[ARROW-9325] [C++][Dataset][Python] ParquetDataset typecast on read

• Type: New Feature
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 0.17.1
• Component/s: C++, Python
• Labels: dataset
• External issue URL: https://github.com/apache/arrow/issues/25411
• Assignee: Unassigned
• Reporter: Carson Eisenach
• Created: 04/Jul/20 19:38
• Updated: 11/Jan/23 08:06

Description

When reading large Parquet tables, it would be useful to have the option to cast columns to a different type. Consider a large table with double precision types (float64 and int64), the user might prefer to read these in as single precision if double precision is not required. 

Current behavior: One must first read the table and then cast

Desired behavior: provide an additional kwarg that allows the user to specify a target schema. This would be propagated through to ParquetFileFragment, and each fragment can be cast as soon as it is read.

Impact: In cases where the user wants to cast all columns to single precision and the dataset has many partitions, this feature would reduce max memory required by roughly 50%.

--------------

I've already implemented a POC using the old Dataset API, and can reimplement using the v2 dataset API, and then submit a patch.

A couple questions:

1. Does this feature fit in with the Arrow roadmap?

2. Alternatively, is there a way to accomplish this already in v0.17 that I am missing?