Benoit Rostykus created ARROW-6844: -------------------------------------- Summary: List<scalar type> columns read broken with 0.15.0 Key: ARROW-6844 URL: https://issues.apache.org/jira/browse/ARROW-6844 Project: Apache Arrow Issue Type: Bug Components: C++, Python Affects Versions: 0.15.0 Reporter: Benoit Rostykus
Columns of type `array<primitive type>` (such as `array<int32>`, `array<int64>`...) are not readable anymore using `pyarrow == 0.15.0` (but were with `pyarrow == 0.14.1`) when the original writer of the parquet file is `parquet-mr 1.9.1`. ``` import pyarrow.parquet as pq pf = pq.ParquetFile('sample.gz.parquet') print(pf.read(columns=['profile_ids'])) ``` with 0.14.1: ``` pyarrow.Table profile_ids: list<element: int64> child 0, element: int64 ... ``` with 0.15.0: ``` Traceback (most recent call last): File "<string>", line 1, in <module> File "/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/pyarrow/parquet.py", line 253, in read use_threads=use_threads) File "pyarrow/_parquet.pyx", line 1131, in pyarrow._parquet.ParquetReader.read_all File "pyarrow/error.pxi", line 78, in pyarrow.lib.check_status pyarrow.lib.ArrowInvalid: Column data for field 0 with type list<item: int64> is inconsistent with schema list<element: int64> ``` I've tested parquet files coming from multiple tables (with various schemas) created with `parquet-mr`, couldn't read any `array<primitive type>` column anymore. I _think_ the bug was introduced with [this commit|[https://github.com/apache/arrow/commit/06fd2da5e8e71b660e6eea4b7702ca175e31f3f5]]. I think the root of the issue comes from the fact that `parquet-mr` writes the inner struct name as `"element"` by default (see [here|[https://github.com/apache/parquet-mr/blob/b4198be200e7e2df82bc9a18d54c8cd16aa156ac/parquet-column/src/main/java/org/apache/parquet/schema/ConversionPatterns.java#L33]]), whereas `parquet-cpp` (or `pyarrow`?) assumes `"item"` (see for example [this test|[https://github.com/apache/arrow/blob/c805b5fadb548925c915e0e130d6ed03c95d1398/python/pyarrow/tests/test_schema.py#L74]]). The round-tripping tests write/read in pyarrow only obviously won't catch this. -- This message was sent by Atlassian Jira (v8.3.4#803005)