Causal Consistency For Large Neo4j Clusters by Jim Webber at Big Data Spain 2017

An overview of the Raft algorithm and how Neo4j uses it to provide strong consistency at scale.

https://www.bigdataspain.org/2017/talk/causal-consistency-for-large-neo4j-clusters

Big Data Spain 2017
November 16th - 17th Kinépolis Madrid

Transcript

2. Causal Consistency For Large Neo4j Clusters Dr. Jim Webber Chief Scientist, Neo4j
5. ads to a social graph
7. Motivation Why do we need Neo4j clusters?
8. Massive Throughput
9. Data Redundancy
10. Data Redundancy
11. Data Redundancy
12. Data Redundancy
13. High Availability
14. High Availability
15. High Availability Error! 503: Service Unavailable
16. High Availability Error! 503: Service Unavailable
17. High Availability Error! 503: Service Unavailable
18. High Availability Error! 503: Service Unavailable
19. High Availability ✓ Error! 503: Service Unavailable
20. Data Massive High
21. Data Massive High 3.0
22. Data Massive High 3.0 Bigger Clusters Consensus Commit Built-in load balancing 3.1 + Causal Clusterin g
23. No Free Lunch* Consistency makes clusters tricky
24. Register Login You need to login in to continue your purchase!
25. Register Login You need to login in to continue your purchase! Username: Password: Create Account
26. Register Login You need to login in to continue your purchase! Username: jim_w Password: ******** Create Account
27. Register Login You need to login in to continue your purchase! Username: Password: Login
28. Username: jim_w Password: ******** Login
29. Purchase Login Successful Try again No account found! Username: jim_w Password: ******** Login
30. Roles for Safety and Scale Divide and conquer complexity
31. Design Trade-off Design Trade-off Availability Reliability
32. Read Replicas Core
33. • Small group of Neo4j databases • Fault-tolerant Consensus Commit • Responsible for data safety Core
34. Writing to the Core Cluster Neo4j Driver Neo4j Cluster
35. Writing to the Core Cluster Neo4j Driver CREATE ( : User {. . . }) ✓ Neo4j Cluster
36. Writing to the Core Cluster Neo4j Driver CREATE ( : User {. . . }) ✓ Neo4j Cluster
37. Writing to the Core Cluster Neo4j Driver CREATE ( : User {. . . }) ✓ ✓ ✓ Neo4j Cluster
38. Writing to the Core Cluster Neo4j Driver CREATE ( : User {. . . }) ✓ ✓ ✓ Neo4j Cluster
39. Writing to the Core Cluster Neo4j Driver CREATE ( : User {. . . }) ✓ ✓ ✓ Neo4j Cluster
40. Writing to the Core Cluster Neo4j Driver ✓ ✓ ✓ Success Neo4j Cluster
41. Writing to the Core Cluster Neo4j Driver ✓ ✓ ✓ Success Neo4j Cluster ✓ ✓
42. Raft Protocol Non-Blocking Consensus for Humans
43. EASY TO UNDERSTAND
44. Raft Protocol https://github.com/ongardie/raftscope
45. Raft in a Nutshell • Raft keeps logs tied together • Logs contain entries for both the data and cluster membership • Entries are appended and subsequently committed if a simple majority agree • Implication: majority agree with the log as proposed • Anyone can call an election: highest term (logical clock) wins, followed by highest committed, followed by highest appended • Appended but uncommitted entries can be truncated, but this is safe (transaction aborted)
46. Consensus Log → Committed Transactions → Updated Graph 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11 12 13 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 Transaction log: the same transactions appear in the same order on all members Consensus log: stores both committed and uncommitted transactions Uncommitt ed entries may differ between members Transactions are only appended to the transaction log when committed according to Raft Transactions are applied, updating the graph Neo4j Raft implementation
47. • Small group of Neo4j databases • Fault-tolerant Consensus Commit • Responsible for data safety Core
48. • For massive query throughput • Read-only replicas • Not involved in Consensus Commit • Disposable, suitable for auto-scaling Read Replicas
49. Propagating updates to the Read Replicas Neo4j Driver Neo4j Cluster
50. Propagating updates to the Read Replicas Neo4j Driver Neo4j Cluster Write
51. Propagating updates to the Read Replicas Neo4j Driver Neo4j Cluster Write
52. Reading from the Read Replicas Neo4j Driver Neo4j Cluster Read
53. Updating the graph Querying the graph
54. Read Repli cas Cor e Updating the graph Queries, analysis, reporting
55. Building an App Where CS meets software eng
56. App Server Neo4j Driver Bolt protocol
57. Java < dependency> < gr oupI d> org. neo4j . dri ver< / gr oupI d> < art i f act I d> neo4j -j ava-dri ver< / art i f act I d> < / dependency> Python pi p i nst al l neo4j -dri ver .NET PM > I nst al l - Package N eo4j . D ri ver JavaScript npm i nst al l neo4j -dri ver
58. https://neo4j.com/developer/language-guides
59. bolt:// G raphD at abase. dri ver( "bol t : / / aServer" )
60. bolt+routing:// G raphD at abase. dri ver( "bol t + r out i ng: / / aCor eServer" )
61. G raphD at abase. dri ver( "bol t + r out i ng: / / aCor eServer" ) Bootstrap: specify any core server to route load across the bolt+routing://
62. Application Server Neo4j Driver Max Jim Jane Mar k
63. Routed write statements dri ver = G raphD at abase. dri ver( "bol t + r out i ng: / / aCor eServer" ) ; t ry ( Sessi on sessi on = dri ver. sessi on( AccessM ode. W RI TE ) ) { t ry ( Transact i on t x = sessi on. begi nTransact i on( ) ) { t x. run( "M ERG E ( user : User {userI d: {userI d}}) ", param et ers( "userI d", userI d ) ) ; t x. success( ) ; } }
64. Routed read queries dri ver = G raphD at abase. dri ver( "bol t + r out i ng: / / aCor eServer" ) ; t ry ( Sessi on sessi on = dri ver. sessi on( AccessM ode. READ ) ) { t ry ( Transact i on t x = sessi on. begi nTransact i on( ) ) { t x. run( "M ATCH ( user : User {userI d: {userI d}}) - [ *] - ( : Pr oduct ) RETURN *", param et ers( "userI d", userI d ) ) ; t x. success( ) ; } }
65. Consistency Models Can you read what you write?
66. Cluster members slightly “ahead” or “behind” of each other 0 1 2 3 4 5 6 7 8 9 10 1 1 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 If I query this server I won’t see the updates from transaction . If I query this server, I’ll see all updates from all committed transactions 1 1 1 1 This is normal behaviour
67. Register Login You need to login in to continue your purchase!
68. Register Login You need to login in to continue your purchase! Username: Password: Create Account
69. Register Login You need to login in to continue your purchase! Username: jim_w Password: ******** Create Account
70. Register Login You need to login in to continue your purchase! Username: Password: Login
71. Username: jim_w Password: ******** Login
72. Purchase Login Successful Try again No account found! Username: jim_w Password: ******** Login
73. Username: jim_w Password: ******** A few moments later... ✓ Login
74. Purchase Login Successful Username: jim_w Password: ******** Login A few moments later... ✓
75. Q Why didn’t this work? A Eventual Consistency
76. 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 Create Account App Serve r A Drive r
77. 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r
78. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r
79. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r 11
80. 0 1 2 3 4 5 6 7 8 9 10 11 CREATE ( : User ) Create Account App Serve r A Drive r 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 11
81. 0 1 2 3 4 5 6 7 8 9 10 11 CREATE ( : User ) Create Account App Serve r A Drive r 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 11
82. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r 11
83. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r M ATCH ( : User) Login App Serve r B Drive r 11
84. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r M ATCH ( : User) Login App Serve r B Drive r 11
85. Bookmark Session token String (for portability) Opaque to application Represents ultimate user’s most recent view of the graph More capabilities to come
86. Let’s try again, with Causal Consistency
87. 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 Create Account App Serve r A Drive r
88. 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r
89. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r
90. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r 11
91. 0 1 2 3 4 5 6 7 8 9 10 11 CREATE ( : User ) Create Account App Serve r A Drive r 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 11
92. 0 1 2 3 4 5 6 7 8 9 10 11 CREATE ( : User ) Create Account App Serve r A Drive r 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 11
93. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r 11
94. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 CREATE ( : User ) Create Account App Serve r A Drive r M ATCH ( : User) Login App Serve r B Drive r 11
95. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 CREATE ( : User ) Create Account M ATCH ( : User) Login App Serve r A App Serve r B Drive r Drive r
96. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 CREATE ( : User ) Create Account M ATCH ( : User) Login App Serve r A App Serve r B Drive r Drive r 11
97. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 CREATE ( : User ) Create Account M ATCH ( : User) Login App Serve r A App Serve r B Drive r Drive r 11
98. Obtain bookmark t ry ( Sessi on sessi on = dri ver. sessi on( AccessM ode. W RI TE ) ) { t ry ( Transact i on t x = sessi on. begi nTransact i on( ) ) { t x. run( "CREATE ( user : User {userI d: {userI d}, passw or dHash: {passw or dHash}) ", param et ers( "userI d", userI d, "passw or dH ash", passw or dH ash ) ); t x. success( ) ; } St ri ng bookm ark = sessi on. l ast Bookm ark( ) ; }
99. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 CREATE ( : User ) Create Account M ATCH ( : User) Login App Serve r A App Serve r B Drive r Drive r 11 Obtain bookmark
100. Use a bookmark t ry ( Sessi on sessi on = dri ver. sessi on( AccessM ode. READ ) ) { t ry ( Transact i on t x = sessi on. begi nTransact i on( bookm ark ) ) { t x. run( "M ATCH ( user : User {userI d: {userI d}}) RETURN *", param et ers( "userI d", userI d ) ) ; t x. success( ) ; } }
101. 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 CREATE ( : User ) Create Account M ATCH ( : User) Login App Serve r A App Serve r B Drive r Drive r 11 Use bookmark
102. Thank you for listening @jimwebber