-
Notifications
You must be signed in to change notification settings - Fork 4k
Closed
Description
I found that the following does not roundtrip:
[('generated_primitive', DataType(uint32)), ('generated_primitive', DataType(uint32))]
[('generated_primitive_no_batches', DataType(uint32)), ('generated_primitive_no_batches', DataType(uint32))]
[('generated_primitive_zerolength', DataType(uint32)), ('generated_primitive_zerolength', DataType(uint32))]The exact code I am using for this
import os
import pyarrow.ipc
import pyarrow.parquet as pq
def get_file_path(file: str):
return f"../testing/arrow-testing/data/arrow-ipc-stream/integration/1.0.0-littleendian/{file}.arrow_file"
def _expected(file: str):
return pyarrow.ipc.RecordBatchFileReader(get_file_path(file)).read_all()
def check_file(file):
expected = _expected(file)
path = f"{file}.parquet"
pq.write_table(expected, path, compression=None, write_statistics=False)
table = pq.read_table(path)
os.remove(path)
failing = []
for c1, c2 in zip(expected, table):
if c1 != c2:
failing.append((file, c1.type))
return failing
for file in [
"generated_primitive",
"generated_primitive_no_batches",
"generated_primitive_zerolength",
"generated_null",
"generated_null_trivial",
"generated_primitive_large_offsets",
]:
failing = check_file(file)
if failing:
print(failing)Note: I generated the same parquet using the experimental parquet2 and the roundtrip succeeds, suggesting that the potential error is in writing.
Upon further investigation, it seems that the only difference is the type: c1's type is uint32, c2's type is int64.
Reporter: Jorge Leitão / @jorgecarleitao
Related issues:
- [C++][Python] Switch default Parquet version to 2.4 (is superceded by)
Note: This issue was originally created as ARROW-13214. Please see the migration documentation for further details.