Prove the Raft demo's safety claims under seeded chaos - #18
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The evidence half of the Raft demo. Four safety invariants (election safety, log matching, leader completeness, state-machine safety) are checked after every simulated run against an ordered record of elections and applies; a seeded chaos campaign — randomized partitions, process restarts, message loss and duplication — holds them across 50,000 seeds in about 15 minutes on a laptop, and a smaller sweep now runs nightly next to the jobqueue's.
The ablation table is the point: switch off any one safeguard (vote ledger, log-freshness check, persistence before reply, stale-term rejection, the own-term commit gate) and the explorer finds a violating schedule within a few seeds, replayable from its seed and shrinkable to the scheduling decisions that mattered — one out of 3,514 in the sharpest case. Two safeguards get standalone safe-proofs so the defense-in-depth claim is earned in both directions.
Writing the replay-stability check also caught a real determinism bug in the demo (connection handlers cancelled in set-iteration order); fixed here with a regression guard, and every number in the README was re-measured on the fixed code.