This also contains a few fixes for the `common_subset` module:
* Rename `common_subset::Output` to `Message` to avoid confusing it
with the value that the algorithm outputs as a result.
* Implement dispatch of messages to the right instance within
`CommonSubset`, in a way that is transparent to the user.
This fixes several issues with the broadcast algorithm and enables the
previously ignored tests that now pass:
* Don't decide on a root hash based on anyone's `Echo` message.
* Work around the `merkle` crate's inability to produce the proof of the
`i`-th leaf for a given index `i`.
* Ignore messages from unknown nodes.
* Avoid decoding multiple times.
* Document the full algorithm.
* Don't count multiple `Echo` or `Ready` messages coming from the same
node.
* Fix index computation for a given proof.
* Move `BroadcastMessage` into `broadcast` to make the module more
self-contained.
The example now only executes a single broadcast instance, expecting the
first node (the one with the lexicographically lowest address) to
propose a value. A shell script is added that runs for example nodes.
Use env_logger instead of simple_logger, so the log level can be controlled
with an environment variable. You can e.g. log all output from the broadcast
test and the crate itself in debug mode with:
RUST_LOG=hbbft=debug,broadcast=debug
Some debugging messages are more concise now and use hexadecimal
notation instead of printing arrays of decimal values.
An indentation error in the Travis script is also fixed.
I reduced the socket IO tasks to mere message forwarders. The algorithm
complexity lies in stage modules. Example: broadcast/mod.rs. Communication is
set up and modules are run from the node module.
There is a problem with this commit. std:🧵:spawn imposes a static
lifetime guarantee on type T in Message<T>.
Peter van Nostrand
Andreas Fackler
Vladimir Komendantskiy