Richard Xia created LIBCLOUD-903:
------------------------------------
Summary: AWS S3 upload_object_via_stream fails on non-file
iterable due to missing Content-Length header
Key: LIBCLOUD-903
URL: https://issues.apache.org/jira/browse/LIBCLOUD-903
Project: Libcloud
Issue Type: Bug
Reporter: Richard Xia
The issue I am seeing appears to be due to the incorrect integration of 4
separate libraries, but I believe the real problem is here in libcloud, in the
{{upload_object_via_stream()}} method on the S3 storage driver.
I am using Python 3.5.1 and the the four libraries I am using are:
* Django 1.10.6
* django-storages 1.5.2
* libcloud v2.0.0rc1-tentative
* requests 2.13.0
Specifically, when I try to use a Django
[ContentFile|https://docs.djangoproject.com/en/1.10/ref/files/file/#django.core.files.base.ContentFile],
Django's own file-like wrapper for strings, to save a new file to S3 via the
Libcloud backend of django-storages, I get the following error:
{code:xml}
<?xml version="1.0"
encoding="UTF-8"?>\n<Error><Code>NotImplemented</Code><Message>A header you
provided implies functionality that is not
implemented</Message><Header>Transfer-Encoding</Header><RequestId>A2FC4D5109083076</RequestId><HostId>K9WGhd18iqQHyIyv+GxWcxHexvapVSidTtHzSqujtT9nT5LhmIEygMKOfR/7F0v7ujnlE/CoYiM=</HostId></Error>
{code}
The reason this happens is because Libcloud is generating an HTTP request to
AWS S3 that is missing the {{Content-Length}} header. AWS S3 requires the
{{Content-Length}} header for file uploads *unless* if it is a multi-part
upload. This is why this used to work on the 1.5.0 release of {{libcloud}},
because even single-part uploads were done as a one-part multi-part upload.
I've traced my bug down through all four libraries and have determined exactly
why the {{Content-Length}} header is missing in my particular use case. The
{{upload_object_via_stream()}} has an {{iterator}} argument that should yield
the content body data, and it eventually passes that argument directly to the
{{requests}} library. The {{requests}} library will actually [try very hard to
add the {{Content-Length}}
header|https://github.com/kennethreitz/requests/blob/c43fefa7ed535c41ba7d58021f0f16ed5ba1d584/requests/models.py#L471],
even for certain types of iterator streams. In particular it can determine the
length of file-like objects which support stat operations and it can handle
StringIO/BytesIO objects. However, the Django {{ContentFile}} is neither, and
{{requests}} cannot extract the length of the stream without consuming the
iterator, so it does not try.
Here's some (Python 3) code to demonstrate the bug:
{code:python}
from io import BytesIO
class MyWrapper(object):
"""A contrived wrapper that acts similar to BytesIO."""
def __init__(self, content):
self.content = BytesIO(content)
def __iter__(self):
self.content.seek(0)
yield self.content.read()
# Assume driver is already set to some S3 provider w/ credentials
container = driver.get_container(container_name='my-container')
driver.upload_object_via_stream(iterator=iter(MyWrapper(b'hello world')),
container=container,
object_name='my_file.txt')
{code}
I think the proper solution to this will require all calls to the S3
{{upload_object_via_stream()}} to use the multi-part uploader in order to
eschew the need for the {{Content-Length}} header. If desired, you could make
the same optimizations that the request library makes by checking for certain
common cases where you do know the file size and only using the multi-part
uploader when necessary.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
