Epic: workload modeler (DB-style benchmarking of Sei EVM)

[bdchatham]

Tracking epic for evolving sei-load from a throughput generator into a DB-style workload modeler for benchmarking Sei's EVM under controlled key-contention, tx-size, and load distributions.

Thesis: Does a change move Sei's EVM throughput/latency under controlled key-contention and tx-size — and can I trust the comparison? Primary use case: A/B benchmarking (PR-vs-main) on harbor.

Design (signed off, merged): https://github.com/sei-protocol/platform/blob/main/docs/designs/sei-load-workload-modeler.md

Approach borrows the established canon — YCSB distribution taxonomy (hot keys / size buckets / op mix), wrk2/Vegeta open-loop + HdrHistogram (coordinated-omission-correct latency), Diablo/Caliper blockchain metrics.

Phase 1 (MVP — tunable contention + size + trustworthy A/B numbers)

• Phase 1: open-loop measurement core (coordinated-omission fix) #34 — Open-loop measurement core (coordinated-omission fix) — foundation, non-negotiable
• Phase 1: Distribution primitive (uniform/zipfian, seeded) #35 — Distribution primitive (uniform/zipfian, seeded)
• Phase 1: StorageRW contention contract + key distribution #36 — StorageRW contention contract + key distribution
• Phase 1: tx-size distribution (size buckets) #37 — Tx-size distribution (size buckets)
• Phase 1: reproducible run-summary contract #38 — Reproducible run-summary contract

Blocked / upstream

• Parallel-exec abort-rate surfacing (blocked on node-side metrics) #39 — Parallel-exec abort-rate surfacing (needs node-side sei_occ_*/block-stm metrics)

Deferred (trigger-gated — see design "Deferred" section, not yet active)

Multi-contract/hot-contract sets · --workload presets · seiload_run_info{} gauge · wallet-skew distribution · full YCSB A–F matrix + latency-over-time · Poisson/bursty arrival.

Backwards compatibility (contract — applies to every Phase 1 PR)

Additive: optional fields default to today's behavior; deprecate-don't-delete existing scenarios (StorageRW added alongside ERC20Conflict); fence the intentional latency-semantics change behind methodology_version.

Suggested order

#34 (measurement foundation) → #35 (distribution) → #36 + #37 (contention + size, parallelizable) → #38 (run-summary ties it together). #39 unblocks once the sei-chain metric names land.