iotaledger / iota / 30596652036

70%

Build Type

push

github

Committed by web-flow

Commit Message

feat(consensus): leader schedule redesign (#12474)

# Description of change

Redesign of the Starfish leader schedule, aimed at removing crashed or
degraded validators from the leader rotation faster. Two independently
flag-gated changes, both off by default:

- **Sliding-window reputation scoring**
(`consensus_enable_sliding_window_leader_schedule`): replaces V2's batch
scoring (accumulate votes for a full schedule period, then rebuild from
scratch) with a running window aggregate updated incrementally on every
commit. The schedule can rotate much more frequently (e.g. every 60
commits over a 300-commit window) with no per-rotation rebuild cost; on
restart the window is rebuilt by replaying committed subdags. Under
`consensus_starfish_speed`, a strong vote (attesting the voter holds the
leader's transaction data, not just its header) counts double in the
sliding-window scorer, so validators that vote on time but consistently
lag on payloads score lower.
- **Absolute-score bad-node selection**
(`consensus_enable_absolute_score_leader_schedule`): replaces the fixed
bottom-stake-rank cut with thresholds on scores normalized to the
window's top performer, so a failing validator is excluded on its own
score instead of having to out-decay the chronically lowest-ranked ones.
Applies to both scoring paths.

Landed parts:

- #11875 sliding-window leader scorer
- #11975 wiring behind the protocol flag
- #12076 metric rename
- #12123 skip V2 vote scoring on the sliding-window path
- #12264 skip sliding-window recovery during fast sync
- #12295 leader-schedule window invariant check
- #12350 certifier scan refactor
- #12408 absolute-score bad-node selection
- #12304 rebuild the sliding window from a recovered scoring backlog
- #12470 double strong votes in sliding-window reputation scoring

Measured on a 20-validator network with crash injection: post-crash
leader timeouts roughly halve versus the current V2 + stake-rank
configuration (~2×, idle and under 100 TP... (continued)

Coverage Stats

1108 of 1123 new or added lines in 10 files covered. (98.66%)

2713 existing lines in 72 files now uncovered.

165405 of 239390 relevant lines covered (69.09%)

464949.73 hits per line