Pineapple: Unifying Multi-Paxos and Atomic Shared Registers
summary
Pineapple implements a unified Multi-Paxos plus atomic-register design that uses logical timestamps (tag and slot) so any replica can serve reads and blind writes while Multi-Paxos handles one-shot transactions. It offloads reads/writes to followers, achieves one-RTT reads for n=3 and nonblocking execution, incurs one extra RTT for conditional writes, and integrates with etcd. Evaluation against Gryff, Multi-Paxos/Raft and EPaxos shows higher throughput and lower tail latency (median latency reduced >50% vs optimized Raft in a balanced workload).
AI-written summary · about
chapters (5)
- 0:00Pineapple introduction and motivation
- 3:37Improving consensus performance
- 8:59Pineapple design and architecture
- 12:11Pineapple timestamp and protocol
- 16:23Evaluation results and comparisons
related talks
20mGryff: Unifying Consensus and Shared Registers
16mSwiftPaxos: Fast Geo-Replicated State Machines
12mEPaxos Revisited
20mHigh Throughput Replication with Integrated Membership Management
47mRiak Ensemble A Paxos based Framework for Scalable, Consistent Systems
32mWas Leslie Lamport Right?
01h 13mPaxos vs Raft: Have we reached consensus on distributed consensus?
22mLiberating Distributed Consensus