docs/database.md

# Database Reference

This project keeps local SQLite metadata databases for Danbooru, Gelbooru, and a
deduplicated quality union.

## Databases

| DB | Purpose |
|----|---------|
| `quality_80_union.db` | Main query DB: Gelbooru 80 selection union Danbooru 80 selection, deduped by md5 |
| `gelbooru.db` | Normalized Gelbooru source metadata |
| `danbooru2025_candidate.db` | Danbooru source metadata with md5 |
| `danbooru.db` | Older Danbooru DB retained for compatibility |

SQLite `-shm` and `-wal` files are runtime sidecars. They are not part of the
durable data model.

## quality_80_union.db

### images

2,988,704 rows. Primary key: `md5`.

| Column | Type | Meaning |
|--------|------|---------|
| `md5` | TEXT PK | File hash used for deduplication |
| `canonical_rating` | TEXT | Stricter merged rating |
| `canonical_file_url` | TEXT | Unified original file URL |
| `canonical_preview_url` | TEXT | Unified preview URL, mostly Gelbooru |
| `canonical_sample_url` | TEXT | Unified sample URL, mostly Gelbooru |
| `canonical_source` | TEXT | Unified source URL |
| `canonical_width` | INTEGER | Unified image width |
| `canonical_height` | INTEGER | Unified image height |
| `present_in_gelbooru` | INTEGER | 1 if this md5 exists in Gelbooru |
| `present_in_danbooru` | INTEGER | 1 if this md5 exists in Danbooru |
| `selected_by_gelbooru80` | INTEGER | 1 if selected by Gelbooru 80th-threshold rule |
| `selected_by_danbooru80` | INTEGER | 1 if selected by Danbooru 80th-threshold rule |
| `quality_score` | REAL | `max(gelbooru_score_pct, danbooru_score_pct)`, range `0..1` |
| `quality_tier` | TEXT | Current tier label, `p80` |
| `gelbooru_id` | INTEGER | Gelbooru post ID, nullable |
| `gelbooru_rating` | TEXT | Gelbooru rating, nullable |
| `gelbooru_score` | INTEGER | Gelbooru raw score, nullable |
| `gelbooru_score_pct` | REAL | Gelbooru rating-bucket percentile, nullable |
| `gelbooru_file_url` | TEXT | Gelbooru original file URL, nullable |
| `gelbooru_preview_url` | TEXT | Gelbooru preview URL, nullable |
| `gelbooru_sample_url` | TEXT | Gelbooru sample URL, nullable |
| `gelbooru_source` | TEXT | Gelbooru source URL/text, nullable |
| `gelbooru_tags_text` | TEXT | Gelbooru raw tag string, nullable |
| `gelbooru_status` | TEXT | Gelbooru post status, nullable |
| `danbooru_id` | INTEGER | Danbooru post ID, nullable |
| `danbooru_rating` | TEXT | Danbooru rating mapped to full names, nullable |
| `danbooru_score` | INTEGER | Danbooru raw score, nullable |
| `danbooru_score_pct` | REAL | Danbooru rating-bucket percentile, nullable |
| `danbooru_file_url` | TEXT | Danbooru original file URL, nullable |
| `danbooru_source` | TEXT | Danbooru source URL/text, nullable |
| `danbooru_pixiv_id` | INTEGER | Pixiv artwork ID from Danbooru, nullable |
| `danbooru_fav_count` | INTEGER | Danbooru favorite count, nullable |
| `danbooru_file_size` | INTEGER | Danbooru file size in bytes, nullable |
| `danbooru_file_ext` | TEXT | Danbooru file extension, nullable |
| `image_parquet_file` | TEXT | Future local image parquet shard filename, nullable |

Source-specific fields are nullable because the DB is a union. Use
`present_in_gelbooru` and `present_in_danbooru` to interpret which side fields
are expected to exist.
`image_parquet_file` is intentionally nullable until image parquet shards are
written by a later storage pipeline.

### URL Fields

| Field | Non-empty rows |
|-------|---------------:|
| `canonical_file_url` | 2,988,704 |
| `canonical_preview_url` | 2,696,467 |
| `canonical_sample_url` | 1,667,457 |
| `canonical_source` | 2,766,129 |
| `gelbooru_file_url` | 2,696,467 |
| `gelbooru_preview_url` | 2,696,467 |
| `gelbooru_sample_url` | 1,667,457 |
| `gelbooru_source` | 2,451,875 |
| `danbooru_file_url` | 2,428,840 |
| `danbooru_source` | 2,386,988 |

Use `canonical_file_url` for a single original image URL. Use source-specific
URL fields when the caller needs provenance.

### Tags

| Table | Meaning |
|-------|---------|
| `gelbooru_image_tags(md5, tag_name)` | Gelbooru tag associations |
| `danbooru_image_tags(md5, tag_name, tag_type)` | Danbooru tag associations |

### Quality Score

`quality_score` uses `rating_bucket_cume_dist_v1`.

```text
pct(score, rating) = count(score <= current_score within rating) / count(rating)
quality_score = max(gelbooru_score_pct, danbooru_score_pct)
```

Current verified counts:

| Metric | Count |
|--------|------:|
| Total images | 2,988,704 |
| `quality_score >= 0.8` | 2,973,255 |
| `quality_score < 0.8` | 15,449 |
| `quality_score = 1.0` | 9 |
| Selected by Gelbooru 80 | 2,034,916 |
| Selected by Danbooru 80 | 1,850,329 |
| Selected by both | 896,541 |
| Selected by neither | 0 |
| Present in both sources | 2,136,603 |

The DB is still the union of both raw 80th-threshold selections. The `<0.8`
rows are boundary rows caused by discrete raw-score buckets after true
percentile repair.

## gelbooru.db

Verified counts:

| Metric | Count |
|--------|------:|
| posts | 9,956,288 |
| tags | 931,306 |
| post_tags | 321,358,388 |

Main schema:

```text
posts(id, created_at, score, width, height, md5, directory, image,
      rating, source, change, owner, creator_id, parent_id, sample,
      preview_height, preview_width, tags, title, has_notes, has_comments,
      file_url, preview_url, sample_url, sample_height, sample_width,
      status, post_locked, has_children, attr_limit, attr_offset, attr_count)

tags(id, name)
post_tags(post_id, tag_id)
```

## danbooru2025_candidate.db

Verified counts:

| Metric | Count |
|--------|------:|
| posts | 9,113,284 |
| tags | 859,440 |
| post_tags | 330,953,940 |

Main schema:

```text
posts(id, pixiv_id, source, rating, score, fav_count,
      image_width, image_height, file_size, file_ext,
      md5, file_url, created_at, updated_at, uploader_id, parent_id)

tags(id, name, type)
post_tags(post_id, tag_id)
```

Danbooru raw ratings are `g`, `s`, `q`, `e`; the union DB maps them to
`general`, `sensitive`, `questionable`, `explicit`.

## Rebuild and Repair

```bash
/Users/charslee/miniconda3/bin/python3 convert_gelbooru.py --overwrite

/Users/charslee/miniconda3/bin/python3 convert_danbooru.py \
  --output-db /Users/charslee/Repo/private/pixiv_spider/danbooru2025_candidate.db

/Users/charslee/miniconda3/bin/python3 build_quality_union_db.py --overwrite

/Users/charslee/miniconda3/bin/python3 repair_quality_percentiles.py
```

## Known Technical Debt

- `quality_80_union.db` keeps staging tables:
  `selected_md5`, `selected_gelbooru_posts`, `selected_danbooru_posts`.
- `quality_80_union.db` keeps percentile lookup tables:
  `gel_score_percentiles`, `dan_score_percentiles`.
- Do not vacuum or rebuild the 26 GB union DB by default on low free disk
  space. A safe cleanup needs enough temporary room for SQLite to rewrite the
  database.