huggingface / huggingface/datasets

Errors loading dataset using fs = a gcsfs.GCSFileSystem

Open
#3,034 0 comments 0 reactions 0 assignees View on GitHub
bug
Dominant language
Python
Stars
22k
Forks
3.4k
Avg merge
5d 7h
Merged PRs (30d)
17

Description

## Describe the bug

Cannot load dataset using a `gcsfs.GCSFileSystem`. I'm not sure if this should be a bug in `gcsfs` or here...

Basically what seems to be happening is that since datasets saves datasets as folders and folders aren't "real objects" in gcs, gcsfs raises a 404 error. There are workarounds if you use gcsfs directly to download the file, but as is I can't get `load_from_disk` to work.

## Steps to reproduce the bug
```python
from datasets import load_dataset
# load some dataset
dataset = load_dataset("squad", split="train")

# save it to gcs
import gcsfs
fs = gcsfs.GCSFileSystem(project="my-gs-project")
dataset.save_to_disk("gs://my-bucket/squad", fs=fs)

# try to load it from gcs
from datasets import load_from_disk
dataset2 = load_from_disk("my-bucket/squad", fs=fs)
```

## Expected results
`dataset2` would be a copy of `dataset` but loaded from my bucket.

## Actual results
Long traceback but essentially it's a 404 error from gcsfs saying the object `my-bucket/squad` doesn't exist when this is called:
https://github.com/huggingface/datasets/blob/9c81b7d2e6d9feae69a084a3abda265a4ca07fb5/src/datasets/arrow_dataset.py#L977

This is because there is no actual object called `my-bucket/squad`, there are objects called `my-bucket/squad/dataset.arrow`, etc.

Note that *this* works fine, since it's explicitly saying "download all the objects with this prefix":
```python
fs.download(src_dataset_path + "/*", dataset_path.as_posix(), recursive=True)
```
For example, I can do a workaround this way:
```python
import tempfile

with tempfile.TemporaryDirectory() as temppath:
fs.download("gs://my-bucket/squad/*", temppath)
dataset2 = load_from_disk(temppath)
```

It's unclear to me if it's `gcsfs`'s responsibility to say "hey that's folder not a file, I should try to get objects inside of it not the object itself", or if that's `datasets`'s responsibility... I'm leaning towards the latter since you're never loading a dataset from one file using this function/method, only a dataset folder?

Another minor thing that should maybe should be rolled into this bug...

https://github.com/huggingface/datasets/blob/9c81b7d2e6d9feae69a084a3abda265a4ca07fb5/src/datasets/arrow_dataset.py#L968

These fail if you pass in a `gs://` path, e.g.

```python
dataset2 = load_from_disk("gs://my-bucket/squad", fs=fs)
```
Because at this point, `dataset_info_path` is `gs:/my-bucket/squad/dataset_info.json`, gcsfs throws a:
```
Invalid bucket name: 'gs:'
```
error

## Environment info

- `datasets` version: 1.12.1
- Platform: macOS Big Sur 11.6
- Python version: 3.7.12
- PyArrow version: 5.0.0

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.