ResultSet Arrow Tests¶

Tests for ResultSet.to_arrow().

The whole module skips when pyarrow is not installed.

Test Cases¶

1) to arrow basic types¶

name and n read back exactly; n, x, and name have integer, floating-point, and string Arrow types; ok is selected but not checked.

2) to arrow keeps int64 with nulls¶

A nullable integer stays an integer here; to_columns turns it into float64.

3) to arrow keeps bool with nulls¶

A nullable boolean stays a bool column rather than becoming a list.

4) to arrow nullable strings¶

Strings are wrapped from the offsets+blob buffer, nulls included.

5) to arrow multi batch¶

More rows than one batch: chunks must concatenate, not truncate.

6) to arrow multi batch int then float column¶

A column that is int in one batch and float in the next is a legal result set, since ArcadeDB is schemaless per document, and each batch's type is inferred on its own. The chunks are widened to float64 so they concatenate (#7108).

7) to arrow multi batch mixed type degrades to string¶

Numeric in one batch and a string in another has no common numeric type, so the column degrades to string rather than raising at concatenation.

8) to arrow multi batch non representable int forces string fallback¶

int64 values outside float64's exact ±2^53 range next to a batch of floats: pyarrow's cast is safe by default and refuses the lossy widening, so the column falls back to string instead of raising.

9) to arrow empty¶

An empty result is an empty table, not None and not an error.

10) to arrow matches to columns when not null¶

With no nulls the two paths must agree; only null handling differs.

Schemaless and DECIMAL data (`test_columnar_readers.py`)¶

to_columns(), to_dataframe(), and to_arrow() read rows over the bridge in batches, and these tests run at batch sizes 25,000, 2, and 1 because the fixes are about the answer not depending on where the batch boundaries fall. They need the bridge jar built from the tree (the wheel build does that); on the bridge jar and Python of 2026-10-02, 29 of the 31 fail.

a property the first row lacks is kept (humemai/arcadedb-embedded-python#113): three events where reason first appears on the second; to_columns() and to_arrow() return n, kind, reason with reason == [None, "timeout", "user"] at every batch size, and to_dataframe() keeps the column.
a late numeric property keeps its type: level absent from the first three rows; to_columns() is float64 with NaN, to_arrow() int64 with nulls, at every batch size.
Arrow types do not depend on batch size (#114): five readings where level is missing on rows 3 and 4 and tags is empty on rows 2 and 4; level is int64 and tags is list<string> at every batch size (they were strings at batch size 2 and 1), and to_columns() agrees.
a column of mixed types becomes strings in Arrow: an int in one row and 'seven' in the next, in one batch, raised a raw ArrowInvalid; it is now the string column ["1", "seven"].
a DECIMAL column keeps every digit (#115): a 36-digit amount, whole amounts only, whole and fractional, a whole amount above 2**63, and one with a null, each at batch size 25,000 and 1: an object array of exact Decimal from to_columns() and to_dataframe(), a decimal Arrow column from to_arrow().
batches of different scale unify: 2, 2.5, 30 at batch size 1 is one decimal type.
wider than decimal128 stays exact: a 41-digit value is decimal256(41, 1), an 80-digit one an exact string.
pinned columns are read as given: columns=["reason", "n"] returns exactly those, in that order, and a name no row has is an all-null column.
export_to_csv header is the union (#113): the header names every property of the rows; a column that first appears after the first batch of 10,000 rows raises ArcadeDBError naming it.

Running¶

uv run pytest bindings/python/tests/test_resultset_arrow.py -v
uv run pytest bindings/python/tests/test_columnar_readers.py -v