HTTP 500 error in fetch_mldata mauna-loa-atmospheric-co2

[rth]

There have been several PRs (#11100 (review), #11106) where CircleCI arbitrarly fails due to HTTP 500 errors when calling fetch_mldata('mauna-loa-atmospheric-co2'),

Partial traceback below,

Traceback (most recent call last):
  File "/home/circleci/project/examples/gaussian_process/plot_gpr_co2.py", line 75, in <module>
    data = fetch_mldata('mauna-loa-atmospheric-co2').data
  File "/home/circleci/project/sklearn/datasets/mldata.py", line 154, in fetch_mldata
    mldata_url = urlopen(urlname)
  File "/home/circleci/miniconda/envs/testenv/lib/python3.6/urllib/request.py", line 223, in urlopen
[...]

  File "/home/circleci/miniconda/envs/testenv/lib/python3.6/urllib/request.py", line 650, in http_error_default
    raise HTTPError(req.full_url, code, msg, hdrs, fp)
urllib.error.HTTPError: HTTP Error 500: Internal Server Error

If this keeps repeating, a possible workaround could be,

• increasing the number of download attempts in fetch_mldata
• investigating the failures upstream with mldata.org
• copying this particular dataset to figshare (Upload all datasets we have loaders for to figshare? #7425) which seems to have a better quality of service and adding a fallback URL there used if the download from the mldata website fails..

[lesteve]

My understanding is that the long-term solution is the openml fetcher #9543 (not 100% sure what the status is).

mldata.org has historically not been extremely reliable but if this is just temporary glitches I would say we should ignore them as we have done so far. The feeling I got when investigating #8588 is that mldata.org maintenance is not very active (no disrespect intended, just saying that there is not a staff of 10 full-time people behind it). Edit: more details about who maintains mldata.org: #8588 (comment).

If it starts to be too annoying to be ignored, we could probably implement a retry mechanism, but someone should double-check that it actually fixes the problem. For example when a glitch happens it may actually last for a few minutes, in which case a retry mechanism may not be a great fit.

[qinhanmin2014]

Apart from fetch_mldata('mauna-loa-atmospheric-co2'), fetch_mldata('MNIST original') also fails in #11106.
I'm +1 for ignoring the issue since I think it doesn't occur very often (in the previous year at least).
I also doubt whether the retry mechanism will introduce unnecessary complexity for the examples.

[rth]

Agreed. Let's close this for now and re-open later if needed.

[amueller]

@lesteve status is it's on my todo and I'm back from the dead (aka teaching)

[jnothman]

or completing the openml PR and loading from there

…

On 18 May 2018 9:26 pm, "Roman Yurchak" ***@***.***> wrote: There have been several PRs (#11100 (review) <#11100 (review)>, #11106 <#11106>) where CircleCI arbitrarly failes due to HTTP 500 errors when calling fetch_mldata('mauna-loa-atmospheric-co2'), Partial traceback below, Traceback (most recent call last): File "/home/circleci/project/examples/gaussian_process/plot_gpr_co2.py", line 75, in <module> data = fetch_mldata('mauna-loa-atmospheric-co2').data File "/home/circleci/project/sklearn/datasets/mldata.py", line 154, in fetch_mldata mldata_url = urlopen(urlname) File "/home/circleci/miniconda/envs/testenv/lib/python3.6/urllib/request.py", line 223, in urlopen [...] File "/home/circleci/miniconda/envs/testenv/lib/python3.6/urllib/request.py", line 650, in http_error_default raise HTTPError(req.full_url, code, msg, hdrs, fp)urllib.error.HTTPError: HTTP Error 500: Internal Server Error If this keeps repeating, possible workaround could be, - increasing the number of download attemps in fetch_mldata - investigating the failures upstream with mldata.org - copying this particular dataset to figshare (#7425 <#7425>) which seems to have a better quality of service and adding a fallback URL there <https://github.com/scikit-learn/scikit-learn/blob/a24c8b464d094d2c468a16ea9f8bf8d42d949f84/sklearn/datasets/mldata.py#L29> used if the download from the mldata website fails.. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#11108>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz6xmCMKlMcJsxGDe8KlwDO0mYi-0Rks5tzq_UgaJpZM4UEkg6> .