James Nugent - Event Store Internals

Transcript

1. Dev Day Riga 21 January 2015 James Nugent @jen20
2. The Plan • Event Sourcing & CQRS • Event Store Internals
3. Event Sourcing & CQRS
4. The domain model pattern
5. ORMs and the correctness problem
6. State changes should be an explicit part of our model, and should have names
7. Line Items x n Line Items x n Shopping Cart Line Items x n Shipping Information
8. Shopping Cart Created Added 2 shirts Added 5 pairs of socks Removed 1 shirt Checked Out
9. We can persist our domain model by persisting just the events instead of current state!
10. Current state
    is a
    left fold
    of
    previous behaviours
11. Current state
    is a
    left fold
    of
    previous behaviours
12. Since we can rebuild current state from a stream of events, storing it is just an optimisation
13. We can project any structural model from our events by applying a different fold operation
14. Testing
15. Testing • At the Aggregate level: • Given past behaviours

16. Testing • At the Aggregate level: • Given past behaviours • When method is called

17. Testing • At the Aggregate level: • Given past behaviours • When method is called • Then events
18. Testing • At the Aggregate level: • Given past behaviours • When method is called • Then events • And nothing else!

19. Example Tests [Test] public void MyTest() { new CommandScenarioFor<MilesAccount>(MilesAccount.Factory) .Given(new MilesAccountCreated(...), new MilesAddedToAccount(...), new TierPointsAddedToAccount(...)) .When(sut => sut.RecordFlightTaken(...)) .Then(new MilesAddedToAccount(...), new CustomerPromotedToGoldStatus(...), new TierPointPeriodReset(...)) .Assert(); }
20. Example Test [Test] public void ReviewApproval() { new CommandScenarioFor<Review>(Review.Factory) .Given(new ReviewSubmitted(...), new ReviewSentForModeration(...), new ReviewApprovedByModerators(...), new ReviewDisplayOnWebApproved(...)) .When(sut => sut.RecordComplaintFromSubject(...)) .Then(new ReviewDisplayOnWebSuspended(...), new ReviewManualArbitrationProcessRequested(...)) .Assert(); }
21. Querying
22. Command - Query Responsibility Segregation
23. Service reads one way and writes another way.
24. It cannot be that one model is optimal for transaction processing and all types of queries
25. Construct read models by applying a left fold to the events produced by the write model
26. Many Read Models Transaction Log Document DB Graph DB SQL Star Schema
27. Event Store Internals
28. What is Event Store?
29. • Implements a transaction log for event sourced systems • Supports three basic operations: • Append to a stream • Read all or part of a stream • Subscribe to a stream
30. TCP, HTTP and Direct interfaces
31. Over HTTP, streams are exposed as Atom feeds
32. Fully asynchronous TCP protocol using length- framed serialised protocol buffers
33. Liberally licensed open source software! https://github.com/EventStore/EventStore
34. Overall Architecture
35. Desirable Properties of a Concurrent Server
36. Be modular enough that replacing a component doesn’t mean “the” big rewrite
37. Support lots of clients
38. Degrade gracefully under load
39. Operations Friendly
40. Support self-tuning
41. Traditional ways of writing concurrent servers
42. Thread-based
43. Request N Network Dispatcher Network Request 1 Request 2 Request 3 … dispatch send
44. Request N Network Dispatcher Network Request 1 Request 2 Request 3 … dispatch send
46. Event-driven
47. Request N FSM Scheduler Network Request 1 FSM Request 2 FSM Request 3 FSM … events Disk
48. Request N FSM Scheduler Network Request 1 FSM Request 2 FSM Request 3 FSM … events Disk
50. SEDA
51. Stage
52. Event Handler
53. Stage Inbound Events Outbound Events
54. TCP Protocol HTTP Authorization Disk Writer Statistics Disk Reader Request Mgr
55. (GET /streams/myStream/20/10) TCP Protocol HTTP Authorization Disk Writer Statistics Disk Reader Request Mgr
56. ReadStreamForward TCP Protocol HTTP Authorization Disk Writer Statistics Disk Reader Request Mgr
57. ReadEvents TCP Protocol HTTP Authorization Disk Writer Statistics Disk Reader Request Mgr
58. ReadEventsCompleted TCP Protocol HTTP Authorization Disk Writer Statistics Disk Reader Request Mgr
59. AuthoriseRequest TCP Protocol HTTP Authorization Disk Writer Statistics Disk Reader Request Mgr
60. AuthorisationDenied TCP Protocol HTTP Authorization Disk Writer Statistics Disk Reader Request Mgr
61. ReadStreamForwardCompleted TCP Protocol HTTP Authorization Disk Writer Statistics Disk Reader Request Mgr
62. 403 Forbidden TCP Protocol HTTP Authorization Disk Writer Statistics Disk Reader Request Mgr
63. Stages may only communicate by passing messages via a queue
64. Continuations are explicit
65. Messages are usually immutable
66. Messages are processed by stages in the order of arrival
67. Threads are the main source of concurrency
68. We have a limited number of threads per stage rather than one- per-client
69. Blocking calls should be minimised
70. Amortise expensive operations across requests
71. Amortise expensive operations across requests (“Smart Batching”)
72. Monitor at the boundaries between stages
74. “But this is all really complicated and my juniors will never understand it all!” — Anonymous
76. TCP Protocol = Journal Replicate Business Logic Deserialize
77. Important Code •Message •Envelope et al •QueuedHandler •MultiQueuedHandler •InMemoryBus
78. Storage
79. Transactions are appended to a log file
80. Actually written in chunks
81. Log file consists of prepare, commit and system records
82. public abstract class LogRecord { public static readonly byte[] NoData = Empty.ByteArray; public readonly LogRecordType RecordType; public readonly byte Version; public readonly long LogPosition; public class PrepareLogRecord: LogRecord, IEquatable<PrepareLogRecord> { public const byte PrepareRecordVersion = 0; public readonly PrepareFlags Flags; public readonly long TransactionPosition; public readonly int TransactionOffset; public readonly int ExpectedVersion; public readonly string EventStreamId; public readonly Guid EventId; public readonly Guid CorrelationId; public readonly DateTime TimeStamp; public readonly string EventType; public readonly byte[] Data; public readonly byte[] Metadata;
83. Length Record Data Length Position
84. A writer checkpoint keeps track of where we have flushed
85. Immutable files
86. Scavenge for garbage collection
87. chunk-0.0 chunk-1.0 chunk-2.0
88. chunk-0.0 chunk-1.0 chunk-2.0 chunk-0.1
89. Important Code •TFChunkDB •LogRecord (and derivatives) •TFChunk •TFChunkReader •TFChunkWriter
90. Index
91. Only ONE index
92. Index entries are 16 bytes each
93. Index Entries • Hash of stream name • Event number • Logical position within storage
94. Starts off in-memory
96. Memtable (up to 1m)
97. Memtable (up to 1m) P-Table (up to 1m)
98. Memtable (up to 1m) P-Table (up to 1m) CP
99. What happens as the number of P-Tables grows?
100. P-Table (1m entries) Memtable (1m entries)
101. P-Table (1m entries) P-Table (1m entries) Memtable (1m entries)
102. P-Table (1m entries) P-Table (1m entries) P-Table (1m entries) Memtable (1m entries)
103. Merge 4 PTables together
104. P-Table (1m entries) P-Table (1m entries) P-Table (1m entries) Memtable (1m entries) P-Table (4m entries)
105. Memtable (1m entries) P-Table (4m entries)
106. P-Table (1m entries) Memtable (1m entries) P-Table (4m entries)
107. P-Table (1m entries) P-Table (1m entries) Memtable (1m entries) P-Table (4m entries)
108. P-Table (1m entries) P-Table (1m entries) P-Table (1m entries) Memtable (1m entries) P-Table (4m entries)
109. Merge
110. P-Table (1m entries) P-Table (1m entries) P-Table (1m entries) Memtable (1m entries) P-Table (4m entries) P-Table (4m entries)
111. Memtable (1m entries) P-Table (4m entries) P-Table (4m entries)
112. Memtable (1m entries) P-Table (4m entries) P-Table (4m entries) P-Table (4m entries)
113. Merge
114. Memtable (1m entries) P-Table (16m entries)
115. Important Code •PTable •IndexEntry •MergeResult •IndexMap •TableIndex •Murmur3AUnsafe
116. Network Protocols
117. High Availability Clustering
118. HA • Fully consistent quorum-based replication • No shared storage requirements • Works best with an odd number of servers • Nodes use Paxos to elect a master
119. Discovery • Nodes discover each other via DNS • One DNS entry points to each node • Nodes gossip amongst themselves • Client can discover nodes via DNS
120. HA Client Slave Node Slave Node Master Node
121. HA Client Slave Node Slave Node Master Node X Starts Election
122. HA Client Slave Node Slave Node Master Node X Starts Election
123. HA Client Master Node Slave Node Master Node X Starts Election
124. HA Client Master Node Slave Node Master Node X Starts Election
125. Important Code •MasterReplicationService •ReplicaService •ClusterVNodeController •ClusterStorageWriterService •ElectionService
126. More Information • Event Store: http://geteventstore.com • Github: http://github.com/EventStore/EventStore • Twitter: @jen20, @GetEventStore • Docs: http://github.com/EventStore/EventStore/wiki
127. If you liked this talk, I’m @jen20 on twitter. If you didn’t like this talk, I’m @stack72 on twitter.
128. If you liked this talk, I’m @jen20 on Twitter. If you didn’t like this talk, I’m @sbellware on Twitter.
129. We’re Hiring!
130. Event Store as a Read Model
131. Temporal Query • I as a doctor want to find patients for my clinical trial who: • Were diagnosed with condition Y inside the last 6 months • Had treatment A at least twice but no more than four times • Had an abnormal blood test where the white blood cell count was more than 10% lower than before treatment
132. Temporal Query SELECT * FROM …. WHERE x IN (SELECT * FROM … WHERE y IN (SELECT * FROM … WHERE z IN (SELECT * FROM … WHERE λ…))))
133. This gets very ugly very fast.
134. Projections can query historical data and then switch to continue running into the future
135. Projections use JavaScript to configure a left fold.
136. transform(h(g(f(seed, e0), e1), e2))
137. Code
138. Everything is a stream.
139. Sentiment Analysis
140. Sentiment Analysis (sort of)
141. Query • I as someone who doesn’t live in San Francisco want to find users who said the word “starbucks” and the word “happy” in proximity to one another across any chat rooms they are in, so I know who not to ask for advice about coffee
142. Code
143. Indexing
144. Indexing … … chat-GeneralChat user-James user-Rob
145. linkTo(‘newStream', anEvent);
146. Code
147. Using this can dramatically reduce the size of your code base
148. Global Projection State
149. Data locality matters
150. Expressing your data model in the language of the domain is powerful