Broken round synchronous.
rantan opened this issue · 4 comments
rantan commented
I found a scenario which breaks round synchronous.
Scenario
Assumption
- There are three signers A, B, and C.
- B and C run with sync and A is stopped.
- A gets back to run.
- The order of round-robin is A→B→C.
Initial status
- B and C are working in a round that the master of the round is A.
- A is working in a round that the master of the round is C.
Execution result
- B, C's round is timeout. Then, they start a new round that master is B.
- A's round is timeout. Then, A starts a new round that master is A.
- B progress the round to the end with C. B broadcast completedblock message.
- B, C stats a new round that the master is C.
- A is sleeping. (In the current implementation, when the round that the node is master starts, the node is going to get to sleep for the round interval.)
- A awakes up. A progress the round to the end with B. A broadcast completedblock message.
After this, the block proof generation countiue with the broken round synchronous.
How to fix
- Make tapyrus-siner stop getting into sleep when the round the node is master gets started.
- Change the round master assumption when the node gets received key messages which relates round start and end like
candidateblock,completedblock.
Yamaguchi commented
I have some questions.
- What behavior is expected?
- What does 'the broken round synchronous' mean?
- What is 'the round master assumption'?
rantan commented
I have created a visualization for the scenario.
rantan commented
- What behavior is expected?
I have 2 points.
- Blocks should be going to be created with correct intervals.
- All living nodes should be able to join each round. In the diagram, only two nodes can join each round because of sleeping one node.
- What does 'the broken round synchronous' mean?
It is the status that is shown in the scenario visualization.
- What is 'the round master assumption'?
I meant that it is master_index in NodeState.