Part of #33. The highest-value Sei-specific signal — deferred because it's blocked on a node-side dependency, not on sei-load work.
Problem
Block-stm aborts/re-executions under hot-key contention are invisible client-side: a re-executed tx still returns a success receipt. The signal must come from node-side sei_occ_*/block-stm metrics (re-executions, conflict aborts, effective parallelism) on the validator /metrics.
Blocker / upstream
Confirm the exact metric names are exposed by seid — file/track a sei-chain issue. Un-defer the moment names are confirmed.
Do NOT
Do not label client-side metrics/txs as "expected to conflict." Conflict is emergent (same-block co-execution + scheduler), not deterministic per-tx. Record contention pressure (zipf_theta, recordcount) as the independent variable; correlate measured aborts as the outcome — never pre-label.
P1 proxy (until unblocked)
Latency-under-contention + %-committed/goodput as θ rises — observe the cost of contention without attributing it to aborts.
Design: https://github.com/sei-protocol/platform/blob/main/docs/designs/sei-load-workload-modeler.md
Part of #33. The highest-value Sei-specific signal — deferred because it's blocked on a node-side dependency, not on sei-load work.
Problem
Block-stm aborts/re-executions under hot-key contention are invisible client-side: a re-executed tx still returns a success receipt. The signal must come from node-side
sei_occ_*/block-stm metrics (re-executions, conflict aborts, effective parallelism) on the validator/metrics.Blocker / upstream
Confirm the exact metric names are exposed by
seid— file/track a sei-chain issue. Un-defer the moment names are confirmed.Do NOT
Do not label client-side metrics/txs as "expected to conflict." Conflict is emergent (same-block co-execution + scheduler), not deterministic per-tx. Record contention pressure (
zipf_theta,recordcount) as the independent variable; correlate measured aborts as the outcome — never pre-label.P1 proxy (until unblocked)
Latency-under-contention + %-committed/goodput as θ rises — observe the cost of contention without attributing it to aborts.
Design: https://github.com/sei-protocol/platform/blob/main/docs/designs/sei-load-workload-modeler.md