Serverless Is Ruby Future

#rubyrussia 2019

Transcript

1. SERVERLESS
   IS
   UBY
   FUTURE
   NIKOLAY SVERCHKOV
   @ssnickolay
   

   View Slide

2. View Slide

3. evilmartians.com
   New
   

   View Slide

4. ME
   Community
   Speech
   TWO
   OPTIONS
   

   View Slide

5. ME
   Community
   Speech
   VOTE
   
   26 %
   
   74 %
   

   View Slide

6. SERVERLESS AXIOMS
   Event-Driven Architecture
   

   View Slide

7. SERVERLESS AXIOMS
   Event-Driven Architecture
   Stateless
   

   View Slide

8. SERVERLESS AXIOMS
   Event-Driven Architecture
   Stateless
   ♾ Infinite Scaling
   

   View Slide

9. SERVERLESS AXIOMS
   Event-Driven Architecture
   Stateless
   ♾ Infinite Scaling
   Pay when the code is running
   

   View Slide

10. SERVERLESS AXIOMS
    Event-Driven Architecture
    Stateless
    ♾ Infinite Scaling
    Pay when the code is running
    

    View Slide

11. SERVERLESS
    UNDER
    THE
    MICROSCOPE
    

    View Slide

12. LOG
    $ serverless invoke -f index -l
    

    View Slide

13. LOG
    $ serverless invoke -f index -l
    START RequestId: Version: $LATEST
    END RequestId:
    REPORT RequestId:
    Duration: 1.48 ms Billed Duration: 100 ms
    Memory Size: 1024 MB Max Memory Used: 31 MB
    

    View Slide

14. LOG
    $ serverless invoke -f index -l
    START RequestId: Version: $LATEST
    END RequestId:
    REPORT RequestId:
    Duration: 1.48 ms Billed Duration: 100 ms
    Memory Size: 1024 MB Max Memory Used: 31 MB
    

    View Slide

15. SERVERLESS AXIOMS
    Event-Driven Architecture
    Stateless
    ♾ Infinite Scaling
    Pay when the code is running
    

    View Slide

16. LOG
    $ serverless invoke -f index -l
    START RequestId: Version: $LATEST
    END RequestId:
    REPORT RequestId:
    Duration: 1.48 ms Billed Duration: 100 ms
    Memory Size: 1024 MB Max Memory Used: 31 MB
    

    View Slide

17. LOG
    $ serverless invoke -f index -l
    START RequestId: Version: $LATEST
    END RequestId:
    REPORT RequestId:
    Duration: 1.48 ms Billed Duration: 100 ms
    Memory Size: 1024 MB Max Memory Used: 31 MB
    

    View Slide

18. WHAT ABOUT
    CPU?
    

    View Slide

19. AWS Lambda allocates CPU power
    proportional to the memory…
    (с) Amazon Docs
    

    View Slide

20. AWS Lambda allocates CPU power
    proportional to the memory…
    For example, if you allocate 256 MB
    memory, your Lambda function will receive
    twice the CPU share than if you allocated
    only 128 MB.
    (с) Amazon Docs
    

    View Slide

21. (CPU, CST)
    ""<=>
    MEMORY
    http://bit.ly/2UZ7xDL
    Completely Fair Scheduler (CFS)
    

    View Slide

22. https://aws.amazon.com/lambda/pricing/
    6$ "=> 28.8M/128mb!
    

    View Slide

23. COST
    SPEED ⏫
    https://amzn.to/2Ix2t2O
    

    View Slide

24. COST
    ""<=>
    LANGUAGE
    

    View Slide

25. $ wrk -t10 -c400 -d20s -R40000
    $ curl http://endpoint.url/[ruby/go]
    "=> {“success”: true}
    VS
    CHALLENGE:
    

    View Slide

26. 0
    250
    500
    750
    1000
    0 1 2 4 5
    MAX
    AVG
    AVG ~80ms
    31712 requests in 20.01s
    Requests/sec: 1584.83
    BATTLE (128 MB)
    

    View Slide

27. 0
    250
    500
    750
    1000
    0 1 2 4 5
    0
    250
    500
    750
    1000
    0 1 2 4 5
    MAX
    AVG
    AVG ~1.8ms
    AVG ~80ms
    31712 requests in 20.01s
    Requests/sec: 1584.83
    50347 requests in 20.00s
    Requests/sec: 2517.30
    BATTLE (128 MB)
    

    View Slide

28. 0
    250
    500
    750
    1000
    0 1 2 4 5
    0
    250
    500
    750
    1000
    0 1 2 4 5
    MAX
    AVG
    AVG ~1.8ms
    AVG ~80ms
    31712 requests in 20.01s
    Requests/sec: 1584.83
    50347 requests in 20.00s
    Requests/sec: 2517.30
    W
    IN
    BATTLE (128 MB)
    

    View Slide

29. 0
    250
    500
    750
    1000
    0 1 2 4 5
    0
    250
    500
    750
    1000
    0 1 2 4 5
    MAX
    AVG
    AVG ~1.8ms
    AVG ~80ms
    31712 requests in 20.01s
    Requests/sec: 1584.83
    50347 requests in 20.00s
    Requests/sec: 2517.30
    W
    IN
    *
    BATTLE (128 MB)
    

    View Slide

30. BATTLE (128 MB)
    AVG ~1.8ms
    AVG ~80ms
    “Billed Duration: 100 ms” (c) Lambda Docs
    

    View Slide

31. WHEN YOU REWRITE EVERYTHING ON GO
    EXPECTATION REALITY
    

    View Slide

32. View Slide

33. FREE!
    Free Enterprise
    

    View Slide

34. FREE!
    Free Enterprise
    LYING
    http://bit.ly/2lCt2dA
    

    View Slide

35. PREDICTABLY
    

    View Slide

36. PREDICTABLY
    

    View Slide

37. PREDICTABLY
    

    View Slide

38. http://serverlesscalc.com/
    $2400
    450M!!
    $918
    1GB
    

    View Slide

39. BATTLE (128 MB)
    AVG ~1.8ms
    AVG ~80ms
    ❓❓❓❓❓❓❓❓❓
    ❓❓❓❓❓❓❓❓❓❓
    ❓
    ❓
    ❓
    ❓
    ❓
    

    View Slide

40. http://rubyonjets.com/
    

    View Slide

41. RAILS
    ON
    SERVERLESS
    Tung Nguyen
    

    View Slide

42. PURE RUBY LAMBDA
    # handler.rb
    require 'json'
    def heartbeat(event:, context:)
    { statusCode: 200, body: JSON.generate(success: true) }
    end
    

    View Slide

43. BATTLE (128 MB)
    AVG ~39ms
    AVG ~80ms
    

    View Slide

44. COST
    ""<=>
    DEPENDENCIES
    

    View Slide

45. MINIMIZE THE COMPLEXITY OF
    YOUR DEPENDENCIES
    

    View Slide

46. MINIMIZE THE COMPLEXITY OF
    YOUR DEPENDENCIES
    

    View Slide

47. MINIMIZE THE COMPLEXITY OF
    YOUR DEPENDENCIES
    

    View Slide

48. NO RAILS
    GOOD!
    

    View Slide

49. CHEAPLY
    NICE
    NO RAILS
    GOOD!
    

    View Slide

50. CHEAPLY
    NICE
    EVENTS
    AWESOME
    NO RAILS
    GOOD!
    

    View Slide

51. CHEAPLY
    NICE
    EVENTS
    AWESOME
    NO RAILS
    GOOD!
    SCALING
    AMAZING
    

    View Slide

52. TALK ABOUT
    THE
    PROBLEMS
    TELL THE TRUTH
    Sponsored by @stacey_om
    

    View Slide

53. IS
    INFINITE
    SCALING
    ALWAYS
    GOOD?
    

    View Slide

54. # before
    $ puma -t 8:32 -w 1
    # after
    $ puma -t 8:32 -w 2
    HOW RUBY DEV SCALES APP
    

    View Slide

55. View Slide

56. SERVERLESS AXIOMS
    Event-Driven Architecture
    Stateless
    ♾ Infinite Scaling
    Pay when the code is running
    

    View Slide

57. When you write your Lambda function code, do not
    assume that AWS Lambda automatically reuses the
    execution context for subsequent function invocations….
    (c) AWS Docs
    AWS LAMBDA
    EXECUTION CONTEXT
    

    View Slide

58. EXECUTION CONTEXT
    # 1M symbols
    work = proc { 123_456_789 "** 123_456 }
    def heartbeat(event:, context:)
    { statusCode: 200, body: … }
    end
    

    View Slide

59. EXECUTION CONTEXT
    # 1M symbols
    work = proc { 123_456_789 !** 123_456 }
    def heartbeat(event:, context:)
    value = work.() # not cached
    {
    statusCode: 200,
    body: JSON.generate(value: value)
    }
    end
    

    View Slide

60. EXECUTION CONTEXT
    # 1M symbols
    work = proc { 123_456_789 !** 123_456 }
    value = work.() # cached
    def heartbeat(event:, context:)
    {
    statusCode: 200,
    body: JSON.generate(value: value)
    }
    end
    

    View Slide

61. CONNECTION POOL
    pool = DB"::ConnectionPool.new()
    def heartbeat(event:, context:)
    connection = pool.get()
    # use connection
    connection.release()
    end
    

    View Slide

62. https://stackoverflow.com/q/54110697
    

    View Slide

63. SOLUTIONS*
    CLASSIC ($)
    

    View Slide

64. SOLUTIONS*
    CLASSIC ($) MODERN ($$)
    Serverless
    

    View Slide

65. SCALE
    ⬇
    DOWSCALE
    ⬇
    SCALE
    ⬇
    …
    

    View Slide

66. BURSTY TRAFFIC
    BURSTY LOAD
    BURSTY ???
    
    
    
    

    View Slide

67. EXAMPLE
    ML
    

    View Slide

68. EXAMPLE: ETL
    

    View Slide

69. EXAMPLE: ETL
    

    View Slide

70. EXAMPLE: BLOG
    

    View Slide

71. GOTO 0;
    SERVERLESS
    IS RUBY
    FUTURE
    

    View Slide

72. GOTO 0;
    SERVERLESS
    IS RUBY
    FUTURE
    

    View Slide

73. DREAMS
    ABOUT
    SERVERLESS
    

    View Slide

74. DREAMS
    ABOUT
    SERVERLESS
    Incredible
    Processing Power
    

    View Slide

75. SERVERLESS
    Massive User Job
    

    View Slide

76. SERVERLESS
    Backends
    Massive User Job
    DIFFERENT BACKENDS
    

    View Slide

77. SERVERLESS
    Backends
    Massive User Job
    10_000.times do
    Thread.new(job)
    end
    

    View Slide

78. SERVERLESS
    Backends
    Burst-Parallel Job
    

    View Slide

79. SERVERLESS
    Burst-Parallel Job
    Video Encoding: VP8
    

    View Slide

80. Burst-Parallel Job
    Video Encoding: Classic
    encode([, , …, ]) "-> keyframe + interframe(2:n)
    

    View Slide

81. Burst-Parallel Job
    Video Encoding: Existing
    encode([, , …, ]) "-> keyframe + interframe(2:n)
    1MB +25kb * n
    

    View Slide

82. Burst-Parallel Job
    Video Encoding: Existing
    encode([, , …, ]) "-> keyframe + interframe(2:n)
    decode(keyframe + interframe[2:n]) "-> [, , …, ]
    1MB +25kb * n
    

    View Slide

83. Burst-Parallel Job
    Video Encoding: Existing
    encode(i[1:10]) "-> keyframe1 + interframe(2:10)
    encode(i[11:20]) "-> keyframe2 + interframe(12:20)
    encode(i[21:30]) "-> keyframe3 + interframe(22:30)
    1MB * N
    

    View Slide

84. VIDEO ENCODING: ALFALA
    http://bit.ly/2lZFrbO
    

    View Slide

85. Burst-Parallel Job
    preprocessor(compiler
    (assembler(linker
    (archiver(indexer
    (strip(source_file)…)
    Software Compilation: 7 Steps
    

    View Slide

86. Burst-Parallel Job
    Software Compilation
    

    View Slide

87. Burst-Parallel Job
    Software Compilation
    Video Encoding
    

    View Slide

88. DIRECTED ACYCLIC GRAPH
    BRO NOT A BRO
    

    View Slide

89. DIRECTED ACYCLIC GRAPH
    Problem: Different Nodes
    g
    f
    h
    m
    k
    l
    p
    r
    

    View Slide

90. DAG "=> FUNCTIONS
    preprocessor(compiler
    (assembler(linker
    (archiver(indexer
    (strip(source_file)…)
    

    View Slide

91. f(g(h(p(…)…)
    

    View Slide

92. f(g(i)) !== λ(f, λ(g, i))
    THE NEW ABSTRACTION: λ
    

    View Slide

93. f(g(i)) !== λ(f, λ(g, i))
    f = proc { |a| a + 1 }
    g = proc { |a| a + 2 }
    f.call(g.call(0))
    # "=> 3
    THE NEW ABSTRACTION: λ
    

    View Slide

94. f(g(i)) !== λ(f, λ(g, i))
    f = proc { |a| a + 1 }
    g = proc { |a| a + 2 }
    f.call(g.call(0))
    # !=> 3
    λ = proc { |f, i| f.(i) }
    λ.call(f, λ.call(g, 0))
    # "=> 3
    THE NEW ABSTRACTION: λ
    

    View Slide

95. f(g(h(p(…)…) !==
    λ(f, λ(g, λ(h, λ(p(…)…)
    

    View Slide

96. f(g(h(p(…)…) !==
    λ(f, λ(g, λ(h, λ(p(…)…)
    λ(ƒ,arg) !=> arg
    WHERE
    

    View Slide

97. λ(ƒ,arg) !=> arg
    

    View Slide

98. λ(ƒ,arg) !=> arg
    λ(ƒ,arg) "=> Θ,arg
    

    View Slide

99. λ(ƒ,arg) !=> arg
    λ(ƒ,arg) !=> Θ,arg
    λ(T) "=> T
    

    View Slide

100. DIRECTED ACYCLIC GRAPH
     T2
     T1
     T3
     T5
     T4
     T6
     T7
     T8
     λ(T) "=> T, T{i} "<- T
     

     View Slide

101. SERVERLESS
     Backends
     Burst-Parallel Job
     

     View Slide

102. SERVERLESS
     Backends
     Burst-Parallel Job
     λ(T) "=> T
     

     View Slide

103. SERVERLESS
     Backends
     Burst-Parallel Job
     λ(T) "=> T
     λ
     

     View Slide

104. SERVERLESS
     Backends
     Burst-Parallel Job
     λ(T) "=> T
     λ
     IR
     T
     

     View Slide

105. SERVERLESS
     Backends
     Burst-Parallel Job
     IR
     λ(T) "=> T
     T
     λ
     STATIC
     

     View Slide

106. SERVERLESS
     Backends
     Burst-Parallel Job
     λ
     IR
     T
     Frontends
     T generator
     λ(T) "=> T
     

     View Slide

107. SERVERLESS
     Backends
     Burst-Parallel Job
     λ
     IR
     T
     Frontends
     T generator
     λ(T) "=> T
     

     View Slide

108. Burst-Parallel Job
     IR
     Frontends
     Coordinator
     HTTP
     λ
     

     View Slide

109. SERVERLESS
     Backend
     Burst-Parallel Job
     IR
     Frontends
     Coordinator
     STATELESS
     HTTP
     

     View Slide

110. Burst-Parallel Job
     IR
     Frontends
     Coordinator
     S3
     HTTP
     λ
     

     View Slide

111. THE FUTURE
     IS HERE!
     

     View Slide

112. gg
     Backends
     Burst-Parallel Job
     gg IR
     Frontends
     gg
     Coordinator
     C++
     HTTP
     ProtoBuf
     S3
     Stanford Systems
     and
     Networking Research
     

     View Slide

113. Fib(4)
     fib(4)
     fib(3) fib(2)
     fib(2) fib(1)
     fib(1) fib(0)
     fib(1) fib(0)
     +
     +
     +
     +
     

     View Slide

114. Fib(4)
     fib(4)
     fib(3) fib(2)
     fib(2) fib(1)
     fib(1) fib(0)
     fib(1) fib(0)
     +
     +
     +
     +
     Hash#fib1
     Hash#fib2
     

     View Slide

115. Нихуя не понял
     Но очень интересно!
     чего
     

     View Slide

116. gg FRAMEWORK
     https://stanford.io/2mtmiiH
     

     View Slide

117. DEMO
     

     View Slide

118. DEMO
     http://bit.ly/2m0G3Oo
     https://asciinema.org/a/270154
     

     View Slide

119. RESULTS
     gg
     real
     user
     sys
     1m17.670s
     1m49.040s
     0m7.820s
     real
     user
     sys
     0m43.354s
     0m3.830s
     0m5.490s
     real
     user
     sys
     0m29.376s
     0m0.950s
     0m1.400s
     

     View Slide

120. RESULTS
     gg
     real
     user
     sys
     1m17.670s
     1m49.040s
     0m7.820s
     real
     user
     sys
     0m43.354s
     0m3.830s
     0m5.490s
     real
     user
     sys
     0m29.376s
     0m0.950s
     0m1.400s
     

     View Slide

121. RESULTS
     gg
     Total: 1m56s Total: 0m38s
     ~X3 faster
     

     View Slide

122. COMPARISON OF COLD-CACHE BUILD TIMES
     https://cs.stanford.edu/~matei/papers/2019/usenix_atc_gg.pdf
     

     View Slide

123. DARK
     MAGIC!
     

     View Slide

124. https://github.com/StanfordSNR/gg
     

     View Slide

125. GOTO 0;
     SERVERLESS
     IS RUBY
     FUTURE
     

     View Slide

126. “As a computing substrate, we suspect cloud
     functions are in a similar position to Graphics
     Processing Units in the 2000s.
     At the time, GPUs were designed solely for 3D
     graphics, but the community gradually
     recognized that they had become programmable
     enough to execute some parallel algorithms
     unrelated to graphics.” (c)
     

     View Slide

127. GOTO 0;
     SERVERLESS
     IS RUBY
     FUTURE
     

     View Slide

128. SORBET?
     

     View Slide

129. IT saṃsāra
     Java
     Erlang
     C
     C++
     "=> Scala
     "=> Elixir
     "=> Go
     "=> Rust
     

     View Slide

130. RUBY
     SELF-IDENTIFICATION
     

     View Slide

131. RUBY IS
     Fast
     Coding
     

     View Slide

132. RUBY IS
     Startups
     (Rails)
     Fast
     Coding
     

     View Slide

133. RUBY IS
     Startups
     (Rails)
     GIL
     Sensitive
     Fast
     Coding
     

     View Slide

134. RUBY IS
     Startups
     (Rails)
     オブジェクト指向スクリプト⾔言語 Ruby
     GIL
     Sensitive
     Fast
     Coding
     

     View Slide

135. RUBY IS
     Startups
     (Rails)
     GIL
     Sensitive
     The Object-oriented
     Scripting Language Ruby
     Fast
     Coding
     

     View Slide

136. RUBY IS
     Startups
     (Rails)
     GIL
     Sensitive
     The Object-oriented
     Scripting Language Ruby
     Fast
     Coding
     

     View Slide

137. RUBY IS
     Startups
     (Rails)
     GIL
     Sensitive
     The Object-oriented
     Scripting Language Ruby
     Fast
     Coding
     

     View Slide

138. RUBY IS
     Startups
     (Rails)
     GIL
     Sensitive
     The Object-oriented
     Scripting Language Ruby
     Fast
     Coding
     

     View Slide

139. RUBY IS
     Startups
     (Rails)
     GIL
     Sensitive
     The Object-oriented
     Scripting Language Ruby
     Fast
     Coding
     

     View Slide

140. 60 SECNDS
     OF SUMMARY
     

     View Slide

141. @ssnickolay
     Nikolay Sverchkov
     @ssnickolay
     THANK YOU!
     @evilmartians
     http://evl.ms/blog
     http://evl.ms/telegram
     

     View Slide