[ARROW-13518] Identify selected row when using filters

• Type: New Feature
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: C++, Parquet, Python
• External issue URL: https://github.com/apache/arrow/issues/18768
• Assignee: Unassigned
• Reporter: Yair Lenga
• Created: 01/Aug/21 12:21
• Updated: 11/Jan/23 08:33

Description

I created a proposed enhancement to speed up reading of specific rows arrow-13517 https://issues.apache.org/jira/browse/ARROW-13517

proposing extending the functions that provides filter parquet.read_table (https://arrow.apache.org/docs/python/generated/pyarrow.parquet.read_table.html#pyarrow.parquet.read_table) to support returning actual row numbers (e.g, row_group and row_index). 

with the proposed enhancement, this can provide for faster reading of the data (e.g. by caching the return indices, and reading the full data when needed). 

proposed implementation will be to add 2 pseudo columns, which can be requested in the columns list. E.g., columns=[ ‘$row_group’, ‘$row_index’, ‘dealid’, …] or similar.

• $row_group - 0 based row group index
• $row_index - 0  based position within the row group
• $row_file_index - 0 based position in the file (not critical), can be constructed from the other two

 

not sure if this requires change to the c++ interface, or just to the python part of pyarrow.

 

Issue Links

• Dependent: ARROW-13517 Selective reading of rows for parquet file