[Paper Review] CRDTs: Consistency without concurrency control

Mihai Leţia, Nuno Preguiça+1 more|ArXiv.org|Jul 6, 2009

Distributed systems and fault tolerance13 references36 citations

TL;DR

This paper introduces CRDTs (Commutative Replicated Data Types) as a scalable alternative to traditional concurrency control in distributed systems, enabling eventual consistency without consensus. It presents Treedoc, a practical CRDT for concurrent document editing that uses a naming tree to ensure commutative operations, compact identifiers, and efficient causal ordering, achieving high performance and scalability in large-scale, dynamic environments with minimal coordination overhead.

ABSTRACT

A CRDT is a data type whose operations commute when they are concurrent. Replicas of a CRDT eventually converge without any complex concurrency control. As an existence proof, we exhibit a non-trivial CRDT: a shared edit buffer called Treedoc. We outline the design, implementation and performance of Treedoc. We discuss how the CRDT concept can be generalised, and its limitations.

Motivation & Objective

• To address the scalability and consistency challenges in large-scale distributed systems with mutable shared data.
• To eliminate the need for complex concurrency control mechanisms like consensus or vector clocks in replicated data types.
• To design a practical, efficient, and scalable CRDT that supports concurrent insertions and deletions in an ordered sequence.
• To overcome practical limitations such as indefinite growth, identifier bloat, and garbage collection in CRDTs.
• To demonstrate that CRDTs can be used in real-world systems like collaborative editing with performance comparable to production systems.

Proposed method

• Designing a CRDT abstraction for an ordered set with insert-at-position and delete operations using unique, stable identifiers.
• Using a naming tree structure to assign compact, globally unique identifiers to atoms, ensuring commutativity across replicas.
• Implementing a two-tier architecture with core and nebula sites: core maintains the global state, nebula sites handle local updates and propagate them via a reliable broadcast protocol.
• Applying a tree restructuring algorithm that preserves commutativity by transforming subtrees into major nodes when needed.
• Generating update operations (insert/delete) only for black nodes and tombstones in the final tree, ensuring idempotent and commutative propagation.
• Using implicit causal ordering encoded in the tree structure to avoid the need for precise vector clocks, enabling scalable approximate causal tracking.

Experimental results

Research questions

• RQ1 Can a non-trivial, practical, and efficient CRDT be designed for concurrent editing of ordered sequences?
• RQ2 How can CRDTs be made scalable in systems with dynamic, varying numbers of writable replicas?
• RQ3 What mechanisms ensure commutativity of operations without relying on consensus or vector clocks?
• RQ4 How can identifier size and data structure bloat be minimized in CRDTs while preserving correctness?
• RQ5 Can garbage collection be decoupled from the critical path of application updates in CRDTs?

Key findings

• Treedoc successfully implements a CRDT for concurrent document editing that ensures eventual convergence across replicas without consensus or complex conflict resolution.
• The use of a naming tree enables compact, globally unique identifiers that remain stable across flattening operations, preventing identifier explosion.
• Tree restructuring is performed in a way that maintains commutativity, allowing dynamic scaling without breaking consistency.
• Causal ordering is implicitly encoded in the tree structure, reducing reliance on expensive vector clocks and enabling scalable approximate tracking.
• Garbage collection is feasible and non-blocking, as it operates outside the critical path of application updates, though it requires consensus.
• Performance evaluation using Wikipedia traces shows that Treedoc scales well and performs competitively with production systems, even under high update loads.