Building a collaborative authoring experience with CRDTs

Do you intend to provide a collaborative and distributed authoring experience to your users?

Join our speaker Alexandre Piveteau to learn the fundamentals, key properties, and use cases of CRDTs, data structures that automatically resolve synchronization conflicts.

Discover Markdown Party, Alexandre's peer-to-peer collaborative web text editor, which is highly available, eventually consistent, and enables its users to work concurrently on a tree of documents.

Learn how to adapt and design CRDTs that match your application domain-specific rules.

Transcript

1. Building a collaborative authoring experience with CRDTs AEM Engineering Coffee Alexandre Piveteau 8 November 2022 1
4. OUTLINE Introduction Operational Transformation and Conflict-free Replicated Data Types Replicated trees and sequences Summary 4
5. Introduction
6. REQUIREMENTS Functional • Web client • Unique project-specific URL for sharing • Tree of Markdown text documents Non-functional • Edits possible when the network is partitioned • Distributed protocol • Deleting unnecessary meta-data 5
7. CHALLENGES Implementing a collaborative text editor that: 1. is partition-tolerant 2. eventually converges 3. preserves user intent 6
8. CHALLENGES Implementing a collaborative text editor that: 1. is partition-tolerant 2. eventually converges 3. preserves user intent 7
9. PARTITION TOLERANCE – TOPOLOGY Figure 1: Centralized Figure 2: Tree-struct. Figure 3: Distributed 8
10. PARTITION TOLERANCE – TOPOLOGY Figure 1: Centralized Figure 2: Tree-struct. Figure 3: Distributed 8
11. CHALLENGES Implementing a collaborative text editor that: 1. is partition-tolerant 2. eventually converges 3. preserves user intent 9
12. EVENTUAL CONVERGENCE Initial text : Thanks … Local edits • Site 1 : Thanks Alice, • Site 2 : Thanks Bob, 10
13. EVENTUAL CONVERGENCE Initial text : Thanks … Local edits • Site 1 : Thanks Alice, • Site 2 : Thanks Bob, 10
14. EVENTUAL CONVERGENCE Initial text : Thanks … Local edits • Site 1 : Thanks Alice, • Site 2 : Thanks Bob, 10
15. EVENTUAL CONVERGENCE Initial text : Thanks … Local edits • Site 1 : Thanks Alice, • Site 2 : Thanks Bob, After merge • Site 1 : Thanks Alice, Bob, • Site 2 : Thanks Alice, Bob, 10
16. EVENTUAL CONVERGENCE Initial text : Thanks … Local edits • Site 1 : Thanks Alice, • Site 2 : Thanks Bob, After merge • Site 1 : Thanks Alice, Bob, • Site 2 : Thanks Bob, Alice, 10
17. CHALLENGES Implementing a collaborative text editor that: 1. is partition-tolerant 2. eventually converges 3. preserves user intent 11
18. PRESERVING USER INTENT Concurrent edits • Site 1 : Thanks Alice, Bob, – del. Alice@[7:13] • Site 2 : Thanks Alice, Bob, – del. Alice@[7:13] Possible results • Expected : “Thanks Bob,” • Unexpected : “Thanks” – range [7:13] removed twice 12
19. CHALLENGES Implementing a collaborative text editor that: 1. is partition-tolerant 2. eventually converges 3. preserves user intent 13
20. Operational Transformation and Conflict-free Replicated Data Types
21. CAP THEOREM1 “Any distributed data store can provide only two of the following three guarantees: • Consistency: all participants see the most recent data; • Availability: every request receives a response; • Partition-tolerance: the system continues to operate despite an arbitrary number of messages being lost.” 1See https://en.wikipedia.org/wiki/CAP_theorem 14
22. OPERATIONAL TRANSFORMATION Operational transformation • Server transforms operations before forwarding them • Centralized server =⇒ Single point of failure • Notoriously difficult to implement Figure 4: Centralized OT 15
23. CONFLICT-FREE REPLICATED DATA TYPES Characterizing CRDTs • Exist as CvRDTs and CmRDTs • CvRDTs have a function m(s1,s2) with the 3 following properties : • Associativity – m(m(a,b),c) = m(a,m(b,c)) • Commutativity – m(a,b) = m(b,a) • Idempotence – m(a,a) = a • All states si form a semi-lattice • CRDTs are composable 16
24. CvRDT – EXAMPLE Figure 5: Grow-only set (GSet) example 17
25. CvRDT – EXAMPLE Figure 5: Grow-only set (GSet) example 17
26. CvRDT – EXAMPLE Figure 5: Grow-only set (GSet) example 17
27. CvRDT – EXAMPLE Figure 5: Grow-only set (GSet) example 17
28. CvRDT – EXAMPLE Figure 5: Grow-only set (GSet) example 17
29. How can we model a generic CvRDT? 18
30. REPLICATED EVENT LOG • Operations are totally ordered • State is computed by applying operations • Deltas are stored to revert operations Figure 6: ”Accordion-like” movement to insert operations 19
31. REPLICATED EVENT LOG Totally causally ordered log Each event has a globally unique Lamport (site, seqno) timestamp. • Timestamps define a total order for all events • All events respect causal ordering • The log is just a set (GSet) of timestamped events =⇒ it’s a CvRDT! 20
32. REPLICATED EVENT LOG Totally causally ordered log Each event has a globally unique Lamport (site, seqno) timestamp. • Timestamps define a total order for all events • All events respect causal ordering • The log is just a set (GSet) of timestamped events =⇒ it’s a CvRDT! 20
33. REPLICATION ALGORITHM Log replication 1. A sends the list of all known participants Ii . 2. B sends, for each participant Ii , the next expected sequence number Si and the number Ni ≥ 0 of expected events. 3. While Ni > 0, A sends the events (Ii,Sj ) with Sj > Si and Sj minimal, decrements N and sets Si = Sj . 4. Replication stops when A or B disconnects. 21
34. Replicated trees and sequences
35. Demo – Replicated trees 22
36. TREES – OPERATIONS Type Arguments Create text file – Create folder – Move File: file op. ref. To Parent: file op. ref. Set name File: file op. ref. Name: text Delete File: file op. ref. Table 1: Operations supported by replicated trees 23
37. TREES – CONCURRENT UPDATES Figure 7: Conflicting concurrent Moves Locally valid operations may produce graphs that don’t respect tree invariants. 24
38. TREES – CONCURRENT UPDATES Figure 7: Invalid operations are ignored Locally valid operations may produce graphs that don’t respect tree invariants. 24
39. TREES – DELETING ELEMENTS Figure 8: Deleted elements are moved in a special trash can node 25
40. Demo – Text replication 26
41. REPLICATED GROWABLE ARRAY RGAs – operations Type Arguments Insert Prev: text op. ref. Char: letter Remove Id: text op. ref. Table 2: Operations on RGA 27
42. REPLICATED GROWABLE ARRAY RGAs – operations Type Arguments Insert Prev: text op. ref. Char: letter Remove Id: text op. ref. Table 2: Operations on RGA Figure 9: RGA structure 27
43. Playing with timestamps 28
44. HYBRID LOGICAL CLOCKS We can improve concurrent updates using HLCs : • Specialization of Lamport timestamps • Physical time is used to order concurrent events • Happens-before relationship preserved, as with Lamport timestamps • Implemented at the event log level 29
45. COMPRESSION USING RUN-LENGTH ENCODING Site Seq. no Letter 1 1 'a' 1 2 'l' 1 3 'e' 1 4 'x' Table 3: COMPRESSION USING RLE 30
46. COMPRESSION USING RUN-LENGTH ENCODING Site ∆ seq. no Letter 1 (+1) 'a' 1 (+1) 'l' 1 (+1) 'e' 1 (+1) 'x' Table 3: COMPRESSION USING RLE 30
47. COMPRESSION USING RUN-LENGTH ENCODING Site First seq. no Text 1 1 "alex" Table 3: COMPRESSION USING RLE 30
48. COMPRESSION USING RLE – DETAILS • Implemented in Automerge2 • Not compatible with HLC • Especially relevant for text • Fixed-size operations • Most operations are consecutive 2Martin Kleppmann, github.com/automerge 31
49. Technologies used and architecture 32
50. MODULE HIERARCHY Figure 10: Markdown Party components 33
51. Summary
52. SUMMARY • A replicated log can implement any CRDT • Invariants can be enforced by skipping ops • Hybrid logical clocks • Lamport timestamps run-length encoding • Many paths to explore! 34
53. THANK YOU! Me: Alexandre Piveteau (alexandre@piveteau.email) Live demo: https://markdown.party GitHub: https://github.com/markdown-party More about CRDTs: https://crdt.tech 35