fix(leaderboard): match period rank tiebreakers across profile and embed surfaces

[junhoyeo]

What

Three surfaces show a user's rank for a finite period — the leaderboard's week/month/custom tab, the public profile's 7d/30d view, and the embed card's week/month view — and until now they did not agree on what a tie means.

getLeaderboard.ts builds its rank expression from a sequentialRanks flag. The period path passes true and emits ROW_NUMBER() OVER (ORDER BY total_tokens DESC, total_cost DESC, LOWER(username) ASC, user_id ASC) (a cost sort swaps the two metrics), which gives every user a distinct position and gives pagination a stable order. The all-time path passes false and emits RANK() OVER (ORDER BY total_tokens DESC), so tied users share a position.

publicProfileData.ts and getUserEmbedStats.ts each rolled their own period rank with a bare RANK() OVER (ORDER BY total_tokens DESC). So two users on the same period total read #1 and #2 on the leaderboard's week tab, but #1 and #1 — followed by a gap — on their own profile pages and on their embed cards, from the same data over the same window at the same moment.

Change

Both finite-period rank queries now use the leaderboard's period expression and its tiebreak columns, matched to each surface's own ranking metric (the embed's cost sort orders total_cost DESC, total_tokens DESC, the tokens sort the other way around, and the profile ranks by tokens only).

The tiebreak needs username and user_id in scope, so each period CTE now also selects u.username (and, on the profile side, SUM(CAST(d.cost AS DECIMAL(18,4))) AS total_cost for the secondary term) and groups by s.user_id, u.username. users.id is the primary key the join runs through, so the added grouping column produces the same groups and the embed's SELECT COUNT(*)::int FROM rankable denominator is unchanged. The u.leaderboard_hidden = false filter and the date-range bounds are untouched — the same rows are ranked, and only the order within a tie moves. The Number(...) coercion of the returned bigint, the unstable_cache keys and tags, and the profile's empty-window skip all stay as they were.

The all-time paths deliberately stay on shared RANK(). They already agree with the leaderboard's all-time tab, and moving them to ROW_NUMBER would introduce exactly the same divergence in the other direction. Both files now carry a comment saying so, and the tests below pin it.

Tests

__tests__/api/usersProfile.test.ts and __tests__/lib/getUserEmbedStats.test.ts mock next/cache and drizzle-orm's sql tag, so they can assert on the SQL each path emits. Added there:

• ranks the profile period window the way the leaderboard's period tab does — asserts the period rank emits the full ROW_NUMBER() OVER (ORDER BY total_tokens DESC, total_cost DESC, LOWER(username) ASC, user_id ASC) clause, that no RANK() OVER is emitted on that path, that the hidden-user filter and both date bounds are unchanged, and that the cache key still carries the anchored window.
• ranks the finite embed window the way the leaderboard's period tab ranks $sortBy — the same check for the embed card, once per sort, pinning the metric order for each and the of N denominator.
• keeps the lifetime profile rank on shared RANK and the renamed keeps the lifetime embed rank on shared RANK with no tie-breakers — assert the all-time paths still emit RANK() and emit no ROW_NUMBER at all, so the mirror-image bug fails the suite too.

Four existing assertions that matched the old period SQL shape were updated from RANK() OVER to ROW_NUMBER() OVER. The embed test's SQL serializer now recurses into nested sql fragments — a conditional ORDER BY is interpolated as a value, and the old serializer flattened it to [object Object], which hid the clause these tests are about.

Verified by falsification in both directions: reverting the source change fails 7 tests (the 3 new period tests plus the 4 updated ones) and leaves the two all-time tests passing; flipping the all-time queries to ROW_NUMBER instead fails exactly those two and nothing else.

Verification

From packages/frontend, on the branch head:

• npx tsc --noEmit — exit 0
• bun run test — exit 0, 925 passed | 6 skipped (931) across 87 passing / 1 skipped test files
• bun run lint — exit 0, 17 pre-existing warnings, none in the changed files

No database integration test: the repo has no harness for one on this path, so these tests pin the emitted SQL rather than Postgres' ordering of a real tie.

---

Summary by cubic

Aligns finite-period ranks on profile and embed with the leaderboard’s period tab. Previously, profile/embed used shared RANK on period totals (ties showed the same position); now they use sequential ROW_NUMBER with the same tie-breakers as the leaderboard, so tied users read distinct positions consistently. All-time ranks remain on shared RANK.

• Period queries now rank by ROW_NUMBER over tokens/cost plus tie-breakers (tokens sort: total_tokens DESC, total_cost DESC, LOWER(username) ASC, user_id ASC; cost sort swaps the first two). CTEs include u.username (and profile includes summed total_cost) and group by s.user_id, u.username. Filters, date bounds, cache keys/tags, and the embed “of N” denominator are unchanged; only tie ordering moves. Tests assert emitted SQL for each path and update the SQL serializer to recurse nested fragments.

^{Written for commit ddf66e6. Summary will update on new commits.}

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
tokscale	Ready	Preview	Aug 24, 2026 10:48pm