[
https://issues.apache.org/jira/browse/ARROW-12083?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17309029#comment-17309029
]
Weston Pace commented on ARROW-12083:
-------------------------------------
Related: ARROW-12080
Schema evolution (ARROW-11003) is also a closely related topic.
Allowing the schema to evolve as a dataset is read is challenging. The
single-file CSV reader does this and it comes with a fair amount of complexity
as well as CPU cost (have to go back and cast previous chunks). It's not even
"single-file" with datasets though. The streaming CSV reader uses the first
block (typically ~1MB) of the CSV file to determine the schema. A lot of the
dataset scanning, project, etc. code relies on knowing the schema up front.
It's probably an interesting question for the ML (or the query engine doc).
A pretty decent workaround is to store the evolved schema alongside your data
files as a really short IPC or parquet file. I believe Spark refers to these
as "summary files".
> [R] schema use in open_dataset
> ------------------------------
>
> Key: ARROW-12083
> URL: https://issues.apache.org/jira/browse/ARROW-12083
> Project: Apache Arrow
> Issue Type: Improvement
> Components: R
> Affects Versions: 3.0.0
> Environment: Windows
> Reporter: Shaun Nielsen
> Priority: Minor
>
> I have a directory of split .csvs that I'm importing with open_dataset().
> Between files, a column is imported as either int64 (e.g. -2) and the other
> string (1986CD), and this throws an error when {{unify_schemas = T}}
> {{ arrow::open_dataset('./split-csvs/nswcr/', format = 'csv', unify_schemas =
> T)}}
> {{Error: Invalid: Unable to merge: Field SEIFACalcMethod has incompatible
> types: int64 vs string}}
> If I use the schema parameter, and only want to specify this column, I only
> am able to import this column
> {{arrow::open_dataset('./split-csvs/nswcr/', }}{{format = 'csv', }}{{schema =
> schema(SEIFACalcMethod = string()))}}
> {{ }}
> {{FileSystemDataset with 45 csv files}}
> {{SEIFACalcMethod: string}}
> I was expecting that could set the class of a select few columns, while the
> rest would be imported as-is. Similar to readr::read_csv(col_types = cols())
> approach.
> Not sure if this is expected behaviour, a bug, or a possible avenue for
> improvement. I've tagged this as the latter. (y)
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
