Weston Pace created ARROW-9935:
----------------------------------
Summary: New filesystem API unable to read empty S3 folders
Key: ARROW-9935
URL: https://issues.apache.org/jira/browse/ARROW-9935
Project: Apache Arrow
Issue Type: Bug
Affects Versions: 1.0.0
Reporter: Weston Pace
Attachments: arrow_453.py, arrow_9935.py
When an empty "folder" is created in S3 using the online bucket explorer tool
on the management console then it creates a special empty file with the same
name as the folder.
(Some more details here:
[https://docs.aws.amazon.com/AmazonS3/latest/user-guide/using-folders.html)]
If parquet files are later loaded into one of these directories (with or
without partitioning subdirectories) then this dataset cannot be read by the
new dataset API. The underlying s3fs `find` method returns a "file" object
with size 0 that pyarrow then attempts to read. Since this file doesn't truly
exist a FileNotFoundError is thrown.
Would it be safe to simply ignore all files with size 0?
As a workaround I can wrap s3fs' find method and strip out these objects with
size 0 myself.
I've attached a script showing the issue and a workaround. It uses a public
bucket that I'll leave up for a few months.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
