`PreparedRequest.prepare_url` incorrectly applied IDNA encoding to the
URLs entire `netloc`. It should only be encoding the hostname portion
of the URL. IDNA encoding was limiting the user info, host, and port
segments to be a maximum of 63 characters which causes problems for
all by the most trivial user + password combinations.
- Replaced usage of `urlparse` in `PreparedRequest.prepare_url` with
`urllib3` equivalent.
- Modified IDNA encoding section so that it only encodes the host
portion of the URL.
This allows for blocking when using threading to prevent the pool from creating more connections that the max-size allows.
Specifically was seeing the following errors without the block=True:
WARNING:requests.packages.urllib3.connectionpool:HttpConnectionPool is full, discarding connection: www.example.com
Fixes#649 and #1329 by making Session.headers a CaseInsensitiveDict,
and fixing the implementation of CID. Credit for the brilliant idea
to map `lowercased_key -> (cased_key, mapped_value)` goes to
@gazpachoking, thanks a bunch.
Changes from original implementation of CaseInsensitiveDict:
1. CID is rewritten as a subclass of `collections.MutableMapping`.
2. CID remembers the case of the last-set key, but `__setitem__`
and `__delitem__` will handle keys without respect to case.
3. CID returns the key case as remembered for the `keys`, `items`,
and `__iter__` methods.
4. Query operations (`__getitem__` and `__contains__`) are done in
a case-insensitive manner: `cid['foo']` and `cid['FOO']` will
return the same value.
5. The constructor as well as `update` and `__eq__` have undefined
behavior when given multiple keys that have the same `lower()`.
6. The new method `lower_items` is like `iteritems`, but keys are
all lowercased.
7. CID raises `KeyError` for `__getitem__` as normal dicts do. The
old implementation returned
6. The `__repr__` now makes it obvious that it's not a normal dict.
See PR #1333 for the discussions that lead up to this implementation
Kenneth Reitz