# Transformer — Funciones que limpian, refinan o agregan datos

Tag: `transformer`. Grupo de funciones que **transforman datos**: clean, dedup, aggregate, feature-engineer, filter, impute, normalize. Input + output ambos datos (no efectos externos). Segundo eslabon del flujo en `data_factory` (analogia Factorio = assemblers).

Filtro MCP: `mcp__registry__fn_search query="" tag="transformer"`.

## Funciones del grupo

| ID | Lang | Que hace |
|---|---|---|
| [aggregate_by_group_py_datascience](../../python/functions/datascience/aggregate_by_group.md) | py | GROUP BY + agg |
| [deduplicate_entities_py_datascience](../../python/functions/datascience/deduplicate_entities.md) | py | Dedup entities |
| [deduplicate_relations_py_datascience](../../python/functions/datascience/deduplicate_relations.md) | py | Dedup relations |
| [align_relations_to_entities_py_datascience](../../python/functions/datascience/align_relations_to_entities.md) | py | Reconciliacion |
| [clip_py_datascience](../../python/functions/datascience/clip.md) | py | Cap valores fuera de rango |
| [clip_go_datascience](../../functions/datascience/clip.md) | go | Cap valores fuera de rango |
| [impute_py_datascience](../../python/functions/datascience/impute.md) | py | Rellenar nulls |
| [impute_go_datascience](../../functions/datascience/impute.md) | go | Rellenar nulls |
| [histogram_py_datascience](../../python/functions/datascience/histogram.md) | py | Bin frequencies |
| [histogram_go_datascience](../../functions/datascience/histogram.md) | go | Bin frequencies |
| [group_by_go_datascience](../../functions/datascience/group_by.md) | go | Group rows by key |
| [coerce_types_py_core](../../python/functions/core/coerce_types.md) | py | Convert column types |
| [csv_to_parquet_duckdb_py_core](../../python/functions/core/csv_to_parquet_duckdb.md) | py | Format conversion |
| [diff_entities_py_datascience](../../python/functions/datascience/diff_entities.md) | py | Compare entity snapshots |
| [diff_relations_py_datascience](../../python/functions/datascience/diff_relations.md) | py | Compare relation snapshots |

## Ejemplo canonico

Pipeline limpiar dataframe: impute nulls -> clip outliers -> dedup -> aggregate.

```python
from datascience import impute, clip, deduplicate_entities, aggregate_by_group

# 1. Rellenar nulls
df = impute(df, columns=["price"], strategy="median")

# 2. Cap valores extremos
df = clip(df, column="price", lower=0.0, upper=10000.0)

# 3. Deduplicar por clave
df = deduplicate_entities(df, key_columns=["sku"])

# 4. Agregar por categoria
out = aggregate_by_group(df, group_by=["category"], aggs={"price": "mean", "qty": "sum"})
```

## Fronteras del grupo

NO cubre:
- **Extract** (leer de fuente externa) -> [[extractor]].
- **Sink** (escribir a destino externo) -> [[sink]].
- **Validacion** (range/null/drift checks) -> [[validator]]. Transformers ASUMEN datos validos en entrada.
- ML training (eso es `trainer`, deferred a v2 de data_factory).

## Cuando NO usar `transformer`

- Si la funcion solo lee/escribe pero no transforma -> es extractor o sink.
- Si solo verifica una condicion (bool/list) -> es validator.
- Si solo encadena varias funciones -> es `pipeline` (kind), no transformer plano.

## Pureza

La mayoria de transformers son **puros** (mismo input -> mismo output). Algunos son impuros por usar SDKs con caches (duckdb, polars). El tag `transformer` no impone pureza — verifica `purity` en frontmatter de cada funcion.

## Consumidores

- `data_factory` — tab Transformers.
- `dag_engine` DAG steps con `function: <transformer_id>`.
- Notebooks analysis/ pipelines.

## Notas

- Transformers Go/Py duplicados intencionalmente — eleccion por stack y rendimiento.
- Si el registry incorpora `polars`/`duckdb` masivamente, considerar crear sub-grupos.