This is a repost of this blog post originally published on the Nomadic Labs blog.
The consensus algorithm is a crucial part of any blockchain project. Because of the distributed nature of blockchains, different nodes can have different ideas of what the current state of the blockchain is supposed to be. The role of the consensus algorithm is to decide which of these possible states, called forks or branches, will be selected globally. In the world of distributed systems, there are two distinctive families of consensus algorithms: Nakamoto-style and BFT-style. Most blockchain solutions use Nakamoto-style algorithms that allow the existence of any number of forks of any length, but make longer forks increasingly unstable, so that they eventually collapse to a single branch. We say that these algorithms have probabilistic finality. Byzantine fault tolerance (BFT) algorithms have deterministic finality. They stipulate definite conditions that must be fulfilled for a block to become final.
Nomadic Labs is intending to propose Tenderbake — a BFT-style algorithm — as the next consensus algorithm of Tezos. Its deterministic finality allows us to make solid claims about the period of time that should pass for a transaction to become final. In Tenderbake, a block becomes final when there are two blocks on top of it — this is the only condition. So, if the system produces one block per 15 seconds, a transaction will become final in about 30 seconds. This is the kind of performance that users expect from a successful blockchain solution.
Historically, Tenderbake started as a variant of Tendermint and was subsequently adapted to fit into the existing Tezos system. However, as time passed, the description from the paper and its nascent implementation in the Tezos codebase started to differ. Thus, for a person starting working on Tenderbake, reading the paper is not enough. It is also hard to study the algorithm by reading the Tezos code, because the consensus algorithm is not implemented in isolation from other components of the system. This is where the Tenderbake simulator project comes into play.
Nomadic Labs asked Tweag to develop a framework that would be general enough to model any consensus algorithm (be it BTF-style, Nakamoto-style or in styles yet to be invented) in a clear way to facilitate onboarding of newcomers and for exploration of consensus algorithms in the future. The results of our work can be found in this repository. The language of choice at Nomadic Labs is OCaml and we saw no reason not to use it for this project. Similar to the Tezos code itself, the simulator is distributed under the MIT license.
We provide here a simplified description that nevertheless should give the reader an idea of what the framework can do. The project comes with a guide that explains in detail how the simulation framework works and how to implement a consensus algorithm in it.
The simulator allows us to observe the evolution of a system that comprises a collection of nodes — independent processes that do not share memory but can exchange messages by means of asynchronous broadcast that can be fine-tuned by the user if desired.
Consensus algorithms in the framework are implemented as event handlers — functions that are called when an event occurs at a particular node. A call of an event handler is called an iteration. Event handlers have the following type signature:
type event_handler = Algorithm.params -> Time.t -> Event.t -> Signature.private_key -> Algorithm.node_state -> Effect.t list * Algorithm.node_state
Let’s go over the arguments of the function:
Algorithm.paramsis the parameters of the algorithm such as e.g. round duration in seconds.
Time.tis the current time.
Event.tis the event the node needs to react to. Currently there are two kinds of events: reception of a message and a “wake up” call that the node can schedule for itself. Message types are defined per consensus algorithm.
Signature.private_keyis the private key that every node magically knows. It is used for signing of messages. This is important because the framework allows us to program and use Byzantine versions of nodes, too.
Algorithm.node_stateis the node state. The type is defined per consensus algorithm.
The return type of an
event_handler is an effect list and an updated node
state. An effect in the list can be one of the following:
- Broadcast a message to all nodes.
- Schedule a wake up call.
- Shut down.
We have written two kinds of tests: stress tests and scenario tests.
Stress tests are about letting an algorithm run for a number of iterations with a large enough network and realistic propagation of messages, including messages getting lost and messages arriving out of order. The framework allows us to specify a set of predicates that must hold at each iteration in such a test. This way we can determine if an algorithm satisfies liveness and safety properties. According to the tests, all models that we have written satisfy both properties.
Scenario tests are about adjusting propagation of messages and/or lifetime of nodes in order to model a situation of interest. We can then inspect execution logs and check whether the nodes behaved in the expected way. It is easy to do because simulations return typed logs that we can pattern match on.
We have implemented four consensus algorithms (listed below roughly in order of increasing complexity), and applied stress tests and scenario tests as discussed above:
- Leader election, see
- Ouroboros (the simple BFT version), see
- Emmy+, see
src/emmy_plus; this is the current consensus algorithm used by Tezos.
- Tenderbake, see
src/tenderbake; this is the algorithm Nomadic Labs is planning to propose as a future amendment to the Tezos blockchain.
Every algorithm is explained in its
README.md. Our focus was not only
explaining how a particular algorithm works in principle, but also how it
translates to code that the simulator framework can run.
The simulator has already proven useful. People who have tried it out report that it has helped them understand the algorithm of Tenderbake and experiment with it. In the future, we can expect new consensus algorithms to be implemented and explored using this framework. Please, feel free to give it a try and contribute!