[
https://issues.apache.org/jira/browse/ARROW-18001?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17616413#comment-17616413
]
Joris Van den Bossche commented on ARROW-18001:
-----------------------------------------------
Some background: in this specific case when using the _pandas_ (or dask)
{{to_parquet}} method, the {{schema}} keyword gets passed to
{{Table.from_pandas}}, and not the actual parquet write methods.
In general, the type inference happens when converting your python object (eg
pandas dataframe, or a dict, ..) to an Arrow Table, and once you have such
table with a fixed schema, writing to Parquet doesn't do type inference anymore
(since arrow types map to parquet types).
So I think we should reframe the issue as providing a way to specify the type
of a subset of columns for {{from_pandas}}.
Doing a small search for other JIRAs, I noticed that at some point in the past
we actually did support a partial schema (this was accidentally broken at some
point and then fixed again: ARROW-1125, although in the PR it was already noted
that we might prefer doing this in another way:
https://github.com/apache/arrow/pull/790#discussion_r124543809).
Afterwards, the behaviour was changed again (intentionally) in ARROW-3766, now
honoring the exact schema as passed
(https://github.com/apache/arrow/pull/2979#discussion_r234010810)
> [Python] parquet.write_table/parquet.ParquetWriter should except a subset of
> columns
> ------------------------------------------------------------------------------------
>
> Key: ARROW-18001
> URL: https://issues.apache.org/jira/browse/ARROW-18001
> Project: Apache Arrow
> Issue Type: Improvement
> Components: Python
> Reporter: Alenka Frim
> Priority: Major
>
> This question came up in the GitHub issue:
> [https://github.com/apache/arrow/issues/14025] and it would be a good
> improvement to the Parquet part of PyArrow. Haven't found any existing issue
> and so created a new one.
> h6. Description:
> If a user wants to change a type of one single column when using
> {{{}parquet.write_table{}}}/{{{}parquet.ParquetWriter{}}} they currently need
> to specify the schema with all columns included. If a column is not specified
> in the schema, it will not be included in the parquet file.
> h6. Proposal
> There should be a possibility for {{parquet.ParquetWriter}} excepting a
> subset of columns in a Schema and infer everything else.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
