本页尚未翻译

本页内容仍为英文原文，源自官方文档 geneva/udfs/index。中文翻译正在分批进行中。

Understanding Transforms

Understand the three types of user-defined functions in Geneva — UDFs, chunkers (scalar UDTFs), and batch UDTFs — and when to use each.

Geneva provides three types of user-defined functions for transforming data. Each type has a different input/output cardinality and is suited to different workflows.

Choosing the Right Type

• Adding a column to each row? Use a UDF.
• Splitting each row into multiple rows? Use a Chunker.
• Computing across rows with a different output shape? Use a Batch UDTF.

At a Glance

	UDF	Chunker (Scalar UDTF)	Batch UDTF
Cardinality	1:1	1:N	N:M
Decorator	@udf	@chunker	@udtf
Refresh	Incremental	Incremental	Full
Parallelism	Fragment-parallel	Fragment-parallel	Partition-parallel
Inherited columns	N/A — adds to existing rows	Automatic from query	Independent output schema
Registration	table.add_columns()	db.create_udtf_view()	db.create_udtf_view()

UDFs (1:1)

Standard UDFs produce exactly one output value per input row. Use them to add computed columns to existing tables or materialized views.

id	text	embedding
1	"hello world"	→ [0.12, 0.34, ...]
2	"foo bar"	→ [0.56, 0.78, ...]
3	"baz qux"	→ [0.90, 0.11, ...]

Each input row produces exactly one output value. The new column is added to the same table.

Use cases: Embeddings, data enrichment, format conversion, scoring.

See UDFs for the full guide.

Chunkers (Scalar UDTFs, 1:N)

Chunkers — also called scalar UDTFs — expand each source row into multiple output rows. The output is a materialized view that inherits parent columns and supports incremental refresh.

Source: documents

doc_id	title	text
1	"Intro to AI"	"Machine learning is..."
2	"Data Guide"	"Data pipelines are..."

Derived: chunks (1:N expansion via @chunker)

doc_id	title	chunk_index	chunk_text
1	"Intro to AI"	0	"Machine learning..."
1	"Intro to AI"	1	"Neural networks..."
1	"Intro to AI"	2	"Training data..."
2	"Data Guide"	0	"Data pipelines..."
2	"Data Guide"	1	"ETL processes..."

Each source row produces one or more output rows. Parent columns (doc_id, title) are inherited automatically.

Use cases: Document chunking, video segmentation, image tiling.

See Chunkers for the full guide.

Batch UDTFs (N:M)

Batch UDTFs read from a source table (or partition) and produce output with an arbitrary schema and row count. They always perform a full refresh.

Source: sales

product	region	amount
Widget	East	100
Widget	East	250
Widget	West	175
Gadget	East	300
Gadget	West	400
Gadget	West	150

Derived: sales_summary (N:M aggregation via @udtf)

product	total_amount	avg_amount	num_sales
Widget	525	175.0	3
Gadget	850	283.3	3

6 input rows become 2 output rows with a completely different schema. The output shape is determined entirely by the UDTF logic — it could be fewer rows (aggregation), more rows (clustering), or the same count with different columns.

Use cases: Deduplication, clustering, aggregation, cross-row joins.

See Batch UDTFs for the full guide.

API Reference

• UDF — @udf decorator and UDF class
• UDTF — @udtf, @chunker, @batch_udtf decorators and UDTF/Chunker classes
• Table — add_columns(), backfill()
• Connection — create_udtf_view(), create_materialized_view()