A blank cell makes CSV File Scan infer a numeric column as STRING, breaking every downstream numeric operator

[kz930]

What happened?

One blank cell in a numeric CSV column makes CSV File Scan type the whole column as STRING, so every downstream operator that does arithmetic on it fails, including on the rows that hold perfectly good numbers.

The cause is that the schema inference and the execution disagree about what a blank cell is. CSVScanSourceOpDesc.scala:116 sets csvSetting.setNullValue("") before inferring, so a blank reads as an empty string: tryParseDouble("") fails, tryParseBoolean("") fails, and inferField lands on tryParseString(). CSVScanSourceOpExec.scala:95-100 builds its parser without setNullValue, so at execution the same blank reads as null, which is what AttributeTypeUtils.parseField is written to pass through.

Either convention is defensible on its own. What breaks is having one for the schema and another for the data, and the schema side is the one that silently downgrades a numeric column.

Two files that differ only in one blank cell take the same workflow down two different paths:

id,petal_length,petal_width      id,petal_length,petal_width
1,2.6,0.75                       1,2.6,0.75
2,,1.3                           2,3.9,1.3
3,5.2,1.85                       3,5.2,1.85

With the blank, Iris Logistic Regression takes 2 rows, emits 1 and pauses. Without it, the same operator with the same configuration takes 3 and emits 3. The failure is on row 1, whose values are 2.6 and 0.75, because they arrive as array([['2.6', '0.75']], dtype='<U32') rather than as numbers.

Expected: a blank cell should not change the inferred type of a column. Inference and execution should agree on what a blank is.

How to reproduce?

Upload the two CSVs above. Build CSV File Scan to Hugging Face Iris Logistic Regression with Petal Length Cm Attribute = petal_length and Petal Width Cm Attribute = petal_width. Run it once against each file. The one with the blank cell fails on its first row; the one without it completes.

Any operator that does arithmetic on a scanned numeric column shows the same thing; Iris is just the shortest workflow that reaches it.

Version/Branch

1.3.0-incubating-SNAPSHOT (main)

Relevant log output

2026-08-11 00:51:41.181 | ERROR | core.architecture.managers.context:report_exception:104 - ufunc 'subtract' did not contain a loop with signature matching types (dtype('<U32'), dtype('float64')) -> None
                 │          └ [3.72666667, 1.17619048]
                 └ array([['2.6', '0.75']], dtype='<U32')

numpy._core._exceptions._UFuncNoLoopError: ufunc 'subtract' did not contain a loop with signature matching types (dtype('<U32'), dtype('float64')) -> None