    

    View Slide

34. @miciek
    Model
    e.g.
    -a YouTube video
    -instagram post
    -tweet
    Model の実装を考える
    

    View Slide

35. @miciek
    Pure logic
    f
    video #1 stats
    video #2 stats
    aggregated stats
    純粋なロジック
    

    View Slide

36. @miciek
    Pure logic
    

    View Slide

37. @miciek
    Pure logic
    

    View Slide

38. @miciek
    Pure logic
    …
    1 201 882 views, 423 987 likes
    easy to test
    easy to reason about
    PURE WIN!
    

    View Slide

39. @miciek
    Can it be as easy as this?
    f
    input value output value
    NOPE
    では次のような関数は？
    

    View Slide

40. @miciek
    Can it be as easy as this?
    1. get collection collectionId
    2. get videoIds from it
    3. for each videoId, call YouTube API
    4.call our pure calculate function
    (and log some things in between)
    
    
    
    f
    input value output value
    f
    input value output value
    いくつか問題があります。
    

    View Slide

41. @miciek
    Can it be as easy as this?
    1. get collection collectionId
    2. get videoIds from it
    3. for each videoId, call YouTube API
    4. call our calculate function
    NOPE
    because side effects
    
    
    
    
    副作⽤用があるからです。
    

    View Slide

42. @miciek
    Side-effect’y stuff
    

    View Slide

43. @miciek
    Side-effect’y stuff
    1. get collection collectionId
    2. get videoIds from it
    3. for each videoId, call YouTube API
    4. call our pure calculate function
    

    View Slide

44. @miciek
    Side-effect’y stuff
    1. get collection collectionId
    2. get videoIds from it
    3. for each videoId, call YouTube API
    4. call our pure calculate function
    

    View Slide

45. @miciek
    Specific effect type
    特定の作⽤用を表す型が使われています。
    

    View Slide

46. @miciek
    Too powerful effect type :(
    作⽤用型が強すぎます。
    

    View Slide

47. @miciek
    Side-effect’y stuff
    1. get collection collectionId
    2. get videoIds from it
    3. for each videoId, call YouTube API
    4. call our pure calculate function
    F doesn’t do anything
    F is chosen by the user
    モナドの F とだけ弱く指定します。
    

    View Slide

48. @miciek
    synchronous
    asynchronous
    IO effect type
    F[_]
    

    View Slide

49. @miciek
    Constrained?
    A doesn’t do anything
    A is chosen by the user
    

    View Slide

50. @miciek
    Constrained?
    A doesn’t do anything
    A is chosen by the user
    

    View Slide

51. @miciek
    We can’t constrain ourselves
    What we have
    What we want
    F doesn’t do anything
    F is chosen by the user
    制限を後付けすることはできません。
    

    View Slide

52. @miciek
    Constraints Liberate
    F doesn’t do anything
    F is chosen by the user
    Constraints Liberate
    Liberties Constrain
    制限は解放し、⾃自由は制限する
    

    View Slide

53. @miciek
    Get collection collectionId
    

    View Slide

54. @miciek
    Get collection collectionId
    just an algebra
    

    View Slide

55. @miciek
    Get videoIds from it
    

    View Slide

56. @miciek
    For each videoId call YouTube API
    

    View Slide

57. @miciek
    For each videoId call YouTube API
    

    View Slide

58. @miciek
    Call our pure calculate function
    

    View Slide

59. @miciek
    (and log some things in between)
    

    View Slide

60. @miciek
    (and log some things in between)
    

    View Slide

61. @miciek
    Signature is enough
    

    View Slide

62. @miciek
    

    View Slide

63. @miciek
    PURE
    ✓Collections interpreter
    ✓HTTP Web Server
    ✓HTTP Client interpreter
    ✓Collection storage interpreter
    ✓Logging interpreter
    Side-effect’y
    ✓Calculate Stats
    ✓Collections algebra
    ✓HTTP Client algebra
    ✓Logging algebra
    ✓getStats algorithm on F
    

    View Slide

64. @miciek
    algebras
    ✓Collections interpreter
    ✓HTTP Client interpreter
    ✓Collection storage interpreter
    ✓Logging interpreter
    interpreters
    ✓Collections algebra
    ✓HTTP Client algebra
    ✓Logging algebra
    代数 | インタプリタ
    

    View Slide

65. @miciek
    algebra
    interpreter
    

    View Slide

66. @miciek
    What happens next…
    intro
    code
    Java code
    performance
    tests
    t
    we are
    here
    

    View Slide

67. @miciek
    algebra interpreters
    - AkkaHttp
    - Hammock (Apache)
    - Http4s
    - InMem LinkedList
    - InMem TrieMap
    - Log All
    - Max 1k/sec (dropping rest)
    plug & play
    代数 | インタプリタ
    

    View Slide

68. @miciek
    Make it fast
    ⾼高速化
    

    View Slide

69. @miciek
    @miciek
    Warning
    Do not perform any of the following stunts
    on PRODUCTION.
    注意! プロダクションで⾏行行ってはいけません
    

    View Slide

70. @miciek
    wrk -t1 -c16 -d30s --latency URL
    We will use wrk
    1 thread
    16
    connections
    30 sec
    duration
    ベンチマークツール: wrk
    

    View Slide

71. @miciek
    We will use async-profiler
    jps
    cd async-profiler
    ./profiler.sh -d 10 -f
    https://github.com/jvm-profiling-tools/async-profiler
    プロファイラ: async-profiler
    

    View Slide

72. @miciek
    We will use flame graphs
    http://www.brendangregg.com/flamegraphs.html
    read
    from
    bottom
    

    View Slide

73. @miciek
    Version 1
    # requests
    99%
    latency
    avg
    latency
    req/s
    Version 1 ? ?ms ?ms ?
    -akka-http (client + server)
    -log all things
    -in mem linked list state
    

    View Slide

74. @miciek
    Version 1 Results
    # requests
    99%
    latency
    avg
    latency
    req/s
    Version 1 8510 82.41ms 56.47ms 283.12
    

    View Slide

75. @miciek
    Version 1 Flame Graph
    

    View Slide

76. @miciek
    -don’t change the logic
    -just change the Logger interpreter
    -dropping logs if the rate is > 1k per second
    Version 2: Dropping logs
    

    View Slide

77. @miciek
    Version 2 Results
    # requests
    99%
    latency
    avg
    latency
    req/s
    Version 1 8510 82.41ms 56.47ms 283.12
    Version 2 89 757 62.53ms 6.94ms 2986.26
    DroppingLogger
    

    View Slide

78. @miciek
    Version 2 Flame Graph
    

    View Slide

79. @miciek
    -don’t change the logic
    -change the VideoClient type class instance
    Version 3: Change the Video Client
    クライアントを代えてみる
    

    View Slide

80. @miciek
    Version 3 Results
    # requests
    99%
    latency
    avg
    latency
    req/s
    Version 1 8510 82.41ms 56.47ms 283.12
    Version 2 89 757 62.53ms 6.94ms 2986.26
    Version 3 96 564 13.05ms 5.06ms 3218.48
    DroppingLogger
    Hammock (Apache)
    

    View Slide

81. @miciek
    Version 3 Flame Graph
    

    View Slide

82. @miciek
    -don’t change the logic
    -cache results and call logic in separate thread
    Version 4: Statistics Caching
    

    View Slide

83. @miciek
    Version 4 Results
    # requests
    99%
    latency
    avg
    latency
    req/s
    Version 1 8510 82.41ms 56.47ms 283.12
    Version 2 89 757 62.53ms 6.94ms 2986.26
    Version 3 96 564 13.05ms 5.06ms 3218.48
    Version 4 933 303 62.86ms 2.87ms 31 081.82
    DroppingLogger
    Hammock (Apache)
    Stats Caching
    

    View Slide

84. @miciek
    Version 4 Flame Graph
    

    View Slide

85. @miciek
    Version 4 Flame Graph
    

    View Slide

86. @miciek
    -don’t change the logic
    -change the underlying server from Akka HTTP
    to Http4s
    Version 5: Change the Server
    サーバーを代えてみる
    

    View Slide

87. @miciek
    Version 5 Results
    # requests
    99%
    latency
    avg
    latency
    req/s
    Version 1 8510 82.41ms 56.47ms 283.12
    Version 2 89 757 62.53ms 6.94ms 2986.26
    Version 3 96 564 13.05ms 5.06ms 3218.48
    Version 4 933 303 62.86ms 2.87ms 31 081.82
    Version 5 1 158 664 9.05ms 555.80µs 38 596.24
    DroppingLogger
    Hammock (Apache)
    Stats Caching
    Http4s Server
    

    View Slide

88. @miciek
    Version 5 Flame Graph
    

    View Slide

89. @miciek
    Inventions That
    Already Exist
    既にあるものを発明する
    

    View Slide

90. @miciek
    Introducing…
    

    View Slide

91. @miciek
    

    View Slide

92. @miciek
    Introducing…
    Cheaper taxi that drives on fixed routes!
    固定ルートの格安タクシーがあったら...
    

    View Slide

93. @miciek
    This is a BUS. You invented a BUS.
    Cheaper taxi that drives on fixed routes!
    それはバスです。あなたはバスを発⾒見見したのです。
    

    View Slide

94. @miciek
    Introducing…
    Algebras & Interpreters
    

    View Slide

95. @miciek
    Algebras & Interpreters
    This is an INTERFACE. You invented an INTERFACE.
    

    View Slide

96. @miciek
    Scala FP
    

    View Slide

97. @miciek
    Java
    

    View Slide

98. @miciek
    Interfaces
    

    View Slide

99. @miciek
    Interfaces
    difference:
    we need to choose the effect type
    

    View Slide

100. @miciek
     Return type
     difference:
     we need to choose the effect type
     which is very powerful
     

     View Slide

101. @miciek
     Algebras
     

     View Slide

102. @miciek
     Constrained effect type
     

     View Slide

103. @miciek
     Dependencies
     

     View Slide

104. @miciek
     Dependencies
     

     View Slide

105. @miciek
     Algebras & Interpreters
     Not an INTERFACE.
     major difference:
     we can abstract over the effect type
     and constrain ourselves
     using F instead of CompletableFuture…
     

     View Slide

106. @miciek
     Conclusions
     結論
     

     View Slide

107. @miciek
     -changed infrastructure details
     -…but our core logic (and tests) didn’t need to
     change!
     What have we done?
     plug & play
     quickly replacing non-core-logic modules
     インフラの詳細を変えましたが、コアロジックは変えていません!
     

     View Slide

108. @miciek
     Modularity / Separation of concerns
     if we can..
     replace the server implementation
     replace the state implementation
     replace the logging implementation
     replace the video client implementation
     add caching
     without
     changing the core logic (& tests)
     PURE WIN!
     コアロジックを変えることなく実装の差し替えに成功しました
     

     View Slide

109. @miciek
     PURE
     ✓Collections interpreter
     ✓HTTP Web Server
     ✓HTTP Client interpreter
     ✓Collection storage interpreter
     ✓Logging interpreter
     Side-effect’y
     ✓Calculate Stats
     ✓Collections algebra
     ✓HTTP Client algebra
     ✓Logging algebra
     ✓getStats algorithm on F
     unit-tested integration-tested
     very constrained
     just F[_]
     no constraints
     Future, IO, …
     

     View Slide

110. @miciek
     Interpreters integration test
     https://www.testcontainers.org/
     

     View Slide

111. @miciek
     Interpreters integration test
     interpreter
     response
     https://www.testcontainers.org/
     

     View Slide

112. @miciek
     You learned…
     AST &
     UNCTIONAL
     F
     F
     modularity (the λ way!)
     plug & play
     simple real-world
     service in Scala
     モジュール性とプラグアンドプレイ
     

     View Slide

113. @miciek
     Learn More
     https://github.com/miciek/influencer-stats
     

     View Slide

114. @miciek
     λ
     Michał Płachta
     www.michalplachta.com
     @miciek
     AST &
     UNCTIONAL
     F
     F
     Thanks!
     

     View Slide