[forge][experiment] Window-only: proposer window multiplier 10× → 100×#19574
Draft
danielxiangzl wants to merge 3 commits into
Draft
[forge][experiment] Window-only: proposer window multiplier 10× → 100×#19574danielxiangzl wants to merge 3 commits into
danielxiangzl wants to merge 3 commits into
Conversation
Forge harness for measuring commit-latency baseline against leader- reputation experiments. Self-contained: - realistic_env_max_load: ConstTps 4000 with 7 validators, 0 fullnodes - 24-hour epoch_duration_secs (no epoch changes during the test) - 20-min effective duration via with_duration_override(1200) - Last validator (by ordered index) simulates a slow proposer: 1s proposal delay every 10th round, guaranteeing round timeout Adds the duration_override plumbing to ForgeConfig/runner so individual tests can override the CLI duration. No P90 latency or multi-region changes; pure baseline for A/B against latency-weighted leader-rep. Prototype/experiment code -- not for merge to main. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Isolates the contribution of just bumping the leader-reputation proposer window from 10x (default, ~70 blocks) to 100x (~700 blocks / 35s). All other classifier parameters (failed_weight=1, failure_threshold_percent=10) match the baseline. Acts as a control for #19567 (heuristic-only) and a decomposition step for #19566 (which bundles window + threshold + failed_weight changes). Hypothesis: window alone slightly regresses p90 because larger window stabilizes the failure-rate estimate near the threshold boundary (threshold = V6's true 10% rate). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This comment has been minimized.
This comment has been minimized.
Contributor
✅ Forge suite
|
danielxiangzl
changed the title
This was referenced Apr 28, 2026
Draft
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.
Summary
Window-only experiment branch. Bumps
proposer_window_num_validators_multiplierfrom the default10to100(700 blocks / ~35 sec window with 7 forge validators), keeping all other classifier parameters at baseline defaults (failed_weight=1,failure_threshold_percent=10, no latency-weighted heuristic).Why
Acts as a control PR to cleanly decompose the latency-improvement experiment ladder:
Hypothesis (CONFIRMED)
P90 should not improve because the larger window stabilizes V6's measured failure rate near the 10% threshold boundary (= V6's true failure rate), causing more consistent "active" classification rather than helpful oscillation. Confirms that the threshold/failed_weight changes (not the window bump) are the actual fix in #19566 and #19341.
Forge results (run 2026-04-28 21:24-21:40 UTC, 16 min, 4k TPS, 1 slow validator)
Commit-accepted latency vs baseline #19330:
Statistically indistinguishable from baseline. Window increase alone provides no measurable benefit when the threshold equals the true failure rate.
Conclusion
Window size is not a useful lever in isolation. Decision-relevant value of this PR: it isolates "window's solo contribution = ~0%" so the gains in #19566/#19341 can be attributed entirely to the classifier/heuristic changes, not the window bump that they all also include.
⚠ Draft / experiment — not for merge to main.
🤖 Generated with Claude Code