Pat Helland and me: How to build stateful distributed applications that can scale almost infinitely - Salesforce July 2018

In 2007, Pat Helland published “Life Beyond Distributed Transactions: An Apostate’s Opinion,” in which he conducts a thought experiment on how to design a distributed database that can scale almost infinitely. While the paper explicitly addresses distributed database design, Sean Allen shows that the ideas are far more widely applicable, particularly in scaling stateful applications. Sean explores some of Helland’s ideas through practical examples from his experience building data processing systems using tools like Apache Storm and, more recently, developing a stateful distributed stream processor at Wallaroo Labs.

Transcript

1. PAT HELLAND AND ME
   HOW TO BUILD STATEFUL DISTRIBUTED APPLICATIONS THAT CAN SCALE ALMOST INFINITELY
   

   View Slide

2. SEAN T. ALLEN
   VP OF ENGINEERING AT WALLAROO LABS
   MEMBER OF THE PONY CORE TEAM
   AUTHOR OF “STORM APPLIED”
   LOVER OF FRENCH STREET ART
   @SEANTALLEN
   @WALLAROOLABS
   @PONYLANG
   

   View Slide

3. PAT HELLAND
   AND ME
   

   View Slide

4. SOME AXIOMS…
   

   View Slide

5. TO SCALE INFINITELY,
   WE HAVE TO SCALE HORIZONTALLY
   

   View Slide

6. TO SCALE INFINITELY,
   WE MUST AVOID COORDINATION
   

   View Slide

7. DISTRIBUTED TRANSACTIONS ARE
   A FORM OF COORDINATION
   

   View Slide

8. THEREFORE…
   TO SCALE INFINITELY,
   WE CAN’T USE TRANSACTIONS
   

   View Slide

9. WHAT IS SCALING?
   

   View Slide

10. MORE AND
    MORE THINGS
    BUT, THEY DON’T GET
    BIGGER. THERE’S JUST…
    MORE OF THEM. LOTS MORE.
    

    View Slide

11. WE SCALE
    ENTITIES
    ENTITIES:
    LIVE ON A SINGLE MACHINE
    AND ARE MANIPULATED
    INDIVIDUALLY
    

    View Slide

12. WHAT IS AN ENTITY?
    

    View Slide

13. ENTITIES ARE BOUNDARIES OF
    ATOMICITY
    

    View Slide

14. Bob
    6 3
    5
    8
    Alice
    4 2
    
    7
    1
    

    View Slide

15. Bob
    6 3
    5
    8
    Alice
    4 2
    
    7
    1
    

    View Slide

16. Bob
    6 3
    5
    8
    Alice
    4
    7
    1
    2
    
    

    View Slide

17. Bob
    6 3
    5
    8
    Alice
    4 2
    
    7
    1
    

    View Slide

18. Bob
    6 3
    5
    8
    Alice
    4 2
    
    7
    1
    

    View Slide

19. DENORMALIZE..
    ALL THE THINGS!
    

    View Slide

20. TWO-LAYER ARCHITECTURE
    

    View Slide

21. scale-agnostic
    scale-aware
    API
    

    View Slide

22. scale-agnostic
    scale-aware
    API
    

    View Slide

23. scale-agnostic
    scale-aware
    API
    

    View Slide

24. scale-agnostic
    scale-aware
    API
    

    View Slide

25. scale-agnostic
    scale-aware
    API
    

    View Slide

26. scale-agnostic
    scale-aware
    API
    

    View Slide

27. TO SCALE INFINITELY,
    YOUR BUSINESS LOGIC HAS TO BE
    INDEPENDENT OF SCALE
    

    View Slide

28. WALLAROO
    SCALE INDEPENDENT
    COMPUTING FOR PYTHON
    

    View Slide

29. ENTITIES
    BUT WE CALL THEM…
    “STATE OBJECTS”
    

    View Slide

30. TWO-LAYER
    ARCHITECTURE
    BUT WE CALL IT…
    “SCALE INDEPENDENCE”
    

    View Slide

31. user supplied logic
    Wallaroo runtime
    Wallaroo API
    

    View Slide

32. user supplied logic
    Wallaroo runtime
    Wallaroo API
    

    View Slide

33. user supplied logic
    Wallaroo runtime
    Wallaroo API
    

    View Slide

34. user supplied logic
    Wallaroo runtime
    Wallaroo API
    

    View Slide

35. WALLAROO API
    MARKET SPREAD EXAMPLE
    

    View Slide

36. MARKET SPREAD
    REAL-TIME “SOMETHING AIN’T RIGHT” TRADE CHECKS
    Market Spread
    State
    Market
    Data
    Orders
    Update
    APPL
    Check
    MSFT
    Rejections
    

    View Slide

37. MARKET SPREAD
    TWO SOURCES OF DATA
    Market Spread
    State
    Market
    Data
    Orders
    Update
    APPL
    Check
    MSFT
    Rejections
    

    View Slide

38. MARKET SPREAD
    ONE SINK
    Market Spread
    State
    Market
    Data
    Orders
    Update
    APPL
    Check
    MSFT
    Rejections
    

    View Slide

39. MARKET SPREAD
    ORDER PIPELINE
    Market Spread
    State
    Market
    Data
    Orders
    Update
    APPL
    Check
    MSFT
    Rejections
    

    View Slide

40. MARKET SPREAD
    MARKET DATA PIPELINE
    Market Spread
    State
    Market
    Data
    Orders
    Update
    APPL
    Check
    MSFT
    Rejections
    

    View Slide

41. MARKET SPREAD
    APPLICATION DEFINITION
    

    View Slide

42. APPLICATION DEFINITION
    FLOW OF DATA FROM SOURCE TO SINK
    

    View Slide

43. TWO DATA PIPELINES
    ORDERS
    

    View Slide

44. TWO DATA PIPELINES
    MARKET DATA
    

    View Slide

45. DEFINE OUR SOURCES
    1 PER PIPELINE
    

    View Slide

46. DEFINE OUR OPERATIONS
    1 PER PIPELINE
    

    View Slide

47. DEFINE OUR OPERATIONS
    CHECK ORDER AGAINST SYMBOL DATA
    

    View Slide

48. DEFINE OUR OPERATIONS
    UPDATE SYMBOL DATA WITH LATEST MARKET DATA
    

    View Slide

49. DEFINE OUR SINKS
    1 PER PIPELINE
    

    View Slide

50. DEFINE OUR SINKS
    ORDERS PIPELINE MIGHT HAVE OUTPUT
    

    View Slide

51. DEFINE OUR SINKS
    MARKET DATA ONLY UPDATES SYMBOL DATA- NO OUTPUT
    

    View Slide

52. SCALE INDEPENDENT
    ONLY FLOW OF DATA AND OPERATIONS
    

    View Slide

53. USER SUPPLIED
    LOGIC
    

    View Slide

54. UPDATE MARKET DATA STATE COMPUTATION
    UPDATES SYMBOL DATA STATE
    

    View Slide

55. WALLAROO
    RUNTIME
    MESH NETWORK OF
    COOPERATING PROCESSES
    

    View Slide

56. STATE OBJECTS
    ONE BIG MAP?
    

    View Slide

57. STATE OBJECTS
    CONCEPTUALLY ITS LIKE A BIG MAP
    Market
    Data Update State
    

    View Slide

58. STATE OBJECTS
    WITH A KEY FOR EACH OBJECT
    APPL IBM
    MSFT AMZN
    INTC NVDA
    Market
    Data Update
    

    View Slide

59. STATE OBJECTS
    WHERE WE MAY FROM INCOMING DATA’S KEY
    APPL IBM
    MSFT AMZN
    INTC NVDA
    Market
    Data MSFT
    

    View Slide

60. STATE OBJECTS
    TO THE STATE OBJECT FOR THAT KEY
    APPL IBM
    MSFT AMZN
    INTC NVDA
    Market
    Data MSFT
    

    View Slide

61. HASH
    PARTITIONING
    DISTRIBUTING STATE
    OBJECTS ACROSS A
    CLUSTER
    

    View Slide

62. SINGLE WORKER
    ALL SYMBOLS TOGETHER
    APPL
    AMZN
    MSFT
    IBM
    

    View Slide

63. SINGLE WORKER
    ALL SYMBOLS TOGETHER
    APPL
    AMZN
    MSFT
    IBM
    

    View Slide

64. ADD ANOTHER WORKER
    STATE OBJECTS WILL BE REDISTRIBUTED ACROSS THE CLUSTER
    APPL
    AMZN
    MSFT
    IBM
    

    View Slide

65. ADD ANOTHER WORKER
    STATE OBJECTS WILL BE REDISTRIBUTED ACROSS THE CLUSTER
    APPL
    AMZN
    MSFT
    IBM
    

    View Slide

66. ADD ANOTHER WORKER
    STATE OBJECTS WILL BE REDISTRIBUTED ACROSS THE CLUSTER
    APPL
    AMZN
    MSFT
    IBM
    

    View Slide

67. ADD ANOTHER WORKER
    STATE OBJECTS WILL BE REDISTRIBUTED ACROSS THE CLUSTER
    APPL
    AMZN
    MSFT
    IBM
    

    View Slide

68. ADD ANOTHER WORKER
    STATE OBJECTS WILL BE REDISTRIBUTED ACROSS THE CLUSTER
    APPL
    AMZN
    MSFT
    IBM
    

    View Slide

69. ADD ANOTHER WORKER
    STATE OBJECTS WILL BE REDISTRIBUTED ACROSS THE CLUSTER
    APPL
    AMZN
    IBM
    MSFT
    

    View Slide

70. ADD ANOTHER WORKER
    STATE OBJECTS WILL BE REDISTRIBUTED ACROSS THE CLUSTER
    APPL
    AMZN
    IBM
    MSFT
    

    View Slide

71. DATA MODEL
    YOU CAN GET WITH THIS, OR
    YOU CAN GET WITH THAT.
    

    View Slide

72. A WALLAROO STATE OBJECT
    PLAIN OLD PYTHON
    

    View Slide

73. PERFORMANCE
    THERE’S NO REAL WORLD
    SCALING WITHOUT IT
    

    View Slide

74. LEARN MORE
    GITHUB.COM/SEANTALLEN/
    PAT-HELLAND-AND-ME
    

    View Slide