Scalable Atomic Visibility with RAMP Transactions

Slides for "Scalable Atomic Visibility with RAMP Transactions" by Bailis et al., appearing in SIGMOD 2014

This deck also contains a proposal for implementation in Cassandra. If you're interested in implementing RAMP in your own system, don't hesitate to get in touch: pbailis at cs.berkeley.edu or @pbailis

Paper: http://www.bailis.org/papers/ramp-sigmod2014.pdf
Blog post intro: http://www.bailis.org/blog/scalable-atomic-visibility-with-ramp-transactions/
Source code from paper and "executable pseudocode" in Python: https://github.com/pbailis/ramp-sigmod2014-code

Transcript

1. SCALABLE
   ATOMIC VISIBILITY
   WITH
   RAMP TRANSACTIONS
   Peter Bailis, Alan Fekete, Ali Ghodsi,
   Joseph M. Hellerstein, Ion Stoica
   UC Berkeley and University of Sydney
   
   Overview deck with Cassandra discussion
   
   @pbailis
   

   View Slide

2. NOSQL
   

   View Slide

3. NO SQL
   NOSQL
   

   View Slide

4. NO SQL
   DIDN’T WANT SQL
   NOSQL
   

   View Slide

5. NO SQL
   DIDN’T WANT SERIALIZABILITY
   NOSQL
   

   View Slide

6. POOR PERFORMANCE
   NO SQL
   DIDN’T WANT SERIALIZABILITY
   NOSQL
   

   View Slide

7. POOR PERFORMANCE
   NO SQL
   DIDN’T WANT SERIALIZABILITY
   NOSQL
   

   View Slide

8. POOR PERFORMANCE
   DELAY
   NO SQL
   DIDN’T WANT SERIALIZABILITY
   NOSQL
   

   View Slide

9. POOR PERFORMANCE
   DELAY
   PEAK THROUGHPUT: 1/DELAY
   FOR CONTENDED OPERATIONS
   NO SQL
   DIDN’T WANT SERIALIZABILITY
   NOSQL
   

   View Slide

10. POOR PERFORMANCE
    DELAY
    PEAK THROUGHPUT: 1/DELAY
    FOR CONTENDED OPERATIONS
    at .5MS,
    2K TXN/s
    at 50MS,
    20 TXN/s
    NO SQL
    DIDN’T WANT SERIALIZABILITY
    NOSQL
    

    View Slide

11. NO SQL
    DIDN’T WANT SERIALIZABILITY
    NOSQL
    POOR PERFORMANCE
    HIGH LATENCY
    

    View Slide

12. NO SQL
    DIDN’T WANT SERIALIZABILITY
    NOSQL
    POOR PERFORMANCE
    LIMITED AVAILABILITY
    HIGH LATENCY
    

    View Slide

13. STILL DON’T WANT SERIALIZABILITY
    “NOT ONLY SQL”
    

    View Slide

14. STILL DON’T WANT SERIALIZABILITY
    “NOT ONLY SQL”
    (DON’T WANT THE COSTS)
    

    View Slide

15. STILL DON’T WANT SERIALIZABILITY
    “NOT ONLY SQL”
    BUT WANT MORE FEATURES
    (DON’T WANT THE COSTS)
    

    View Slide

16. STILL DON’T WANT SERIALIZABILITY
    “NOT ONLY SQL”
    BUT WANT MORE FEATURES
    This paper!
    (DON’T WANT THE COSTS)
    

    View Slide

17. “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    

    View Slide

18. “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    

    View Slide

19. “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    

    View Slide

20. “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    

    View Slide

21. “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    FRIENDS
    FRIENDS
    

    View Slide

22. as
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    FRIENDS
    FRIENDS
    

    View Slide

23. as
    s
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    

    View Slide

24. as
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    s
    Denormalized Friend List
    

    View Slide

25. as
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    s
    Denormalized Friend List
    Fast reads…
    

    View Slide

26. as
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    s
    Denormalized Friend List
    Fast reads…
    …multi-entity updates
    

    View Slide

27. as
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    s
    Denormalized Friend List
    Fast reads…
    …multi-entity updates
    s
    

    View Slide

28. as
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    s
    Denormalized Friend List
    Fast reads…
    …multi-entity updates
    s
    

    View Slide

29. as
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    s
    Denormalized Friend List
    Fast reads…
    …multi-entity updates
    s
    

    View Slide

30. as
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    s
    Denormalized Friend List
    Fast reads…
    …multi-entity updates
    Not cleanly partitionable
    s
    

    View Slide

31. FOREIGN KEY DEPENDENCIES
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    

    View Slide

32. FOREIGN KEY DEPENDENCIES
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    “On Brewing Fresh Espresso: LinkedIn’s Distributed Data
    Serving Platform” SIGMOD 2013
    

    View Slide

33. FOREIGN KEY DEPENDENCIES
    “TAO: Facebook’s Distributed Data Store for the Social Graph”
    USENIX ATC 2013
    “On Brewing Fresh Espresso: LinkedIn’s Distributed Data
    Serving Platform” SIGMOD 2013
    “PNUTS: Yahoo!’s Hosted Data Serving Platform”
    VLDB 2008
    

    View Slide

34. View Slide

35. ID: 532
    AGE: 42
    ID: 123
    AGE: 22
    ID: 2345
    AGE: 1
    ID: 412
    AGE: 72
    ID: 892
    AGE: 13
    

    View Slide

36. ID: 532
    AGE: 42
    ID: 123
    AGE: 22
    ID: 2345
    AGE: 1
    ID: 412
    AGE: 72
    ID: 892
    AGE: 13
    

    View Slide

37. ID: 532
    AGE: 42
    ID: 123
    AGE: 22
    ID: 2345
    AGE: 1
    ID: 412
    AGE: 72
    ID: 892
    AGE: 13
    

    View Slide

38. ID: 532
    AGE: 42
    ID: 123
    AGE: 22
    ID: 2345
    AGE: 1
    ID: 412
    AGE: 72
    ID: 892
    AGE: 13
    

    View Slide

39. ID: 532
    AGE: 42
    ID: 123
    AGE: 22
    ID: 2345
    AGE: 1
    ID: 412
    AGE: 72
    ID: 892
    AGE: 13
    

    View Slide

40. ID: 532
    AGE: 42
    ID: 123
    AGE: 22
    ID: 2345
    AGE: 1
    ID: 412
    AGE: 72
    ID: 892
    AGE: 13
    Partition by
    primary key (ID)
    

    View Slide

41. ID: 532
    AGE: 42
    ID: 123
    AGE: 22
    ID: 2345
    AGE: 1
    ID: 412
    AGE: 72
    ID: 892
    AGE: 13
    Partition by
    primary key (ID) How should we look up by age?
    

    View Slide

42. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    

    View Slide

43. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    

    View Slide

44. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    

    View Slide

45. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    WRITE ONE SERVER, READ ALL
    

    View Slide

46. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    poor scalability
    WRITE ONE SERVER, READ ALL
    

    View Slide

47. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    poor scalability
    WRITE ONE SERVER, READ ALL
    

    View Slide

48. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    WRITE 2+ SERVERS, READ ONE
    poor scalability
    WRITE ONE SERVER, READ ALL
    

    View Slide

49. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    WRITE 2+ SERVERS, READ ONE
    scalable lookups
    poor scalability
    WRITE ONE SERVER, READ ALL
    

    View Slide

50. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    WRITE 2+ SERVERS, READ ONE
    scalable lookups
    poor scalability
    WRITE ONE SERVER, READ ALL
    

    View Slide

51. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    scalable lookups
    poor scalability
    WRITE 2+ SERVERS, READ ONE
    WRITE ONE SERVER, READ ALL
    

    View Slide

52. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    scalable lookups
    poor scalability
    WRITE 2+ SERVERS, READ ONE
    WRITE ONE SERVER, READ ALL
    OVERVIEW
    

    View Slide

53. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    scalable lookups
    poor scalability
    WRITE 2+ SERVERS, READ ONE
    WRITE ONE SERVER, READ ALL
    OVERVIEW
    INCONSISTENT
    GLOBAL 2i
    

    View Slide

54. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    scalable lookups
    poor scalability
    WRITE 2+ SERVERS, READ ONE
    WRITE ONE SERVER, READ ALL
    OVERVIEW
    INCONSISTENT
    GLOBAL 2i
    INCONSISTENT
    GLOBAL 2i
    

    View Slide

55. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    scalable lookups
    poor scalability
    WRITE 2+ SERVERS, READ ONE
    WRITE ONE SERVER, READ ALL
    OVERVIEW
    INCONSISTENT
    GLOBAL 2i
    INCONSISTENT
    GLOBAL 2i
    (PROPOSED)
    INCONSISTENT
    GLOBAL 2i
    

    View Slide

56. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    scalable lookups
    poor scalability
    WRITE 2+ SERVERS, READ ONE
    WRITE ONE SERVER, READ ALL
    OVERVIEW
    INCONSISTENT
    GLOBAL 2i
    INCONSISTENT
    GLOBAL 2i
    (PROPOSED)
    INCONSISTENT
    GLOBAL 2i
    INCONSISTENT
    GLOBAL 2i
    

    View Slide

57. SECONDARY INDEXING
    Partition by
    primary key (ID) How should we look up by age?
    Option I: Local Secondary Indexing
    Build indexes co-located with primary data
    Option II: Global Secondary Indexing
    Partition indexes by secondary key
    Partition by
    secondary attribute
    scalable lookups
    poor scalability
    WRITE 2+ SERVERS, READ ONE
    WRITE ONE SERVER, READ ALL
    OVERVIEW
    INCONSISTENT
    GLOBAL 2i
    INCONSISTENT
    GLOBAL 2i
    (PROPOSED)
    INCONSISTENT
    GLOBAL 2i
    INCONSISTENT
    GLOBAL 2i
    INCONSISTENT
    GLOBAL 2i
    

    View Slide

58. TABLE:
    ALL USERS
    

    View Slide

59. TABLE:
    ALL USERS
    

    View Slide

60. TABLE:
    ALL USERS
    TABLE:
    USERS OVER 25
    

    View Slide

61. TABLE:
    ALL USERS
    TABLE:
    USERS OVER 25
    

    View Slide

62. TABLE:
    ALL USERS
    TABLE:
    USERS OVER 25
    

    View Slide

63. TABLE:
    ALL USERS
    TABLE:
    USERS OVER 25
    

    View Slide

64. MATERIALIZED VIEWS
    TABLE:
    ALL USERS
    TABLE:
    USERS OVER 25
    

    View Slide

65. MATERIALIZED VIEWS
    TABLE:
    ALL USERS
    TABLE:
    USERS OVER 25
    RELEVANT
    RECENT
    EXAMPLES IN
    GOOGLE PERCOLATOR
    TWITTER RAINBIRD
    LINKEDIN ESPRESSO
    PAPERS
    

    View Slide

66. FOREIGN KEY DEPENDENCIES
    SECONDARY INDEXES
    MATERIALIZED VIEWS
    HOW SHOULD WE CORRECTLY MAINTAIN
    

    View Slide

67. SERIALIZABILITY
    

    View Slide

68. SNAPSHOT ISOLATION
    REPEATABLE READ
    (PL-2.99)
    CURSOR STABILITY
    READ UNCOMMITTED
    READ COMMITTED
    CAUSAL
    PRAM
    RYW
    LINEARIZABILITY
    EVENTUAL CONSISTENCY
    SERIALIZABILITY
    

    View Slide

69. SNAPSHOT ISOLATION
    REPEATABLE READ
    (PL-2.99)
    CURSOR STABILITY
    READ UNCOMMITTED
    READ COMMITTED
    CAUSAL
    PRAM
    RYW
    LINEARIZABILITY
    EVENTUAL CONSISTENCY
    SERIALIZABILITY
    

    View Slide

70. REPEATABLE READ
    (PL-2.99)
    SERIALIZABILITY
    SNAPSHOT ISOLATION
    CURSOR STABILITY
    READ UNCOMMITTED
    READ COMMITTED
    LINEARIZABILITY
    CAUSAL
    PRAM
    RYW
    EVENTUAL CONSISTENCY
    

    View Slide

71. REPEATABLE READ
    (PL-2.99)
    SERIALIZABILITY
    SNAPSHOT ISOLATION
    CURSOR STABILITY
    READ UNCOMMITTED
    READ COMMITTED
    LINEARIZABILITY
    MANY
    SUFFICIENT
    CAUSAL
    PRAM
    RYW
    EVENTUAL CONSISTENCY
    

    View Slide

72. REPEATABLE READ
    (PL-2.99)
    SERIALIZABILITY
    SNAPSHOT ISOLATION
    CURSOR STABILITY
    READ UNCOMMITTED
    READ COMMITTED
    LINEARIZABILITY
    REQUIRE SYNCHRONOUS COORDINATION
    MANY
    SUFFICIENT
    CAUSAL
    PRAM
    RYW
    EVENTUAL CONSISTENCY
    

    View Slide

73. SERIALIZABILITY
    SNAPSHOT ISOLATION
    REPEATABLE READ
    (PL-2.99)
    CURSOR STABILITY
    READ UNCOMMITTED
    READ COMMITTED
    CAUSAL
    PRAM
    RYW
    LINEARIZABILITY
    EVENTUAL CONSISTENCY
    REQUIRE SYNCHRONOUS COORDINATION
    MANY
    SUFFICIENT
    COORDINATION-FREE
    

    View Slide

74. SERIALIZABILITY
    SNAPSHOT ISOLATION
    REPEATABLE READ
    (PL-2.99)
    CURSOR STABILITY
    READ UNCOMMITTED
    READ COMMITTED
    CAUSAL
    PRAM
    RYW
    LINEARIZABILITY
    EVENTUAL CONSISTENCY
    REQUIRE SYNCHRONOUS COORDINATION
    INSUFFICIENT
    MANY
    SUFFICIENT
    COORDINATION-FREE
    

    View Slide

75. SERIALIZABILITY
    SNAPSHOT ISOLATION
    REPEATABLE READ
    (PL-2.99)
    CURSOR STABILITY
    LINEARIZABILITY
    REQUIRE SYNCHRONOUS COORDINATION
    INSUFFICIENT
    MANY
    SUFFICIENT
    COORDINATION-FREE
    Facebook TAO
    LinkedIn Espresso
    Yahoo! PNUTS
    Google Megastore
    Google App Engine
    Twitter Rainbird
    Amazon DynamoDB
    CONSCIOUS
    CHOICES!
    

    View Slide

76. SERIALIZABILITY
    SNAPSHOT ISOLATION
    REPEATABLE READ
    (PL-2.99)
    CURSOR STABILITY
    READ UNCOMMITTED
    READ COMMITTED CAUSAL PRAM
    RYW
    LINEARIZABILITY
    EVENTUAL CONSISTENCY
    COORDINATION-FREE
    INSUFFICIENT
    REQUIRE SYNCHRONOUS COORDINATION
    SUFFICIENT
    

    View Slide

77. SERIALIZABILITY
    SNAPSHOT ISOLATION
    REPEATABLE READ
    (PL-2.99)
    CURSOR STABILITY
    READ UNCOMMITTED
    READ COMMITTED CAUSAL PRAM
    RYW
    LINEARIZABILITY
    EVENTUAL CONSISTENCY
    COORDINATION-FREE
    RAMP (THIS PAPER)
    INSUFFICIENT
    REQUIRE SYNCHRONOUS COORDINATION
    SUFFICIENT
    

    View Slide

78. TRANSACTIONS
    R
    A
    M
    P
    TOMIC
    EAD
    ULTI-
    ARTITION
    

    View Slide

79. TRANSACTIONS
    R
    A
    M
    P
    TOMIC
    EAD
    ULTI-
    ARTITION
    

    View Slide

80. TRANSACTIONS
    RAMP
    EFFICIENTLY MAINTAIN
    

    View Slide

81. TRANSACTIONS
    RAMP
    FOREIGN KEY DEPENDENCIES
    EFFICIENTLY MAINTAIN
    

    View Slide

82. TRANSACTIONS
    RAMP
    FOREIGN KEY DEPENDENCIES
    SECONDARY INDEXES
    EFFICIENTLY MAINTAIN
    

    View Slide

83. TRANSACTIONS
    RAMP
    FOREIGN KEY DEPENDENCIES
    SECONDARY INDEXES
    MATERIALIZED VIEWS
    EFFICIENTLY MAINTAIN
    

    View Slide

84. TRANSACTIONS
    RAMP
    FOREIGN KEY DEPENDENCIES
    SECONDARY INDEXES
    MATERIALIZED VIEWS
    BY PROVIDING
    ATOMIC VISIBILITY
    EFFICIENTLY MAINTAIN
    

    View Slide

85. ATOMIC VISIBILITY
    

    View Slide

86. Informally:
    Either all of each transaction’s updates are visible,
    or none are
    ATOMIC VISIBILITY
    

    View Slide

87. Informally:
    Either all of each transaction’s updates are visible,
    or none are
    ATOMIC VISIBILITY
    

    View Slide

88. Informally:
    Either all of each transaction’s updates are visible,
    or none are
    ATOMIC VISIBILITY
    WRITE X = 1
    WRITE Y = 1
    

    View Slide

89. Informally:
    Either all of each transaction’s updates are visible,
    or none are
    ATOMIC VISIBILITY
    WRITE X = 1
    WRITE Y = 1
    READ X = 1
    READ Y = 1
    

    View Slide

90. Informally:
    Either all of each transaction’s updates are visible,
    or none are
    ATOMIC VISIBILITY
    WRITE X = 1
    WRITE Y = 1
    READ X = 1
    READ Y = 1
    OR
    

    View Slide

91. Informally:
    Either all of each transaction’s updates are visible,
    or none are
    ATOMIC VISIBILITY
    WRITE X = 1
    WRITE Y = 1
    READ X = 1
    READ Y = 1
    READ X = ∅
    READ Y = ∅
    OR
    

    View Slide

92. Informally:
    Either all of each transaction’s updates are visible,
    or none are
    ATOMIC VISIBILITY
    READ X = 1
    READ Y = 1
    READ X = ∅
    READ Y = ∅
    OR
    

    View Slide

93. Informally:
    Either all of each transaction’s updates are visible,
    or none are
    ATOMIC VISIBILITY
    READ X = 1
    READ Y = 1
    READ X = ∅
    READ Y = ∅
    OR
    BUT NOT
    READ Y = ∅
    READ X = 1
    

    View Slide

94. Informally:
    Either all of each transaction’s updates are visible,
    or none are
    ATOMIC VISIBILITY
    READ X = 1
    READ Y = 1
    READ X = ∅
    READ Y = ∅
    OR
    BUT NOT
    READ X = ∅
    READ Y = ∅
    READ X = 1
    OR
    READ Y = 1
    

    View Slide

95. ATOMIC VISIBILITY
    We also provide per-item PRAM guarantees with
    per-transaction regular semantics (see paper Appendix)
    Formally:
    A transaction Tj
    exhibits fractured reads if transaction Ti
    writes versions xm
    and yn
    (in any order,
    with x possibly but not necessarily equal to y), Tj
    reads version xm
    and version yk
    , and k !
    A system provides Read Atomic isolation (RA) if it prevents fractured reads anomalies and also
    prevents transactions from reading uncommitted, aborted, or intermediate data.
    FORMALIZED AS
    READ ATOMIC ISOLATION
    

    View Slide

96. TRANSACTIONS
    RAMP
    GUARANTEE
    ATOMIC VISIBILITY
    

    View Slide

97. TRANSACTIONS
    RAMP
    GUARANTEE
    ATOMIC VISIBILITY
    WHILE ENSURING
    

    View Slide

98. TRANSACTIONS
    RAMP
    GUARANTEE
    ATOMIC VISIBILITY
    WHILE ENSURING
    PARTITION INDEPENDENCE
    

    View Slide

99. TRANSACTIONS
    RAMP
    GUARANTEE
    ATOMIC VISIBILITY
    WHILE ENSURING
    PARTITION INDEPENDENCE
    clients only access servers responsible
    for data in transactions
    

    View Slide

100. TRANSACTIONS
     RAMP
     GUARANTEE
     ATOMIC VISIBILITY
     WHILE ENSURING
     PARTITION INDEPENDENCE
     clients only access servers responsible
     for data in transactions
     W(X=1)
     W(Y=1)
     X Y Z
     

     View Slide

101. TRANSACTIONS
     RAMP
     GUARANTEE
     ATOMIC VISIBILITY
     WHILE ENSURING
     PARTITION INDEPENDENCE
     clients only access servers responsible
     for data in transactions
     W(X=1)
     W(Y=1)
     X Y Z
     

     View Slide

102. TRANSACTIONS
     RAMP
     GUARANTEE
     ATOMIC VISIBILITY
     WHILE ENSURING
     PARTITION INDEPENDENCE
     AND
     SYNCHRONIZATION INDEPENDENCE
     clients only access servers responsible
     for data in transactions
     transactions always commit* and no
     client can cause another client to block
     

     View Slide

103. TRANSACTIONS
     RAMP
     GUARANTEE
     ATOMIC VISIBILITY
     ARE NOT SERIALIZABLE
     DO NOT PREVENT LOST UPDATE
     DO NOT PREVENT WRITE SKEW
     ALLOW CONCURRENT UPDATES
     

     View Slide

104. TRANSACTIONS
     RAMP
     GUARANTEE
     ATOMIC VISIBILITY
     ARE NOT SERIALIZABLE
     DO NOT PREVENT LOST UPDATE
     DO NOT PREVENT WRITE SKEW
     ALLOW CONCURRENT UPDATES
     ARE GUIDED BY REAL WORLD USE CASES
     FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     

     View Slide

105. TRANSACTIONS
     RAMP
     GUARANTEE
     ATOMIC VISIBILITY
     ARE NOT SERIALIZABLE
     DO NOT PREVENT LOST UPDATE
     DO NOT PREVENT WRITE SKEW
     ALLOW CONCURRENT UPDATES
     ARE GUIDED BY REAL WORLD USE CASES
     FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     Facebook TAO
     LinkedIn Espresso
     Yahoo! PNUTS
     Google Megastore
     Google App Engine
     Twitter Rainbird
     Amazon DynamoDB
     

     View Slide

106. STRAWMAN: LOCKING
     X=0 Y=0
     

     View Slide

107. STRAWMAN: LOCKING
     X=0 Y=0
     W(X=1)
     W(Y=1)
     

     View Slide

108. STRAWMAN: LOCKING
     X=0 Y=0
     W(X=1)
     W(Y=1)
     

     View Slide

109. STRAWMAN: LOCKING
     X=0 Y=0
     W(X=1)
     W(Y=1)
     

     View Slide

110. STRAWMAN: LOCKING
     X=1 Y=1
     W(X=1)
     W(Y=1)
     

     View Slide

111. STRAWMAN: LOCKING
     X=1 Y=1
     W(X=1)
     W(Y=1)
     

     View Slide

112. STRAWMAN: LOCKING
     X=1 Y=1
     W(X=1)
     W(Y=1)
     

     View Slide

113. STRAWMAN: LOCKING
     X=1 Y=1
     W(X=1)
     W(Y=1)
     

     View Slide

114. STRAWMAN: LOCKING
     X=1 Y=1
     W(X=1)
     W(Y=1)
     R(X=1)
     

     View Slide

115. STRAWMAN: LOCKING
     X=1 Y=1
     W(X=1)
     W(Y=1)
     R(X=1)
     R(Y=1)
     

     View Slide

116. Y=0
     STRAWMAN: LOCKING
     X=1
     W(X=1)
     W(Y=1)
     

     View Slide

117. Y=0
     STRAWMAN: LOCKING
     X=1
     W(X=1)
     W(Y=1)
     

     View Slide

118. Y=0
     STRAWMAN: LOCKING
     X=1
     W(X=1)
     W(Y=1)
     R(X=?)
     

     View Slide

119. Y=0
     STRAWMAN: LOCKING
     X=1
     W(X=1)
     W(Y=1)
     R(X=?)
     R(Y=?)
     

     View Slide

120. Y=0
     STRAWMAN: LOCKING
     X=1
     W(X=1)
     W(Y=1)
     R(X=?)
     R(Y=?)
     ATOMIC VISIBILITY
     COUPLED WITH
     MUTUAL EXCLUSION
     

     View Slide

121. STRAWMAN: LOCKING
     X=1
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     ATOMIC VISIBILITY
     COUPLED WITH
     MUTUAL EXCLUSION
     

     View Slide

122. STRAWMAN: LOCKING
     X=1
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     ATOMIC VISIBILITY
     COUPLED WITH
     MUTUAL EXCLUSION
     

     View Slide

123. STRAWMAN: LOCKING
     X=1
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     ATOMIC VISIBILITY
     COUPLED WITH
     MUTUAL EXCLUSION
     RTT
     

     View Slide

124. STRAWMAN: LOCKING
     X=1
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     ATOMIC VISIBILITY
     COUPLED WITH
     MUTUAL EXCLUSION
     RTT
     unavailability!
     

     View Slide

125. X=1
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     ATOMIC VISIBILITY
     COUPLED WITH
     MUTUAL EXCLUSION
     at .5 MS
     < 2K TPS!
     unavailable
     during
     failures
     SIMILAR ISSUES IN MVCC,
     PRE-SCHEDULING
     SERIALIZABLE OCC,
     (global timestamp assignment/application)
     (multi-partition validation, liveness)
     (scheduling, multi-partition execution)
     STRAWMAN: LOCKING
     

     View Slide

126. X=1
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     ATOMIC VISIBILITY
     COUPLED WITH
     MUTUAL EXCLUSION
     at .5 MS
     < 2K TPS!
     unavailable
     during
     failures
     SIMILAR ISSUES IN MVCC,
     PRE-SCHEDULING
     SERIALIZABLE OCC,
     (global timestamp assignment/application)
     (multi-partition validation, liveness)
     (scheduling, multi-partition execution)
     FUNDAMENTAL
     TO
     “STRONG”
     SEMANTICS
     STRAWMAN: LOCKING
     

     View Slide

127. BASIC IDEA
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     X=1
     

     View Slide

128. BASIC IDEA
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     X=1
     

     View Slide

129. BASIC IDEA
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     LET CLIENTS RACE, but
     HAVE READERS “CLEAN UP”
     X=1
     

     View Slide

130. BASIC IDEA
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     LET CLIENTS RACE, but
     HAVE READERS “CLEAN UP”
     X=1
     METADATA
     

     View Slide

131. BASIC IDEA
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     LET CLIENTS RACE, but
     HAVE READERS “CLEAN UP”
     X=1
     + LIMITED
     MULTI-VERSIONING
     METADATA
     

     View Slide

132. BASIC IDEA
     W(X=1)
     W(Y=1)
     Y=0
     R(X=?)
     R(Y=?)
     LET CLIENTS RACE, but
     HAVE READERS “CLEAN UP”
     X=1
     + LIMITED
     MULTI-VERSIONING
     METADATA
     FOR NOW:

     READ-ONLY,
     WRITE-ONLY
     TXNS
     

     View Slide

133. last committed stamp for x: 0
     RAMP-Fast
     last committed stamp for y: 0
     

     View Slide

134. last committed stamp for x: 0
     RAMP-Fast
     last committed stamp for y: 0
     

     View Slide

135. last committed stamp for x: 0
     RAMP-Fast
     last committed stamp for y: 0
     

     View Slide

136. last committed stamp for x: 0
     RAMP-Fast
     known versions of x
     last committed stamp for y: 0
     known versions of y
     

     View Slide

137. last committed stamp for x: 0
     RAMP-Fast
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     

     View Slide

138. last committed stamp for x: 0
     RAMP-Fast
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     

     View Slide

139. last committed stamp for x: 0
     RAMP-Fast
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     

     View Slide

140. last committed stamp for x: 0
     RAMP-Fast
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     

     View Slide

141. last committed stamp for x: 0
     RAMP-Fast
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     

     View Slide

142. last committed stamp for x: 0
     RAMP-Fast
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     

     View Slide

143. last committed stamp for x: 0
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     

     View Slide

144. last committed stamp for x: 0
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     timestamp: 242
     e.g., time concat client ID
     concat sequence number
     

     View Slide

145. last committed stamp for x: 0
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     timestamp: 242
     

     View Slide

146. last committed stamp for x: 0
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     timestamp: 242
     

     View Slide

147. last committed stamp for x: 0
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     242 1
     timestamp: 242
     

     View Slide

148. last committed stamp for x: 0
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     242 1
     242 1
     timestamp: 242
     

     View Slide

149. last committed stamp for x: 0
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     

     View Slide

150. RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     last committed stamp for y: 0
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     last committed stamp for x: 242
     timestamp: 242
     

     View Slide

151. RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 242
     timestamp: 242
     

     View Slide

152. RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     R(X=?)
     R(Y=?)
     last committed stamp for x: 242
     last committed stamp for y: 242
     

     View Slide

153. RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     R(X=1)
     R(Y=1)
     last committed stamp for x: 242
     last committed stamp for y: 242
     

     View Slide

154. RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     R(X=1)
     R(Y=1)
     last committed stamp for x: 242
     last committed stamp for y: 242
     

     View Slide

155. R(X=?)
     R(Y=?)
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     last committed stamp for x: 242
     last committed stamp for y: 0
     

     View Slide

156. R(X=?)
     R(Y=?)
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     last committed stamp for x: 242
     last committed stamp for y: 0
     

     View Slide

157. R(X=?)
     R(Y=?)
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     RACE!!!
     last committed stamp for x: 242
     last committed stamp for y: 0
     

     View Slide

158. R(X=?)
     R(Y=?)
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     RACE!!!
     R(X=1)
     R(Y=0)
     last committed stamp for x: 242
     last committed stamp for y: 0
     

     View Slide

159. R(X=?)
     R(Y=?)
     RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     RACE!!!
     R(X=1)
     R(Y=0)
     last committed stamp for x: 242
     last committed stamp for y: 0
     

     View Slide

160. RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     RACE!!!
     R(X=1)
     R(Y=0)
     last committed stamp for x: 242
     last committed stamp for y: 0
     

     View Slide

161. RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     RACE!!!
     R(X=1)
     R(Y=0)
     last committed stamp for x: 242
     last committed stamp for y: 0
     RECORD THE ITEMS
     WRITTEN IN THE
     TRANSACTION
     

     View Slide

162. RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     RACE!!!
     R(X=1)
     R(Y=0)
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     RECORD THE ITEMS
     WRITTEN IN THE
     TRANSACTION
     

     View Slide

163. RAMP-Fast
     W(X=1)
     W(Y=1)
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     1.) Assign unique (logical)
     transaction timestamp.
     2.) Add write to known
     versions on partition.
     3.) Commit and update last
     committed stamp.
     242 1
     242 1
     timestamp: 242
     RACE!!!
     R(X=1)
     R(Y=0)
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     RECORD THE ITEMS
     WRITTEN IN THE
     TRANSACTION
     

     View Slide

164. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     

     View Slide

165. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     

     View Slide

166. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     

     View Slide

167. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     1.) Read last committed:
     

     View Slide

168. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     1.) Read last committed:
     X=1 @ 242, {Y}
     

     View Slide

169. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     1.) Read last committed:
     X=1 @ 242, {Y}
     Y=NULL @ 0, {}
     

     View Slide

170. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     1.) Read last committed:
     2.) Calculate missing versions:
     X=1 @ 242, {Y}
     Y=NULL @ 0, {}
     

     View Slide

171. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     1.) Read last committed:
     2.) Calculate missing versions:
     X=1 @ 242, {Y}
     Y=NULL @ 0, {}
     ITEM HIGHEST TS
     X 242
     Y 242
     

     View Slide

172. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     1.) Read last committed:
     2.) Calculate missing versions:
     X=1 @ 242, {Y}
     Y=NULL @ 0, {}
     ITEM HIGHEST TS
     X 242
     Y 242
     

     View Slide

173. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     1.) Read last committed:
     2.) Calculate missing versions:
     3.) Fetch missing versions.
     X=1 @ 242, {Y}
     Y=NULL @ 0, {}
     ITEM HIGHEST TS
     X 242
     Y 242
     

     View Slide

174. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     1.) Read last committed:
     2.) Calculate missing versions:
     3.) Fetch missing versions.
     X=1 @ 242, {Y}
     Y=NULL @ 0, {}
     ITEM HIGHEST TS
     X 242
     Y 242
     Y=1 @ 242, {X}
     (Send required timestamp in request)
     

     View Slide

175. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     1.) Read last committed:
     2.) Calculate missing versions:
     3.) Fetch missing versions.
     X=1 @ 242, {Y}
     Y=NULL @ 0, {}
     ITEM HIGHEST TS
     X 242
     Y 242
     Y=1 @ 242, {X}
     (Send required timestamp in request)
     2PC ENSURES NO
     WAIT AT SERVER
     

     View Slide

176. RAMP-Fast
     known versions of x
     known versions of y
     TIMESTAMP VALUE METADATA
     0 NULL {}
     TIMESTAMP VALUE METADATA
     0 NULL {}
     242 1
     242 1
     last committed stamp for x: 242
     last committed stamp for y: 0
     {y}
     {x}
     R(X=?)
     R(Y=?)
     1.) Read last committed:
     2.) Calculate missing versions:
     3.) Fetch missing versions.
     X=1 @ 242, {Y}
     Y=NULL @ 0, {}
     ITEM HIGHEST TS
     X 242
     Y 242
     Y=1 @ 242, {X}
     (Send required timestamp in request)
     4.) Return resulting set.
     R(X=1)
     R(Y=1)
     2PC ENSURES NO
     WAIT AT SERVER
     

     View Slide

177. RAMP-Fast
     

     View Slide

178. RAMP-Fast
     2 RTT writes:
     

     View Slide

179. RAMP-Fast
     2 RTT writes:
     2PC, without blocking synchronization
     

     View Slide

180. RAMP-Fast
     2 RTT writes:
     2PC, without blocking synchronization
     ENSURES READERS
     NEVER WAIT!
     

     View Slide

181. RAMP-Fast
     2 RTT writes:
     2PC, without blocking synchronization
     metadata size linear in transaction size
     ENSURES READERS
     NEVER WAIT!
     

     View Slide

182. RAMP-Fast
     2 RTT writes:
     2PC, without blocking synchronization
     metadata size linear in transaction size
     1 RTT reads:
     in race-free case
     ENSURES READERS
     NEVER WAIT!
     

     View Slide

183. RAMP-Fast
     2 RTT writes:
     2PC, without blocking synchronization
     metadata size linear in transaction size
     1 RTT reads:
     in race-free case
     2 RTT reads:
     otherwise
     ENSURES READERS
     NEVER WAIT!
     

     View Slide

184. RAMP-Fast
     2 RTT writes:
     2PC, without blocking synchronization
     metadata size linear in transaction size
     1 RTT reads:
     in race-free case
     2 RTT reads:
     otherwise
     no fast-path
     synchronization
     ENSURES READERS
     NEVER WAIT!
     

     View Slide

185. RAMP-Fast
     2 RTT writes:
     2PC, without blocking synchronization
     metadata size linear in transaction size
     1 RTT reads:
     in race-free case
     2 RTT reads:
     otherwise
     no fast-path
     synchronization
     ENSURES READERS
     NEVER WAIT!
     CAN WE USE LESS
     METADATA?
     

     View Slide

186. RAMP-Small
     

     View Slide

187. RAMP-Small
     2 RTT writes:
     

     View Slide

188. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     

     View Slide

189. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     2 RTT reads
     

     View Slide

190. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     2 RTT reads always
     

     View Slide

191. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     2 RTT reads
     1.) For each item, fetch the highest committed timestamp.
     2.) Request highest matching write with timestamp in step 1.
     always
     

     View Slide

192. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     2 RTT reads
     INTUITION:
     1.) For each item, fetch the highest committed timestamp.
     2.) Request highest matching write with timestamp in step 1.
     always
     

     View Slide

193. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     2 RTT reads
     INTUITION:
     1.) For each item, fetch the highest committed timestamp.
     2.) Request highest matching write with timestamp in step 1.
     X time 523
     always
     

     View Slide

194. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     2 RTT reads
     INTUITION:
     1.) For each item, fetch the highest committed timestamp.
     2.) Request highest matching write with timestamp in step 1.
     X time 523
     Y time 247
     always
     

     View Slide

195. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     2 RTT reads
     INTUITION:
     1.) For each item, fetch the highest committed timestamp.
     2.) Request highest matching write with timestamp in step 1.
     X time 523
     Y time 247
     Z time 842
     always
     

     View Slide

196. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     2 RTT reads
     INTUITION:
     1.) For each item, fetch the highest committed timestamp.
     2.) Request highest matching write with timestamp in step 1.
     X time 523
     Y time 247
     Z time 842
     {247, 523, 842}
     always
     

     View Slide

197. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     2 RTT reads
     partial commits will be in this set
     INTUITION:
     1.) For each item, fetch the highest committed timestamp.
     2.) Request highest matching write with timestamp in step 1.
     X time 523
     Y time 247
     Z time 842
     {247, 523, 842}
     always
     

     View Slide

198. RAMP-Small
     2 RTT writes:
     same basic protocol as RAMP-Fast
     but drop all RAMP-Fast metadata
     2 RTT reads
     partial commits will be in this set
     INTUITION:
     1.) For each item, fetch the highest committed timestamp.
     2.) Request highest matching write with timestamp in step 1.
     X time 523
     Y time 247
     Z time 842
     {247, 523, 842}
     send it to all participating servers
     always
     

     View Slide

199. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(1)
     Bloom filter
     

     View Slide

200. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(1)
     Bloom filter
     

     View Slide

201. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(1)
     Bloom filter
     

     View Slide

202. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(1)
     Bloom filter
     

     View Slide

203. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(1)
     Bloom filter
     

     View Slide

204. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2
     log(2)2
     O([txn len]*log(1/ε))
     

     View Slide

205. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2
     log(2)2
     O([txn len]*log(1/ε))
     

     View Slide

206. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2
     BLOOM FILTER SUMMARIZES WRITE SET
     FALSE POSITIVES: EXTRA RTTs
     log(2)2
     O([txn len]*log(1/ε))
     

     View Slide

207. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2
     BLOOM FILTER SUMMARIZES WRITE SET
     FALSE POSITIVES: EXTRA RTTs
     log(2)2
     O([txn len]*log(1/ε))
     

     View Slide

208. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(B)
     Bloom filter
     

     View Slide

209. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(B)
     Bloom filter
     • AVOID IN-PLACE UPDATES
     

     View Slide

210. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(B)
     Bloom filter
     • AVOID IN-PLACE UPDATES
     • EMBRACE RACES TO IMPROVE CONCURRENCY
     

     View Slide

211. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(B)
     Bloom filter
     • AVOID IN-PLACE UPDATES
     • EMBRACE RACES TO IMPROVE CONCURRENCY
     • ALLOW READERS TO REPAIR PARTIAL WRITES
     

     View Slide

212. RAMP Summary
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(B)
     Bloom filter
     • AVOID IN-PLACE UPDATES
     • EMBRACE RACES TO IMPROVE CONCURRENCY
     • ALLOW READERS TO REPAIR PARTIAL WRITES
     • USE 2PC TO AVOID READER STALLS
     

     View Slide

213. Additional Details
     

     View Slide

214. Additional Details
     Garbage collection:
     limit read transaction duration to K seconds
     GC overwritten versions after K seconds
     

     View Slide

215. Additional Details
     Garbage collection:
     limit read transaction duration to K seconds
     GC overwritten versions after K seconds
     Replication
     paper assumes linearizable masters
     extendable to “AP” systems
     see HAT by Bailis et al., VLDB 2014
     

     View Slide

216. Additional Details
     Garbage collection:
     limit read transaction duration to K seconds
     GC overwritten versions after K seconds
     Failure handling:
     blocked 2PC rounds do not block clients
     stalled commits? versions are not GC’d
     if desirable, use CTP termination protocol
     Replication
     paper assumes linearizable masters
     extendable to “AP” systems
     see HAT by Bailis et al., VLDB 2014
     

     View Slide

217. RAMP PERFORMANCE
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(B)
     Bloom filter
     

     View Slide

218. RAMP PERFORMANCE
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(B)
     Bloom filter
     EVALUATED ON EC2
     cr1.8xlarge instances
     (cluster size: 1-100 servers; default: 5)
     !
     open sourced on GitHub; see link at end of talk
     

     View Slide

219. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     

     View Slide

220. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     

     View Slide

221. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     Doesn’t provide
     atomic visibility
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     

     View Slide

222. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     

     View Slide

223. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     

     View Slide

224. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     

     View Slide

225. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     Also doesn’t provide
     atomic visibility
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     

     View Slide

226. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     Representative of
     coordinated approaches
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     

     View Slide

227. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     

     View Slide

228. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     

     View Slide

229. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     Within ~5% of
     baseline
     !
     Latency in paper
     (comparable)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     

     View Slide

230. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     

     View Slide

231. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     RAMP-F RAMP-S RAMP-H
     RAMP-Small
     

     View Slide

232. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     Always needs
     2 RTT reads
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     RAMP-F RAMP-S RAMP-H
     RAMP-Small
     

     View Slide

233. YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     RAMP-F RAMP-S RAMP-H
     RAMP-Small
     

     View Slide

234. RAMP-F RAMP-S RAMP-H NWNR
     RAMP-Hybrid
     YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     RAMP-F RAMP-S RAMP-H
     RAMP-Small
     

     View Slide

235. RAMP-F RAMP-S RAMP-H NWNR
     RAMP-Hybrid
     YCSB: WorkloadA, 95% reads, 1M items, 4 items/txn
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     0 2000 4000 6000 8000 10000
     Concurrent Clients
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     RAMP-F RAMP-S RAMP-H
     RAMP-Small
     

     View Slide

236. YCSB: WorkloadA, 1M items, 4 items/txn, 5K clients
     0 25 50 75 100
     Percentage Reads
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     

     View Slide

237. YCSB: WorkloadA, 1M items, 4 items/txn, 5K clients
     0 25 50 75 100
     Percentage Reads
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     

     View Slide

238. YCSB: WorkloadA, 1M items, 4 items/txn, 5K clients
     0 25 50 75 100
     Percentage Reads
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     

     View Slide

239. RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     RAMP-F RAMP-S RAMP-H
     RAMP-Small
     YCSB: WorkloadA, 1M items, 4 items/txn, 5K clients
     0 25 50 75 100
     Percentage Reads
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-F RAMP-S RAMP-H NWNR
     RAMP-Hybrid
     

     View Slide

240. RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     RAMP-F RAMP-S RAMP-H
     RAMP-Small
     YCSB: WorkloadA, 1M items, 4 items/txn, 5K clients
     0 25 50 75 100
     Percentage Reads
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-F RAMP-S RAMP-H NWNR
     RAMP-Hybrid
     Linear scaling; due to
     2RTT writes, races
     

     View Slide

241. RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     LWSR LWLR E-PCI
     Serializable 2PL
     NWNR LWNR LWSR LWLR E-PCI
     Write Locks Only
     RAMP-F RAMP-S
     RAMP-Fast
     RAMP-F RAMP-S RAMP-H
     RAMP-Small
     YCSB: WorkloadA, 1M items, 4 items/txn, 5K clients
     0 25 50 75 100
     Percentage Reads
     0
     30K
     60K
     90K
     120K
     150K
     180K
     Throughput (txn/s)
     RAMP-F RAMP-S RAMP-H NWNR
     RAMP-Hybrid
     

     View Slide

242. YCSB: uniform access, 1M items, 4 items/txn, 95% reads
     0 25 50 75 100
     Number of Servers
     0
     2M
     4M
     6M
     8M
     Throughput (ops/s)
     

     View Slide

243. RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control
     YCSB: uniform access, 1M items, 4 items/txn, 95% reads
     0 25 50 75 100
     Number of Servers
     0
     2M
     4M
     6M
     8M
     Throughput (ops/s)
     

     View Slide

244. RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control RAMP-F RAMP-S
     RAMP-Fast
     RAMP-F RAMP-S RAMP-H
     RAMP-Small
     RAMP-F RAMP-S RAMP-H NWNR
     RAMP-Hybrid
     YCSB: uniform access, 1M items, 4 items/txn, 95% reads
     0 25 50 75 100
     Number of Servers
     0
     2M
     4M
     6M
     8M
     Throughput (ops/s)
     

     View Slide

245. RAMP-H NWNR LWNR LWSR LWLR E-PCI
     No Concurrency Control RAMP-F RAMP-S
     RAMP-Fast
     RAMP-F RAMP-S RAMP-H
     RAMP-Small
     RAMP-F RAMP-S RAMP-H NWNR
     RAMP-Hybrid
     YCSB: uniform access, 1M items, 4 items/txn, 95% reads
     0 25 50 75 100
     Number of Servers
     0
     40K
     80K
     120K
     160K
     200K
     operations/s/server
     

     View Slide

246. RAMP PERFORMANCE
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(B)
     Bloom filter
     

     View Slide

247. RAMP PERFORMANCE
     Algorithm Write RTT READ RTT
     (best case)
     READ RTT
     (worst case) METADATA
     RAMP-Fast 2 1 2 O(txn len)
     write set
     summary
     RAMP-Small 2 2 2 O(1)
     timestamp
     RAMP-Hybrid 2 1+ε 2 O(B)
     Bloom filter
     More results in paper:
     Transaction length, contention,
     value size, latency, failures
     

     View Slide

248. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     

     View Slide

249. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     as
     s
     

     View Slide

250. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     as
     s
     

     View Slide

251. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     as
     s
     

     View Slide

252. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     as
     s
     MULTI-PUT
     (DELETES VIA
     TOMBSTONES)
     

     View Slide

253. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     

     View Slide

254. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     

     View Slide

255. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     

     View Slide

256. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     Maintain list of matching record IDs and versions
     e.g., HAS_BEARD={52@512, 412@52, 123@512}
     merge lists on commit/read (LWW by timestamp for conflicts)
     

     View Slide

257. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     Maintain list of matching record IDs and versions
     e.g., HAS_BEARD={52@512, 412@52, 123@512}
     merge lists on commit/read (LWW by timestamp for conflicts)
     LOOKUPs: READ INDEX, THEN FETCH DATA
     

     View Slide

258. FOREIGN KEY DEPENDENCIES
     SECONDARY INDEXING
     MATERIALIZED VIEWS
     HOW RAMP HANDLES:
     Maintain list of matching record IDs and versions
     e.g., HAS_BEARD={52@512, 412@52, 123@512}
     merge lists on commit/read (LWW by timestamp for conflicts)
     LOOKUPs: READ INDEX, THEN FETCH DATA
     SIMILAR FOR
     SELECT/PROJECT
     

     View Slide

259. SERIALIZABILITY
     SNAPSHOT ISOLATION
     REPEATABLE READ
     (PL-2.99)
     CURSOR STABILITY
     READ UNCOMMITTED
     READ COMMITTED
     CAUSAL
     PRAM
     RYW
     LINEARIZABILITY
     EVENTUAL CONSISTENCY
     REQUIRE SYNCHRONOUS COORDINATION
     INSUFFICIENT
     SUFFICIENT
     COORDINATION-FREE
     

     View Slide

260. SERIALIZABILITY
     SNAPSHOT ISOLATION
     REPEATABLE READ
     (PL-2.99)
     CURSOR STABILITY
     READ UNCOMMITTED
     READ COMMITTED
     CAUSAL
     PRAM
     RYW
     LINEARIZABILITY
     EVENTUAL CONSISTENCY
     REQUIRE SYNCHRONOUS COORDINATION
     INSUFFICIENT
     SUFFICIENT
     COORDINATION-FREE
     

     View Slide

261. SERIALIZABILITY
     SNAPSHOT ISOLATION
     REPEATABLE READ
     (PL-2.99)
     CURSOR STABILITY
     READ UNCOMMITTED
     READ COMMITTED CAUSAL
     PRAM
     RYW
     LINEARIZABILITY
     EVENTUAL CONSISTENCY
     COORDINATION-FREE
     INSUFFICIENT
     SUFFICIENT
     REQUIRE SYNCHRONOUS COORDINATION
     

     View Slide

262. SERIALIZABILITY
     SNAPSHOT ISOLATION
     REPEATABLE READ
     (PL-2.99)
     CURSOR STABILITY
     READ UNCOMMITTED
     READ COMMITTED CAUSAL
     PRAM
     RYW
     LINEARIZABILITY
     EVENTUAL CONSISTENCY
     COORDINATION-FREE
     ATOMIC VISIBILITY VIA RAMP
     INSUFFICIENT
     SUFFICIENT
     REQUIRE SYNCHRONOUS COORDINATION
     

     View Slide

263. RAMP IN CASSANDRA
     

     View Slide

264. RAMP IN CASSANDRA
     USES
     

     View Slide

265. RAMP IN CASSANDRA
     USES
     REQUIREMENTS
     

     View Slide

266. RAMP IN CASSANDRA
     USES
     REQUIREMENTS
     IMPLEMENTATION
     

     View Slide

267. RAMP IN CASSANDRA
     STRAIGHTFORWARD USES:
     •Add atomic visibility to atomic batch operations
     •Expose as CQL isolation level
     • USING CONSISTENCY READ_ATOMIC
     •Encourage use in multi-put, multi-get
     •Treat as basis for global secondary indexing
     •CREATE GLOBAL INDEX on users (age )
     

     View Slide

268. RAMP IN CASSANDRA
     REQUIREMENTS:
     •Unique timestamp generation for transactions
     •Use node ID from ring
     •Other form of UUID
     •Hash transaction contents*
     •Limited multi-versioning for prepared and old values
     •RAMP doesn’t actually require true MVCC
     •One proposal: keep a look aside cache
     

     View Slide

269. RAMP IN CASSANDRA
     POSSIBLE IMPLEMENTATION:
     Lookaside cache for prepared and old values
     !
     Standard C* Table stores
     last committed write
     1
     52
     335
     1240
     1402
     2201
     

     View Slide

270. RAMP IN CASSANDRA
     POSSIBLE IMPLEMENTATION:
     Lookaside cache for prepared and old values
     !
     Standard C* Table stores
     last committed write
     Shadow table stores
     prepared-but-not-committed
     and overwritten versions
     1
     52
     335
     1240
     1402
     2201
     64
     335
     2201
     

     View Slide

271. RAMP IN CASSANDRA
     POSSIBLE IMPLEMENTATION:
     Lookaside cache for prepared and old values
     !
     Standard C* Table stores
     last committed write
     Shadow table stores
     prepared-but-not-committed
     and overwritten versions
     1
     52
     335
     1240
     1402
     2201
     64
     335
     2201
     Transparent to end-users
     

     View Slide

272. RAMP IN CASSANDRA
     POSSIBLE IMPLEMENTATION:
     Lookaside cache for prepared and old values
     !
     Standard C* Table stores
     last committed write
     Shadow table stores
     prepared-but-not-committed
     and overwritten versions
     1
     52
     335
     1240
     1402
     2201
     64
     335
     2201
     Overwritten versions have
     TTL set to max read transaction
     time, do not need durability
     Transparent to end-users
     

     View Slide

273. RAMP IN CASSANDRA
     POSSIBLE IMPLEMENTATION:
     

     View Slide

274. RAMP IN CASSANDRA
     POSSIBLE IMPLEMENTATION:
     OPERATION CONSISTENCY LEVEL
     Write Prepare CL.QUORUM
     Write Commit CL.ANY or higher
     First-round Read CL.ANY/CL.ONE
     Second-round Read CL.QUORUM
     

     View Slide

275. RAMP IN CASSANDRA
     To avoid stalling, second-round reads must
     be able to access prepared writes
     POSSIBLE IMPLEMENTATION:
     OPERATION CONSISTENCY LEVEL
     Write Prepare CL.QUORUM
     Write Commit CL.ANY or higher
     First-round Read CL.ANY/CL.ONE
     Second-round Read CL.QUORUM
     

     View Slide

276. RAMP IN CASSANDRA
     To avoid stalling, second-round reads must
     be able to access prepared writes
     POSSIBLE IMPLEMENTATION:
     OPERATION CONSISTENCY LEVEL
     Write Prepare CL.QUORUM
     Write Commit CL.ANY or higher
     First-round Read CL.ANY/CL.ONE
     Second-round Read CL.QUORUM
     

     View Slide

277. RAMP IN CASSANDRA
     To avoid stalling, second-round reads must
     be able to access prepared writes
     POSSIBLE IMPLEMENTATION:
     OPERATION CONSISTENCY LEVEL
     Write Prepare CL.QUORUM
     Write Commit CL.ANY or higher
     First-round Read CL.ANY/CL.ONE
     Second-round Read CL.QUORUM
     

     View Slide

278. RAMP IN CASSANDRA
     POSSIBLE IMPLEMENTATION:
     

     View Slide

279. RAMP IN CASSANDRA
     POSSIBLE IMPLEMENTATION:
     DC1
     

     View Slide

280. RAMP IN CASSANDRA
     POSSIBLE IMPLEMENTATION:
     DC1 DC2
     

     View Slide

281. RAMP IN CASSANDRA
     POSSIBLE IMPLEMENTATION:
     DC1 DC2
     Run algorithms on a per-DC basis, with use of
     CL.LOCAL_QUORUM instead of full CL.QUORUM
     

     View Slide

282. View Slide

283. RAMP TRANSACTIONS:
     

     View Slide

284. RAMP TRANSACTIONS:
     • Provide atomic visibility, as required for
     maintaining FKs, scalable indexing, mat views
     

     View Slide

285. RAMP TRANSACTIONS:
     • Provide atomic visibility, as required for
     maintaining FKs, scalable indexing, mat views
     • Avoid in-place updates, mutual exclusion, any
     synchronous/blocking coordination
     

     View Slide

286. RAMP TRANSACTIONS:
     • Provide atomic visibility, as required for
     maintaining FKs, scalable indexing, mat views
     • Avoid in-place updates, mutual exclusion, any
     synchronous/blocking coordination
     • Use metadata with limited multi versioning,
     reads repair partial writes
     

     View Slide

287. RAMP TRANSACTIONS:
     • Provide atomic visibility, as required for
     maintaining FKs, scalable indexing, mat views
     • Avoid in-place updates, mutual exclusion, any
     synchronous/blocking coordination
     • Use metadata with limited multi versioning,
     reads repair partial writes
     • 1-2RTT overhead, pay only during contention
     

     View Slide

288. RAMP TRANSACTIONS:
     • Provide atomic visibility, as required for
     maintaining FKs, scalable indexing, mat views
     • Avoid in-place updates, mutual exclusion, any
     synchronous/blocking coordination
     • Use metadata with limited multi versioning,
     reads repair partial writes
     • 1-2RTT overhead, pay only during contention
     Thanks!
     http://tiny.cc/ramp-code
     @pbailis
     http://tiny.cc/ramp-intro
     

     View Slide

289. Punk designed by my name is mud from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Queen designed by Bohdan Burmich from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Guy Fawkes designed by Anisha Varghese from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Emperor designed by Simon Child from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Baby designed by Les vieux garçons from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Baby designed by Les vieux garçons from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Gandhi designed by Luis Martins from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Database designed by Anton Outkine from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Girl designed by Rodrigo Vidinich from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Child designed by Gemma Garner from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Customer Service designed by Veysel Kara from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Punk Rocker designed by Simon Child from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Pyramid designed by misirlou from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Person designed by Stefania Bonacasa from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Record designed by Diogo Trindade from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Window designed by Juan Pablo Bravo from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Balloon designed by Julien Deveaux from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Balloon designed by Julien Deveaux from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Balloon designed by Julien Deveaux from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Crying designed by Megan Sheehan from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Sad designed by Megan Sheehan from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Happy designed by Megan Sheehan from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     Happy designed by Megan Sheehan from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     User designed by JM Waideaswaran from the Noun Project Creative Commons – Attribution (CC BY 3.0)
     !
     COCOGOOSE font by ZetaFonts COMMON CREATIVE NON COMMERCIAL USE
     IMAGE/FONT CREDITs
     

     View Slide