Null list<struct<...>> is written and read as an empty list

[slachiewicz]

Apache Iceberg version

main (development)

Please describe the bug 🐞

A null list<struct<...>> is silently rebuilt as an empty list. This is known on
the read path — tests/integration/test_reads.py::test_null_list_and_map currently
asserts the corrupted value, with the correct assertion commented out pending
apache/arrow#38809:

# This should be:
# assert arrow_table["col_list_with_struct"].to_pylist() == [None, [{'test': 1}]]
# Once https://github.com/apache/arrow/issues/38809 has been fixed
assert arrow_table["col_list_with_struct"].to_pylist() == [[], [{"test": 1}]]

Two things seem worth reporting on top of that.

It also affects the write path, where the consequence is worse. The Parquet file
pyiceberg writes contains an empty list, so the null is gone at rest and no reader —
pyiceberg, Spark, Trino — can recover it. On read the file is at least still correct.

It does not depend on the upstream Arrow fix. pa.LargeListArray.from_arrays
takes a mask argument — since well before pyiceberg's pyarrow>=18.0.0 floor — so
this particular null loss can be fixed independently of apache/arrow#38809, which is
still open.

This is the array<struct<>> case from #251. That issue was closed in March 2025 on
the strength of this test existing, but the assertion it makes is the corrupted one;
the array<int> case in the issue body was genuinely fixed by #252, while the
array<struct<test:int>> case in the issue title — which @HonahX flagged as
remaining broken in
#252 (comment) — was not.

Reproduction (write path)

pyiceberg 0.11.1, pyarrow 25.0.1:

import os, shutil, glob
import pyarrow as pa, pyarrow.parquet as pq
from pyiceberg.catalog.sql import SqlCatalog

WH = "/tmp/wh"; shutil.rmtree(WH, ignore_errors=True); os.makedirs(WH)

sch = pa.schema([
 pa.field("id", pa.int32(), nullable=False),
 pa.field("l_struct", pa.list_(pa.field("element", pa.struct([pa.field("x", pa.int32())]), nullable=True)), nullable=True),
 pa.field("l_int", pa.list_(pa.field("element", pa.int32(), nullable=True)), nullable=True),
])
tbl = pa.table({"id": [1, 2, 3, 4],
 "l_struct": [[{"x": 1}], [], None, [{"x": 3}]],
 "l_int": [[1], [], None, [3]]}, schema=sch)

cat = SqlCatalog("r", uri=f"sqlite:///{WH}/c.db", warehouse=f"file://{WH}")
cat.create_namespace("ns")
it = cat.create_table("ns.t", schema=tbl.schema)
it.append(tbl)

out = it.scan().to_arrow()
for c in ("l_struct", "l_int"):
 print(f"{c:9s} in={tbl.column(c).to_pylist()!s:35s} out={out.column(c).to_pylist()}")

# the loss is already in the file on disk, not in the read path
f = glob.glob(f"{WH}/**/*.parquet", recursive=True)[0]
print("raw parquet:", pq.read_table(f).column("l_struct").to_pylist())

Output:

l_struct in=[[{'x': 1}], [], None, [{'x': 3}]] out=[[{'x': 1}], [], [], [{'x': 3}]]
l_int in=[[1], [], None, [3]] out=[[1], [], None, [3]]
raw parquet: [[{'x': 1}], [], [], [{'x': 3}]]

l_int round-trips correctly, and writing the same pa.Table with pq.write_table
preserves the null, so the loss is not pyarrow's.

Cause

ArrowProjectionVisitor.list rebuilds the array when the element is a struct
(pyiceberg/io/pyarrow.py:2078 on main @ 7539661):

if isinstance(value_array, pa.StructArray):
 # This can be removed once this has been fixed:
 # https://github.com/apache/arrow/issues/38809
 list_array = pa.LargeListArray.from_arrays(list_array.offsets, value_array)

from_arrays receives the offsets buffer alone, which cannot express a null list, and
no mask, so the validity bitmap is dropped. That is also why only this one shape is
affected: the struct visitor passes mask=struct_array.is_null(), the map visitor
does not rebuild at all, and a list whose element is a primitive never enters this
branch. The visitor runs on both paths, which is why the same root cause shows up as
the read-side assertion above and as the write-side corruption here.

Fix

Carrying the mask over is enough:

list_array = pa.LargeListArray.from_arrays(list_array.offsets, value_array, mask=list_array.is_null())

With that change the reproduction above returns None for both columns, and
test_null_list_and_map passes with its commented-out assertion restored. I have not
looked at int32-offset or sliced-array handling of this call, which the existing line
already relies on; that appears independent of the mask.

I have this on a branch with a unit test covering the write path and the integration
assertion un-commented, and can open a PR.

Found while testing a third-party Iceberg writer against pyiceberg as a reader.

Willingness to contribute

• I can contribute a fix for this bug independently