Stronger Semantics for Low-Latency Geo-Replicated Storage
summary
A geo-replicated storage design that guarantees local, single-data-center service and implements causal consistency for column families, using Lamport-style logical timestamps to serve consistent snapshots. Prototype forks Cassandra with ~5k added Java lines and publishes code; it delivers single-ms local latency, read-only and write-only transactions that require at most two local parallel rounds, and measures about ~6.5% throughput overhead versus eventually-consistent Cassandra while scaling near-linearly (avg 96% growth up to 128-server clusters).
AI-written summary · about
chapters (16)
- 0:00Eiger: Stronger Semantics for Low-Latency Geo-Replicated Storage
- 0:45Storage Dimensions
- 1:45Sharded, Geo-Replicated Storage
- 2:14Strong Consistency or Low Latency
- 4:57Eiger Ensures Low Latency Keep All Ops Local
- 6:23Causal For Column Families
- 7:17Viewing Data Consistently Is Hard Asynchronous requests distributed data - ?????
- 11:27Transaction Intuition
- 12:28Eiger Provides
- 12:58Eiger Implementation • Fork of open-source Cassandra
- 13:11Evaluation
- 13:50Experimental Setup
- 14:16Facebook Workload Results
- 14:49Eiger Scales Facebook Workload
- 15:52Improving Low-Latency Storage
- 17:16Eiger • Low-latency geo-replicated storage - Causalt for column families -Read-only transactions - Write-only transactions
related talks
16mDEPART: Replica Decoupling for Distributed Key-Value Storage
20mFine Grained Replicated State Machines for a Cluster Storage System
33mGoogle Cloud Spanner: Global consistency at scale
19mBuilding a Performant Configuration-Store Layer
44mAvailability, Latency and Cost: Withstanding Outages
01h 05mRoundtable: Cloud Databases as a Service
01h 15mCAP Theorem — two decades and few clouds later
47mGoogle Cloud Spanner: Global consistency at scale