Causal Consistency at Global Scale: A Practical Re-evaluation

Abstract

Causal consistency has long been proposed as a sweet spot between availability and consistency for geo-replicated systems, yet adoption remains limited. We present the first large-scale empirical study of causal consistency in production, analyzing 18 months of operation across three deployments spanning four continents. We identify the specific workload patterns under which causal protocols degrade, and propose two protocol refinements — lazy metadata compaction and speculative dependency resolution — that reduce tail latency by 3.1x without sacrificing safety guarantees. Our refinements have been implemented in two open-source systems and are in production use.

1. Introduction

The CAP theorem establishes that no distributed system can simultaneously provide consistency, availability, and partition tolerance. Causal consistency has been positioned as a practical compromise: it preserves causality (a stronger property than eventual consistency) while remaining available under partitions. Despite a decade of academic interest, production adoption has been sparse.

This paper investigates why. We instrument three production deployments — a social network, a collaborative document editor, and a financial ledger — and collect detailed traces over 18 months.

2. Background

Causal consistency requires that operations related by causality appear in the same order at all replicas, while concurrent operations may appear in different orders. The classical implementation uses vector clocks, which grow linearly with the number of clients and become impractical at scale.