Importing from urllib3's top-level location causes the namespace to be
urllib3.util.retry.Retry instead of
requests.packages.urllib3.util.retry.Retry. Without this fix, an using
requests with an un-vendored version of urllib3 will break when urllib3's
retry handling kicks in.
Closesshazow/urllib3#567
When working these changes back upstream to pip, we realized that the
previous fix wasn't ideal since unvendoring the packages broke the
imports. For example, if urllib3 were unvendored, then the following
would fail:
from requests.packages import urllib3
While discussion the issue, Donald Stufft (@dstufft) and I realized the
simplest solution is to simply add an alias per vendored dependency. The
resulting changes are simple and effective. It prevents the issue in
2.5.2 and 2.5.3 where the following would work:
from requests.packages import webbrowser
This now appropriately raises an ImportError.
Closes#2465
Instead of only checking one or another type of string-like object that
we accept, let's be able to check both. Previously, we only checked if
the filename was an instance of the native str type which on Python 2
excluded unicode filenames and bytes-like filenames on Python 3.
Fixes#2411
If a session runs long enough (without constant activity) then the server can
expire the nonce the session has negotiated. If that happens the session will
get a new 401 response which we were immediately returning to the user. A user
would then have to essentially reinitialize session.auth each time they get an
unexpected 401.
Also, there's no need for setattr calls when we can simply assign the
attribute on the instance.
e.g. a cherrypy uploaded file behave like a regular file, except that its name attribute is an int and passing it directly to requests fails because of that
RecentlyUsedContainers are threadsafe so they require a lock and as such
cannot be serialized with pickle directly. To handle it, we need to
convert it to a dictionary first and then back when deserializing.
Fixes#2345
Ian Cordasco
kennethreitz