gcsfs: Retry after HttpError code 400

Google Cloud Storage occasionally throws an HTTP Error 400 (which is nominally a ‘bad request’. See the Google Cloud docs on HTTP response 400). But this happens on requests that have worked previously and work again after retrying. I’ve seen these spurious HTTP Error 400s when calling gcs.du and when using dask to read data from Google Cloud.

The error message from GCP is: Error 400 (Bad Request)! That's an error. Your client has issued a malformed or illegal request. That's all we know.

Monkey-patching gcsfs.utils.is_retriable fixes the issue for me:

import gcsfs

# Override is_retriable.  Google Cloud sometimes throws
# a HttpError code 400.  gcsfs considers this to not be retriable.
# But it is retriable!

def is_retriable(exception):
    """Returns True if this exception is retriable."""
    errs = list(range(500, 505)) + [
        # Jack's addition.  Google Cloud occasionally throws Bad Requests for no apparent reason.
        400,
        # Request Timeout
        408,
        # Too Many Requests
        429,
    ]
    errs += [str(e) for e in errs]
    if isinstance(exception, gcsfs.utils.HttpError):
        return exception.code in errs

    return isinstance(exception, gcsfs.utils.RETRIABLE_EXCEPTIONS)

gcsfs.utils.is_retriable = is_retriable

In a perfect world, I guess the best solution would be to ask Google Cloud to not throw spurious HTTP Error 400s. But perhaps a pragmatic approach is to modify gcsfs to retry after HTTP Error 400s 😃

Environment:

• Dask version: 2.28.0
• Python version: 3.8.5
• Operating System: Ubuntu 20.04 on a Google Cloud VM
• Install method: conda