🪵🪵🪵 Raft is not yet good enough. This project intends to improve raft as the next-generation consensus protocol for distributed data storage systems (SQL, NoSQL, KV, Streaming, Graph ... or maybe something more exotic).

Currently, openraft is the consensus engine of meta-service cluster in databend.

• Get started: The guide is the best place to get started, followed by the docs for more in-depth details.

• Openraft FAQ explains some common questions.

• 🙌 Questions? Join the Discord channel or start a discussion.

• Openraft is derived from async-raft with several bugs fixed: Fixed bugs.

Status

• Openraft API is not stable yet. Before 1.0.0, an upgrade may contain incompatible changes. Check our change-log. A commit message starts with a keyword to indicate the modification type of the commit:

  • DataChange: on-disk data types changes, which may require manual upgrade.
  • Change: if it introduces incompatible changes.
  • Feature: if it introduces compatible non-breaking new features.
  • Fix: if it just fixes a bug.

• Branch main has been under active development.

  The main branch is for the 0.8 release.

  • The features are almost complete for building an application.
  • The performance isn't yet fully optimized. Currently, it's about 48,000 writes per second with a single writer.
  • Unit test coverage is 91%.
  • The chaos test is not yet done.

• Branch release-0.8: Latest published: v0.8.6 | Change log v0.8.3 | ⬆️ 0.7 to 0.8 upgrade guide | ⬆️ 0.8.3 to 0.8.4 upgrade guide | release-0.8 Won't accept new features but only bug fixes.

• Branch release-0.7: Latest published: v0.7.6 | Change log v0.7.6 | ⬆️ 0.6 to 0.7 upgrade guide | release-0.7 Won't accept new features but only bug fixes.

• Branch release-0.6: Latest published: v0.6.8 | Change log v0.6 | release-0.6 won't accept new features but only bug fixes.

Roadmap

• 2022-10-31 Extended joint membership

• 2023-02-14 Minimize confliction rate when electing; See: Openraft Vote design; Or use standard raft mode with feature flag single-term-leader.

• 2023-04-26 Goal performance is 1,000,000 put/sec.

• Reduce the complexity of vote and pre-vote: get rid of pre-vote RPC;

• Support flexible quorum, e.g.:Hierarchical Quorums

• Consider introducing read-quorum and write-quorum, improve efficiency with a cluster with an even number of nodes.

Performance

The benchmark is focused on the Openraft framework itself and is run on a minimized store and network. This is NOT a real world application benchmark!!!

Benchmark history:

To access the benchmark, go to the ./cluster_benchmark folder and run make bench_cluster_of_3.

The benchmark is carried out with varying numbers of clients because:

• The 1 client benchmark shows the average latency to commit each log.
• The 64 client benchmark shows the maximum throughput.

The benchmark is conducted with the following settings:

• No network.
• In-memory store.
• A cluster of 3 nodes in a single process on a Mac M1-Max laptop.
• Request: empty
• Response: empty

Features

• It is fully reactive and embraces the async ecosystem. It is driven by actual Raft events taking place in the system as opposed to being driven by a tick operation. Batching of messages during replication is still used whenever possible for maximum throughput.

• Storage and network integration is well defined via two traits RaftStorage & RaftNetwork. This provides applications maximum flexibility in being able to choose their storage and networking mediums.

• All interaction with the Raft node is well defined via a single public Raft type, which is used to spawn the Raft async task, and to interact with that task. The API for this system is clear and concise.

• Log replication is fully pipelined and batched for optimal performance. Log replication also uses a congestion control mechanism to help keep nodes up-to-date as efficiently as possible.

• It fully supports dynamic cluster membership changes with joint config. The buggy single-step membership change algo is not considered. See the dynamic membership chapter in the guide.

• Details on initial cluster formation, and how to effectively do so from an application's perspective, are discussed in the cluster formation chapter in the guide.

• Automatic log compaction with snapshots, as well as snapshot streaming from the leader node to follower nodes is fully supported and configurable.

• The entire code base is instrumented with tracing. This can be used for standard logging, or for distributed tracing, and the verbosity can be statically configured at compile time to completely remove all instrumentation below the configured level.

Who use it

• yuyang0/rrqlite
• raymondshe/matchengine-raft

Contributing

Check out the CONTRIBUTING.md guide for more details on getting started with contributing to this project.

License

Openraft is licensed under the terms of the MIT License or the Apache License 2.0, at your choosing.