The Impossibility of Fast Transactions

Abstract

In this paper, we prove that transactions cannot be fast in an asynchronous system. Specifically, we show that a system cannot be fault-tolerant and provide fast transactions. Our result holds in any system where we require transactions to ensure monotonic writes, or any stronger consistency model, such as, causal consistency. Thus, our result unveils an important, and so far unknown, limitation of fast transactions: they are impossible if we want to tolerate the failure of even one server.

NOC-NOC: Towards Performance-optimal Distributed Transactions

“…• RAMP-F and RAMP-S [5], which are the two original, yet state-of-the-art, read-atomic algorithms; • the RAMP-OPW design [5], which optimizes RAMP with one-phase writes while sacrificing the read-your-writes session property (thus providing a weaker isolation guarantee than RA-NOC2); • LORA [27], which is a SNOW-optimal read-atomic algorithm, missing only C 𝑅 and C 𝑊 ; and • the NOCS-optimal algorithm RA-NOC, which we have designed following the NOCS design objective. 3 We do not consider RAMP-H since its performance lies between that of RAMP-F and RAMP-S [5].…”

“…Design Objectives. Many performance criteria have been proposed for designing highly efficient distributed transactions [3,14,15,24,32,34,49]. These criteria focus on optimizing reads.…”

“…Some design objectives have stronger data freshness requirements (e.g., minimal progress [14]), thus restricting the achievable combinations of performance criteria [3,14,15,49]. NOC-NOC assumes a weaker freshness criterion, as for SNOW and NOCS, which allows returning stale snapshots.…”

“…Substantial research efforts [3,14,15,24,32,34,49] have been devoted to studying performance optimality for distributed transactions with respect to isolation levels. Two representative results are the recent SNOW [32] and NOCS [34] theorems.…”