Investigate missed inference on some nodes (root causes + mitigation) #820
Discussed in https://github.com/gonka-ai/gonka/discussions/817
💬 Comments (3)
Measured data from live network may help narrow root causes here.
Epochs 161–191, 2,503,595 inferences: Miss rate: 3.25% (81,360 misses) Completion rate: mean 90.4%, σ=7.4%, range 72–99%
The σ=7.4% variance is the signal — not the mean. Some nodes miss 28% of assigned inferences while others miss 1%. GetRandomExecutor routes to both equally regardless.
Phase 4 of GiP #860 proposes GetQualityWeightedExecutor — routes traffic proportional to L9 completion rate. Projection: σ ↓40% as high-miss nodes receive less traffic and face economic incentive to improve.
Design + data: docs/specs/inference-quality-protocol.md (PR #859 branch) Discussion: #860
Is this issue still relevant after PR #943 and PR #859 were merged, or are there remaining edge cases to investigate?
🔄 Auto-synced from Issue #820 every hour.
PR: https://github.com/gonka-ai/gonka/pull/843