libhotstuff is a general-purpose BFT state machine replication library with modularity and simplicity, suitable for building hybrid consensus cryptocurrencies.
This repository includes the prototype implementation evaluated in our HotStuff: BFT Consensus in the Lens of Blockchain paper. The consensus protocol is also used by Facebook in Libra (now rebranded to "Diem") project.
Feel free to contact us if you'd like to reproduce the results in the paper, or tweak the code for your own project/product.
- Full paper: https://arxiv.org/abs/1803.05069
- PODC 2019 paper: https://dl.acm.org/citation.cfm?id=3331591
Although we don't know why, many developers seem to be stuck with those intermediate algorithms in the paper (e.g. Basic HotStuff) that are only for theory/research purpose. As already noted in the paper, we recommend the last algorithm ("Event-Driven HotStuff") as it is simple, elegant (there is no "view" at all in the core protocol and only one generic type of QCs) and very close to an actual implementation. It is also the original algorithm we came up with, before adding in Basic/Chained HotStuff that may help some researchers understand its theoretical relation to PBFT, etc.
TL;DR: read the spec (and the proof) of the algorithm in "Implementation" section, if you're only curious about using HotStuff in your system/work.
- Simplicity. The protocol core logic implementation is as simple as the one
specified in our paper. See
consensus.handconsensus.cpp. - Modular design. You can use abstraction from the lowest level (the core protocol logic) to the highest level (the state machine replication service with network implementation) in your application, or override/redefine the detailed behavior to customize your own consensus.
- Liveness decoupled from safety. The liveness logic is entirely decoupled from safety. By defining your own liveness gadget ("PaceMaker"), you can implement your own liveness heuristics/algorithm. The actual performance varies depending on your liveness implementation, but the safety is always intact.
- Friendly to blockchain systems. A PaceMaker could potentially be PoW-based and it makes it easier to build a hybrid consensus system, or use it at the core of some cryptocurrencies.
- Minimal. The project strives to keep code base small and implement just the basic functionality of state machine replication: to deliver a consistent command sequence to the library user. Application-specific parts are not included, but demonstrated in the demo program.
NOTICE: the project is still in-progress. Try at your own risk, and this section may be incomplete and subject to changes.
# install from the repo git clone https://github.com/hot-stuff/libhotstuff.git cd libhotstuff/ git submodule update --init --recursive # ensure openssl and libevent are installed on your machine, more # specifically, you need: # # CMake >= 3.9 (cmake) # C++14 (g++) # libuv >= 1.10.0 (libuv1-dev) # openssl >= 1.1.0 (libssl-dev) # # on Ubuntu: sudo apt-get install libssl-dev libuv1-dev cmake make cmake -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED=ON -DHOTSTUFF_PROTO_LOG=ON make # start 4 demo replicas with scripts/run_demo.sh # then, start the demo client with scripts/run_demo_client.sh # Fault tolerance: # Try to run the replicas as in run_demo.sh first and then run_demo_client.sh. # Use Ctrl-C to terminate the proposing replica (e.g. replica 0). Leader # rotation will be scheduled. Try to kill and run run_demo_client.sh again, new # commands should still get through (be replicated) once the new leader becomes # stable. Or try the following script: # scripts/faulty_leader_demo.sh
See here.
- Rewrite this minimal code base in Rust: this time, with all the experience, without C++ template kung-fu, I plan to have a ready-to-use, blackbox-like libhotstuff implementation as a full library with better encapsulation and interface. The new goal would be any engineer without knowledge of BFT should be able to use it for his/her own application, without changing the library code. Ping me if you like this re-writing idea or you'd like to be part of it.
- Limit the async event callback depth (otherwise in the demo a fresh replica could overflow its callback stack when trying to catch up)
- Add a PoW-based Pacemaker example
- Branch pruning & swapping (the current implementation stores the entire chain in memory)
- Persistent protocol state (recovery?)