523d2ac911
* Rewrite tendermint's message handling loop to much more clearly match the paper No longer checks relevant branches upon messages, yet all branches upon any state change. This is slower, yet easier to review and likely without one or two rare edge cases. When reviewing, please see page 5 of https://arxiv.org/pdf/1807.04938.pdf. Lines from the specified algorithm can be found in the code by searching for "// L". * Sane rebroadcasting of consensus messages Instead of broadcasting the last n messages on the Tributary side of things, we now have the machine rebroadcast the message tape for the current block. * Only rebroadcast messages which didn't error in some way * Only rebroadcast our own messages for tendermint |
||
|---|---|---|
| .. | ||
| src | ||
| tests | ||
| Cargo.toml | ||
| LICENSE | ||
| README.md | ||
Tendermint
An implementation of the Tendermint state machine in Rust.
This is solely the state machine, intended to be mapped to any arbitrary system. It supports an arbitrary signature scheme, weighting, and block definition accordingly. It is not intended to work with the Cosmos SDK, solely to be an implementation of the academic protocol.
Caveats
-
Only SCALE serialization is supported currently. Ideally, everything from SCALE to borsh to bincode would be supported. SCALE was chosen due to this being under Serai, which uses Substrate, which uses SCALE. Accordingly, when deciding which of the three (mutually incompatible) options to support...
-
The only supported runtime is tokio due to requiring a
sleepimplementation. Ideally, the runtime choice will be moved to a feature in the future. -
It is possible for
add_blockto be called on a block which failed (or never went through in the first place) validation. This is a break from the paper which is accepted here. This is for two reasons.- Serai needing this functionality.
- If a block is committed which is invalid, either there's a malicious majority now defining consensus OR the local node is malicious by virtue of being faulty. Considering how either represents a fatal circumstance, except with regards to system like Serai which have their own logic for pseudo-valid blocks, it is accepted as a possible behavior with the caveat any consumers must be aware of it. No machine will vote nor precommit to a block it considers invalid, so for a network with an honest majority, this is a non-issue.
Paper
The paper describes the algorithm with pseudocode on page 6. This pseudocode isn't directly implementable, nor does it specify faulty behavior. Instead, it's solely a series of conditions which trigger events in order to successfully achieve consensus.
The included pseudocode segments can be minimally described as follows:
01-09 Init
10-10 StartRound(0)
11-21 StartRound
22-27 Fresh proposal
28-33 Proposal building off a valid round with prevotes
34-35 2f+1 prevote -> schedule timeout prevote
36-43 First proposal with prevotes -> precommit Some
44-46 2f+1 nil prevote -> precommit nil
47-48 2f+1 precommit -> schedule timeout precommit
49-54 First proposal with precommits -> finalize
55-56 f+1 round > local round, jump
57-60 on timeout propose
61-64 on timeout prevote
65-67 on timeout precommit
The corresponding Rust code implementing these tasks are marked with their related line numbers.