Skip to Main content Skip to Navigation

Exploring the design space of highly-available distributed transactions

Abstract : Today's large-scale cloud services must provide fast response and an "always-on" experience. Failing to do so results in reduced user engagement, which directly impacts revenues. Cloud storage replicates data at multiple sites worldwide. A user connects to her closest site and, in the presence of site failures, she fails over to a healthy one. Each site scatters data across a large number of servers to handle load beyond what a single machine can handle. Transactional isolation hides the complexity of distribution from the application logic, thus reduces development efforts. However, enforcing isolation requires mechanisms that impact latency and availability. This thesis studies how to enforce isolation in a cloud environment without affecting availability and responsiveness. Our first contribution is Cure, a transactional protocol that ensures high level of semantics compatible with availability: Transactional Causal Consistency (TCC), and support for Convergent data types (CRDTs). Experimentally, Cure is as scalable as a weakly-consistent protocol, even though it provides stronger semantics. Transactional protocols like Cure exhibit latency overheads that have impeded their adoption at scale. Our second contribution is to explore how to provide isolation with no extra delays with respect to a non-transactional system. In this quest, we find, quantify and demonstrate a three-way trade-off between read isolation, delay (latency), and data freshness. Motivated by the results of the trade-off we propose two isolation properties: TCC- and PSI-. We create three protocols which exhibit minimal delay. The experimental evaluation of these protocols supports the theoretical results.
Complete list of metadata
Contributor : Alejandro Tomsic Connect in order to contact the contributor
Submitted on : Sunday, January 6, 2019 - 11:56:53 PM
Last modification on : Tuesday, March 23, 2021 - 9:28:03 AM
Long-term archiving on: : Sunday, April 7, 2019 - 12:23:12 PM


Files produced by the author(s)


  • HAL Id : tel-01956321, version 1


Alejandro Tomsic. Exploring the design space of highly-available distributed transactions. Databases [cs.DB]. Sorbonne Universites, UPMC University of Paris 6, 2018. English. ⟨tel-01956321v1⟩



Record views


Files downloads