HTTP error 504 Server Error: Gateway Time-out

[illiyas-sha]

I am trying to push a large dataset(450000+) records with the help of push_to_hub()
While pushing, it gives some error like this.

Traceback (most recent call last):
  File "data_split_speech.py", line 159, in <module>
    data_new_2.push_to_hub("user-name/dataset-name",private=True)
  File "/opt/conda/lib/python3.8/site-packages/datasets/dataset_dict.py", line 951, in push_to_hub
    repo_id, split, uploaded_size, dataset_nbytes = self[split]._push_parquet_shards_to_hub(
  File "/opt/conda/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 3556, in _push_parquet_shards_to_hub
    api.upload_file(
  File "/opt/conda/lib/python3.8/site-packages/huggingface_hub/hf_api.py", line 1017, in upload_file
    raise err
  File "/opt/conda/lib/python3.8/site-packages/huggingface_hub/hf_api.py", line 1008, in upload_file
    r.raise_for_status()
  File "/opt/conda/lib/python3.8/site-packages/requests/models.py", line 953, in raise_for_status
    raise HTTPError(http_error_msg, response=self)
requests.exceptions.HTTPError: 504 Server Error: Gateway Time-out for url: https://huggingface.co/api/datasets/user-name/dataset-name/upload/main/data/train2-00041-of-00064.parquet

Can anyone help me to resolve this issue.

[julien-c]

is pushing directly with git (and git-lfs) an option for you?

[illiyas-sha]

I have installed git-lfs and doing this push with that

[julien-c]

yes but is there any way you could try pushing with git command line directly instead of push_to_hub?

[illiyas-sha]

Okay. I didnt saved the dataset to my local machine. So, I processed the dataset and pushed it directly to the hub. I think I should try saving those dataset to my local machine by save_to_disk and then push it with git command line

[julien-c]

cc @lhoestq @albertvillanova @LysandreJik because maybe I'm giving dumb advice here 😅

[lhoestq]

push_to_hub is the preferred way of uploading a dataset to the Hub, which can then be reloaded with load_dataset. Feel free to try again and see if the server is working as expected now. Maybe we can add a retry mechanism in the meantime to workaround 504 errors.

Regarding save_to_disk, this must only be used for local serialization (because it's uncompressed and compatible with memory-mapping). If you upload a dataset saved with save_to_disk to the Hub, then to reload it you will have to download/clone the repository locally by yourself and use load_from_disk.