A study of CRDTs that do computations

Abstract

A CRDT is a data type specially designed to allow instances to be replicated and modified without coordination, while providing an automatic mechanism to merge concurrent updates that guarantees eventual consistency. In this paper we present a brief study of computational CRDTs, a class of CRDTs whose state is the result of a computation over the executed updates. We propose three generic designs that reduce the amount of information that each replica maintains and propagates for synchronizations. For each of the designs, we discuss the properties that the function being computed needs to satisfy.

Non-uniform Replication

“…In this section we evaluate our data types that follow the non-uniform replication model. To this end, we compare our designs against state-of-the-art CRDT alternatives: delta-based CRDTs [1] that maintain full object replicas efficiently by propagating updates as deltas of the state; and computational CRDTs [19] that maintain non-uniform replicas using a state-based approach.…”

“…Computational CRDTs [19] are an extension of state-based CRDTs where the state of the object is the result of a computation (e.g. the average, the top-K elements) over the executed updates.…”

“…Unlike most operation-based CRDT designs, we do not assume that the system propagates operations in a causal order. These designs were inspired by the state-based computational CRDTs proposed by Navalho et al [19], which also allow replicas to diverge in their quiescent state.…”

“…A possible use of this data type would be for maintaining the summary of classifications given by users in an online shop. Similar approaches could be implemented for data types that return the result of other aggregation functions that can be incrementally computed [19].…”

“…We begin by comparing our Top-K design (NuCRDT ) with a delta-based CRDT set [1] (Delta CRDT ) and the top-K state-based computational CRDT design [19] (CCRDT ).…”