fix(admin): sum global stats costs in numeric

[steebchen]

Problem

On the admin Global Stats page the cost card does not add up: the headline total and the per-mode / per-kind slices printed underneath it are the same rows grouped two ways, yet they disagree by double-digit dollars at all-time scale. Narrow the view to a small enough slice and it reconciles.

Request counts reconcile everywhere; only money drifts, and only at scale. That is the tell.

Every cost column on the stats tables is real (float4), and Postgres accumulates SUM(real) in float4 too — sum(real) → real, about 7 significant digits. Once a running sum is large enough its ulp exceeds a cent, so the result depends on the order and grouping of the rows. The headline (one SUM over everything) and the composition slices (SUM per mode / per kind) therefore diverge. Small slices reconcile because they stay well inside float4's range.

Reproduced standalone:

with r as (select (i % 4) as bucket, 1234.5678::real as c from generate_series(1,800) i)
select (select sum(c)::text from r),                                       -- 987650.7
       (select sum(s)::text from (select sum(c) s from r group by bucket) g), -- 987652.7
       (select sum(c::float8::numeric)::text from r);                      -- 987654.19921875

Three different answers for the same 800 rows.

Fix

Read path (GET /admin/global-stats) — accumulate cost sums in NUMERIC, which is exact, so any grouping of a row set yields a bit-identical total.

The cast chain matters and is not decorative: real::numeric routes through float4's 6-digit display form (1234.5678::real::numeric → 1234.57), losing more than the float4 sum did. real::double precision is the exact stored value, so the cast is real → float8 → numeric, then back to float8 for JSON.

Also stopped narrowing the all-time request/error/cache counts back to int4 — this endpoint is cross-tenant and all-time, so the count will outgrow 2^31 eventually and fail hard.

Write path (apps/worker hourly aggregation) — same bug a layer down: an hour is tens of thousands of log rows, summed in float4 before being stored, so the buckets were already wrong before anything read them. Now cast to double precision first; the target column is still float4, so only the final value is rounded rather than every partial sum along the way. This covers global_*_stats, project_hourly_*, api_key_hourly_* and provider_key_hourly_stats.

While there: data_storage_cost was being summed as cast(... as real) even though log.data_storage_cost is numeric — it was downcast for no reason.

Not in this PR

• No column-type migration. The stats tables keep real columns. Per-row rounding is ~5e-8 of the row's own value; the error that was visible came overwhelmingly from the accumulation, which is what this fixes.
• The other SUM(<real column>) sites across analytics.ts, the model/mapping history rollups and stats-calculator.ts have the same latent bug, but they are per-project or per-model and orders of magnitude smaller, so the drift is currently invisible. Left for a follow-up rather than bundled into a bug fix.
• The "Unattributed" bucket is a separate defect — org_kind was never backfilled when the column was added, and used_mode only partially. That is addressed in feat(scripts): backfill global stats attribution #3549.

Verification

New spec large all-time sums stay exact and grouping-independent builds a synthetic production-magnitude fixture (700 days × 4 mode/kind buckets) and asserts the headline matches the mode composition, the kind composition, and the timeseries.

Against the unfixed code it reproduces the reported symptom at scale:

AssertionError: expected 12098983.5 to be close to 12099147.748046875,
received difference is 164.248046875, but expected 0.005

• apps/api — 930 tests pass (70 files)
• apps/worker — 114 tests pass (9 files)
• pnpm format, pnpm build clean

No screenshots: this is a numeric correctness fix behind an existing card, nothing about the layout changes.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Improved accuracy of cost and usage totals in global statistics.
  • Prevented rounding drift when aggregating large volumes of decimal-based costs.
  • Preserved fractional request, error, and cache counts instead of narrowing them to whole numbers.
  • Updated cost breakdowns across modes, providers, discounts, and storage metrics for more reliable totals.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 11a8da6f-e1dc-4a9b-ab11-09640e761958

📥 Commits

Reviewing files that changed from the base of the PR and between 0f2c22c and 29a6343.

📒 Files selected for processing (1)

• apps/api/src/routes/admin.ts

🚧 Files skipped from review as they are similar to previous changes (1)

• apps/api/src/routes/admin.ts

---

Walkthrough

The change updates admin and worker cost aggregation to use double-precision intermediates. Count aggregates now use float8. New tests validate 700-day global statistics totals against float4-rounded stored costs.

Changes

Cost aggregation precision

Layer / File(s)	Summary
Admin statistics aggregation apps/api/src/routes/admin.ts	Adds sumMoney for precise cost totals and changes request, error, and cache counts to float8.
Worker statistics aggregation apps/worker/src/services/project-stats-aggregator.ts	Uses double-precision accumulation for project, modality, discount, storage, credit, API-key, and provider-key costs.
Large-scale aggregation validation apps/api/src/routes/admin-global-stats.spec.ts	Adds bulk fixtures, cleanup, baseline accounting, and a 700-day aggregation test.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

• theopenco/llmgateway#3247: Modifies shared cost aggregation in apps/api/src/routes/admin.ts for storage-cost metrics and billing.
• theopenco/llmgateway#3453: Modifies global and project statistics aggregation areas that this change updates for greater numeric precision.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly identifies the admin Global Stats cost-summation fix, which matches the primary change.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/global-stats-float4-sums

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}