DEPART: Replica Decoupling for Distributed Key-Value Storage

on YouTube ↗16m

" summary

DEPART implements replica decoupling by storing primaries in an LSM and redundant copies in a separate two-layer log with tunable ordering, implemented as a prototype DEPART atop Cassandra. The system claims up to 1.43x write, 2.43x read, 2.68x scan and 1.44x update throughput under eventual consistency, reduces recovery time by about 38–54%, and lets operators trade read versus write performance by changing the log ordering parameter.

AI-written summary · about

" chapters (4)

1. 0:00Start
2. 3:11Uniform Indexing
3. 7:24How To Manage Redundant Copies
4. 10:30How To Perform Recovery Efficiently

" related talks

• 24mStronger Semantics for Low-Latency Geo-Replicated StorageMichael J. Freedman, Wyatt Lloyd, Michael Kaminsky, David G. Andersen · NSDI 2013
• 14mA Cloud-native Architecture for Replicated Data ServicesHemant Saxena, Jeffrey Pound · HotCloud '20
• 20mFine Grained Replicated State Machines for a Cluster Storage SystemMing Liu, Harsha V. Madhyastha, Karan Gupta, Arvind Krishnamurthy, Rishi Bhardwaj, Chinmay Kamat, Huapeng Yuan, Aditya Jaltade, Roger Liao, Pavan Konka, Anoop Jawahar · NSDI 2020
• 15mMitigating Resource Usage Dependency in Sorting-based KV Stores on Hybrid...Qingyang Zhang, Yongkun Li, Yubiao Pan, Haoting Tang, Yinlong Xu · USENIX ATC '25
• 43mPartitions for Everyone!Kyle Kingsbury · Strange Loop 2013
• 14mLoLKV: The Logless, Linearizable, RDMA-based Key-Value Storage SystemAhmed Alquraan, Sreeharsha Udayashankar, Virendra Marathe, Bernard Wong, Samer Al-Kiswany · 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24)
• 52mThe Computer Science behind a modern distributed data storeMax Neunhöffer · Devoxx Belgium 2017
• 43mWhy We Built Our Own Distributed Column StoreSam Stokes · Strange Loop 2017