feat(power): measured-power multinode support (workers[] + per-stage joules)

[arygupt]

What

App-side support for the InferenceX runner's multinode measured-power telemetry — the dashboard counterpart to SemiAnalysisAI/InferenceX#1574.

Surfaces, end-to-end (ingest → DB → transform → types):

• Per-worker power — workers[], each labeled prefill / decode with num_gpus, hosts[], and per-worker avg_power_w (+ optional temp/util/mem).
• Per-stage joules (disagg) — joules_per_input_token = prefill energy ÷ input; joules_per_output_token = decode energy ÷ output (per-stage, decode GPUs only). joules_per_total_token stays cluster-wide.
• Cluster-wide telemetry — avg_temp_c, peak_temp_c, avg_util_pct, avg_mem_used_mb.

Changes

• db: migration 006_benchmark_results_workers.sql; ETL mapper + ingest + queries + json-provider carry workers[] and the new metric scalars.
• constants: metric-keys lists the measured-power fields.
• app: benchmark-transform + inference types surface the fields; unofficial-run route passes them through; cypress measured-power-overlay e2e.

Schema note (pairs with the runner change)

joules_per_output_token is now the per-stage decode value on disagg (symmetric with joules_per_input_token), matching the reviewer's spec on #1574. The old joules_per_output_token_decode key is removed (folded into joules_per_output_token) — it was only plumbed through, never rendered; the J/output chart already reads joules_per_output_token.

Validation

• app 60 + db 223 unit tests pass
• tsc --noEmit clean; lint (oxlint) 0 warnings/errors

Depends on

Runner PR SemiAnalysisAI/InferenceX#1574 (emits these agg-JSON fields). Land/coordinate together.

🤖 Generated with Claude Code

---

Note

Medium Risk
Adds a schema migration and materialized-view rebuild plus broad ETL/API contract changes; behavior is well-tested but must stay aligned with the companion runner PR and production migration order.

Overview
Adds end-to-end support for multinode / disaggregated measured-power telemetry from the runner: per-worker breakdowns, disagg role-split scalars, and cluster-wide temp/util/memory metrics.

Database & ingest: Migration 006 adds a dedicated benchmark_results.workers JSONB column (kept out of flat metrics), recreates latest_benchmarks, and extends bulk ingest upserts. ETL gains extractWorkers() with validation, excludes workers from numeric metrics, and registers new scalar keys in METRIC_KEYS.

Read path: Benchmark queries, json-provider dumps, API BenchmarkRow, unofficial-run normalization, and benchmark-transform all pass through workers[] plus scalars like prefill_avg_power_w, decode_avg_power_w, joules_per_input_token, and avg_temp_c—preserving undefined vs zero for legacy rows.

App types & tests: WorkerPower / extended AggDataEntry types document disagg semantics (joules_per_output_token as per-stage decode on disagg). Unit tests cover mapper, transform, and unofficial-run routes; Cypress verifies measured-energy Y-axis options on an unofficial-run overlay.

^{Reviewed by Cursor Bugbot for commit db9b15d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview, Comment	May 28, 2026 11:27pm

[claude]

Claude finished @arygupt's task in 2m 58s —— View job

---

Claude Code Review

• Read changed files
• Analyze multinode transform / type changes
• Check tests cover the new transform logic
• Post review feedback

🟢 LGTM — no blocking issues found.

Reviewed the multinode measured-power plumbing end-to-end. The diff is clean and defensive at every JSON boundary:

• extractWorkers (packages/db/src/etl/benchmark-mapper.ts:263) correctly distinguishes "missing field" from "field == 0" via parseNum/parseInt2 returning undefined only for unparseable values — the frontend worker with num_gpus: 0, avg_power_w: 0 case is exercised by the tests.
• rowToAggDataEntry (packages/app/src/lib/benchmark-transform.ts:74) narrows row.workers with Array.isArray so a null JSONB column or a malformed payload survives without poisoning downstream consumers (verified by the 'oops' as any test).
• bulkIngestBenchmarkRows encodes missing workers as JSON null so the heterogeneous jsonb[] unnest stores SQL NULL — correct for the new nullable column.
• Migration 006_benchmark_results_workers.sql correctly drops + recreates the latest_benchmarks materialized view so br.* picks up the new column, and re-creates both indexes (incl. the (config_id, conc, isl, osl) unique index needed for REFRESH … CONCURRENTLY).
• The unofficial-run path (route.ts:60) passes workers through symmetrically, and cypress/e2e/measured-power-overlay.cy.ts exercises the overlay code path against a real GHA run — satisfying CLAUDE.md's overlay-support requirement.

Test coverage is solid: dedicated unit tests for extractWorkers edge cases (non-array, empty, missing required fields, string-numeric coercion, malformed hosts), workers propagation through mapBenchmarkRow/rowToAggDataEntry/normalizeArtifactRows, plus the new role-split scalars and cluster-wide temp/util/mem fields.

Minor observation (non-blocking): the WorkerPower shape is now defined three times — packages/db/src/etl/benchmark-mapper.ts:80, packages/db/src/queries/benchmarks.ts:8 (as BenchmarkWorkerRow), and packages/app/src/components/inference/types.ts:35. They're structurally identical so TypeScript accepts cross-boundary assignment, but a future schema change will need to be made in three places. Worth consolidating into @semianalysisai/inferencex-constants at some point.
· feat/measured-power-multinode-app

[claude]

Claude finished @arygupt's task in 50s —— View job

---

Claude Code Review (Re-review)

• Identify what changed since last review
• Check the new commit
• Verify test fix correctness
• Post re-review summary

🟢 Re-review: LGTM — no blocking issues found.

Only one new commit since the previous review (9847aef): a focused Cypress fix that adds .scrollIntoView() before each should('be.visible') assertion on the gated "Measured Energy" dropdown group in packages/app/cypress/e2e/measured-power-overlay.cy.ts.

Root-cause analysis in the commit message matches the code:

• The searchable-select list is a max-h-72 overflow-y-auto scroll container, so the new "Measured Energy" group renders below the fold.
• Cypress be.visible does not auto-scroll, which is why the assertion failed deterministically on both chrome and firefox e2e shard-2 jobs reported by Blacksmith.
• The second test passed because .click() does auto-scroll its target — consistent with the diagnosis.
• The fix is the smallest possible: .scrollIntoView() chained before each .should('be.visible'). No production code touched.

Previous review feedback on the multinode telemetry plumbing (extractWorkers parsing, JSONB null handling, materialized-view recreation, overlay path symmetry) all still applies and remains addressed. The non-blocking observation about consolidating the WorkerPower shape into @semianalysisai/inferencex-constants is unchanged — still optional, not introduced by this commit.
· feat/measured-power-multinode-app

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 9847aef. Configure here.}

[claude]

Claude finished @arygupt's task in 52s —— View job

---

Claude Code Review (Re-review)

• Identify what changed since last review
• Verify refactor correctness
• Post re-review summary

🟢 Re-review: LGTM — no blocking issues found.

One new commit since the previous review (db9b15d): a focused refactor that aliases BenchmarkWorkerRow to WorkerPower in packages/db/src/queries/benchmarks.ts:11, addressing the Cursor Bugbot "triplicate WorkerPower interface" finding (and the same observation in my prior review).

Verified:

• The two in-package shapes were field-for-field identical (role, worker_idx, hosts?, num_gpus, avg_power_w, avg_temp_c?, peak_temp_c?, avg_util_pct?, avg_mem_used_mb?) — type BenchmarkWorkerRow = WorkerPower is a safe drop-in.
• The public name BenchmarkWorkerRow is preserved, so external consumers (json-provider, BenchmarkRow.workers) are unaffected.
• Scope is correctly limited to the in-db-package duplication. The third copy in packages/app/src/components/inference/types.ts stays — consolidating that would require moving the shape into @semianalysisai/inferencex-constants, which is a larger cross-package change appropriate for a follow-up.

All prior review findings on the multinode telemetry plumbing (extractWorkers parsing, JSONB null handling, materialized-view recreation, overlay path symmetry, test coverage) remain addressed and still apply.
· feat/measured-power-multinode-app