Abstract

Geo-distributed DBMSes often do not support strictly-serializable transactions at all, or limit transactions to only a single region. Those that offer this support force the programmer to either give up low-latency regional writes or use a restricted programming model. We show that these compromises are no longer necessary in modern datacenters for workloads with regional locality, thanks to advances in low-latency datacenter networks and DBMS designs. We present Chablis, a scalable, geo-distributed, multi-versioned transactional key-value store. Chablis provides a general interface supporting range and point reads, as well as writes within multi-step strictly-serializable ACID transactions. It offers fast (i.e., low latency) read-write transactions accessing data homed in a single region, and global strictly-serializable lock-free snapshot reads.

Recognition

Coverage

  1. Columbia Engineering EE Professor Asaf Cidon, CS Professor Junfeng Yang, and Their Team Win the Best Paper Award at CIDR 2024
  2. Murat Demirbas Chardonnay: Fast and General Datacenter Transactions for On-Disk Databases (OSDI'23)