gal-tab/sir_albert/data-review

Boot from os/PREAMBLE.md.

Data Review

A mandatory pre-ship gate. Every dashboard or analysis that leaves your hands gets this review first. Takes ~10 minutes; prevents days of embarrassment.

Announce at start: "Running data-review pre-ship check."

Checklist

1. Schema + Null-Rate Sanity

• [ ] Key columns exist and have the expected types (no silent type coercions).
• [ ] Null rate on critical fields is within acceptable bounds — flag anything >5% unless the spec says otherwise.
• [ ] Row count is in the expected order of magnitude (no silent filter killing most rows).
• [ ] Date range covers the intended window; no off-by-one at period boundaries.

2. Metric-Definition Consistency

• [ ] Every metric name matches the attested spec (to-prd or the team-repo definition).
• [ ] Numerator and denominator are explicitly verified (not inferred from column names).
• [ ] Deduplication logic is present where required (e.g. one row per deal, per user, per event).
• [ ] Conversion rate denominators are non-zero; division-by-zero is handled.

3. Dashboard-Link Freshness

• [ ] The underlying data source last-refreshed timestamp is visible or verifiable.
• [ ] Scheduled refresh is configured and its next run is within the expected cadence.
• [ ] Any hard-coded date filters in the SQL are intentional and documented.

4. Source-Join Correctness

• [ ] Every JOIN has a verified cardinality assumption (1:1, 1:N, N:M — document the expected one).
• [ ] Fan-out risk: if joining on a non-unique key, aggregation is happening after the join, not before.
• [ ] Left-join gaps are accounted for: missing rows are expected / filtered / labelled, not silently zero.
• [ ] Cross-source joins use a consistent grain (e.g. both sides are at deal level, not mixed deal/opportunity).

Boil-the-Ocean Default

Default to the complete cut. Before sharing, add the breakdown that makes the number meaningful:

• by channel / source
• by segment (SMB / MM / ENT)
• by motion (self-serve / sales-assisted / partner)
• trend vs. plan (actual vs. target, WoW / MoM / QoQ)

The complete cut is usually one more GROUP BY clause or one more panel in the dashboard. It almost never materially changes the query cost. Do it. If the audience doesn't need a breakdown, they'll ignore it — but if they do need it and it's missing, you'll be asked for it in the meeting.

Output

After running the checklist, produce a short summary:

DATA REVIEW — <dashboard / analysis name>
Date: <today>
 
PASS / FAIL / WARN items:
- [PASS] Schema: all key columns present, types correct
- [WARN] Null rate on `segment` is 12% — verify if expected
- [PASS] Metric definitions match spec
- [FAIL] LEFT JOIN on `deal_id` has unmapped rows — investigate before sharing
 
Complete-cut status:
- by channel: YES / NO
- by segment: YES / NO
- by motion: YES / NO
- trend vs. plan: YES / NO
 
Recommendation: READY TO SHIP / HOLD — fix [issue] first

Only ship on READY TO SHIP. A HOLD item is not optional.