Add / Change a Stats Dimension

The stats pipeline has one write implementation but several coupled registries. Missing any step below produces silent data loss or drift — none of it fails loudly. Work through the checklist in order.

1. Schema (single source of truth)

• Add the CREATE TABLE / CREATE INDEX statements to apps/collector/src/database/schema.ts (SCHEMA + INDEXES). Nothing else creates tables.
• Existing databases only get IF NOT EXISTS DDL from schema.ts. Column additions and index drops need an explicit migration in apps/collector/src/modules/db/db.ts init() (see the ALTER TABLE ... ADD COLUMN and DROP INDEX IF EXISTS precedents there).
• Do NOT add expression indexes on (total_download + total_upload) — hot queries filter by backend_id first, so the planner never uses them (verified with EXPLAIN QUERY PLAN); they only amplify writes.

2. Write path (exactly one)

• All traffic writes go through batchUpdateTrafficStats in apps/collector/src/database/repositories/traffic-writer.repository.ts. updateTrafficStats is a thin wrapper over it — never add a second write implementation.
• Add your aggregation map in the batch loop and the upsert in the appropriate transaction group. The whole flush is one outer transaction; keep it that way (a partial commit + buffer retry double-counts).
• CSV-style association columns (ips, rules, chains, domains) must dedupe with the delimiter-aware pattern — bare INSTR(list, value) treats a.com as present when aa.com is:

INSTR(',' || col || ',', ',' || @value || ',') > 0

• Respect reduceWrites semantics: detail tables are skipped when ClickHouse is the read source (shouldReduceSqliteWrites in modules/stats/stats-write-mode.ts). Decide whether your table is "core" (always written) or "detail" (skippable) and place it in the matching branch.

3. Rule naming contract (if the dimension involves rules)

• Rule stats key on the top-level policy group (last chain hop) for multi-hop chains — buildRuleName in src/shared/utils/rule-name.ts is the only place this logic lives. The frontend matches these names against gateway rule targets (rule.proxy), so changing the key breaks the flow graph and rule matching (this regressed once in v1.3.9; see docs/dev/deep-review-2026-07.md).

4. Retention

• Register cleanup in apps/collector/src/modules/cleanup/cleanup.service.ts + config.repository.ts. Minute-level tables follow connectionLogsDays (default 7d); hourly tables follow hourlyStatsDays (default 30d). hourly_dim_stats backs all >6h range queries — never tie it to the minute cutoff.
• Cumulative *_stats tables currently have no retention; if your table is a cross-product (rule×ip style), consider a last_seen-based cap from day one.

5. Backend deletion

• Add the table to the explicit delete list in deleteBackendData (apps/collector/src/database/repositories/backend.repository.ts). PRAGMA foreign_keys is OFF, so the declared ON DELETE CASCADE never fires — forgetting this leaves orphan rows forever.

6. ClickHouse parity (if the data is queryable over ranges)

• Writer: apps/collector/src/modules/clickhouse/clickhouse.writer.ts
• Reader: apps/collector/src/modules/clickhouse/clickhouse.reader.ts — the SQLite and CH implementations of the same query must return the same shape and semantics (time format, windows). Routing lives in modules/stats/stats.service.ts (*WithRouting methods).

7. Types and API

• Shared response types go in packages/shared (both apps consume them; never fork the type locally).
• Time keys are UTC ISO slices: minute YYYY-MM-DDTHH:MM:00, hour YYYY-MM-DDTHH:00:00 (toMinuteKey/toHourKey in base.repository.ts). Don't introduce local-time keys.

8. Tests

• Add a Vitest case in apps/collector exercising the batch write + read-back (see traffic-writer.test.ts for patterns, src/__tests__/helpers.ts for DB setup), then run the verify-changes skill.