sdcons
is an implementation of geo-replicated distributed consensus algorithm. It is based on the paper:
SDPaxos: Building Efficient Semi-Decentralized Geo-replicated State Machines (ACM Symposium on Cloud Computing 2018, SoCC '18)
Like original SDPaxos, sdcons
provides one-round-trip latency reading and writing capabilities under realistic configurations (deployed with three or five nodes).
In addition, I ported some optimizations mentioned in the raft algorithm to SDPaxos. Including:
- log term
- leader election
- leadership transfering
- membership configuration (joint consensus)
- batched & pipelined log replication
- lease reading
These optimizations make SDPaxos easier to implement and verify.
However, the current sdcons
has not yet implemented the Straggler detection mentioned in the SDPaxos paper.
Unlike SDPaxos, sdcons
uses the term channel
to represent the paxos commiting progress. The Ordering instance is represented as Index, the Command instance is represented as Entry. Because the term index is used to represent Ordering instance, I choose term id to represent the index in raft.
At present, sdcons has not performed benchmarking and related performance optimization.
See examples/
.
Here is some improvements need to code:
- Applying committed entries without persistence
Ready
add a field to describe the future working- Support asynchornous log entries persisting
Thanks etcd for providing the amazing Go implementation of raft!