It seems the wsgiref package was copied from Python 2.* without any
modifications. There are already 3 issues about that but they only
describe a part of the problem so I decided to start a new one. The
issues was:

http://bugs.python.org/issue3348
http://bugs.python.org/issue3795
http://bugs.python.org/issue4522

The attached patch fix the problem with the following changes:

• Fixed headers handling in wsgiref/simple_server.py;

• Fixed encoding problems. Now WSGI applications must return iterable
  with bytes but start_response() allow status and headers as bytes and as
  strings. "wsgi.input" file-like now use BytesIO instead of StringIO;

• Fixed tests;

• Updated documentation examples;

Phillip, do you have time to take a look at it? We really *must* fix
wsgiref in py3k...

FYI, instead of trying to do exhaustive type checking in _check_type(),
you can just rely on duck typing and catch the TypeError:

>>> str(b"a", "utf-8")
'a'
>>> str(bytearray(b"a"), "utf-8")
'a'
>>> str(memoryview(b"a"), "utf-8")
'a'
>>> str(1, "utf-8")
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: coercing to str: need string or buffer, int found

If you want to change to using bytes, you're going to have to take it to
the Web-SIG and hash out a revision to PEP-333, which at the moment
requires the use of strings, period.

This has nothing to do with the desirability of bytes vs. strings; I am
sure that if Python had bytes from day 1, bytes would've been the way to
go with it. But simply changing the reference library is not the way to
change the spec.

In the meantime, as far as I'm aware, there are two other patches
pending to address these issues, but I'm not in a position to assess
their readiness/correctness since I have yet to even download Py3K. In
principle, I approve their approaches, so if someone else can handle the
code review, those fixes could in principle be put in without changing
the PEP. But to put *this* patch in, the PEP would have to be changed.

If you want to change to using bytes, you're going to have to take it
to the Web-SIG and hash out a revision to PEP-333, which at the moment
requires the use of strings, period.

What was called str in 2.x has become the bytes object in py3k.
What was called unicode in 2.x has become str in py3k.
(roughly)

Given the meaning of the term "string" and its possible acceptions have
dramatically changed between 2.x and py3k, how does this patch violate
the PEP more than any other?

Actually, the PEP says:

    HTTP does not directly support Unicode, and neither does this
    interface. All encoding/decoding must be handled by the
    application; all strings passed to or from the server must be
    standard Python *byte strings*, not Unicode objects.
    [emphasis mine]

So, not accepting bytes in py3k is clearly a violation of the PEP!

Antoine Pitrou wrote:

FYI, instead of trying to do exhaustive type checking in _check_type(),
you can just rely on duck typing and catch the TypeError:

Good point! I'll update the patch soon.

Antoine Pitrou wrote:

> If you want to change to using bytes, you're going to have to take it
> to the Web-SIG and hash out a revision to PEP-333, which at the moment
> requires the use of strings, period.

What was called str in 2.x has become the bytes object in py3k.
What was called unicode in 2.x has become str in py3k.
(roughly)

Agreed, moreover it's time for Python 3.0.1 and we need to decide -
apply a patch or just remove wsgiref completely for now. Keep wsgiref
just as nonworking piece of code is the worse solution which can made
questionable all WSGI effort.

Given that old str has been replaced by bytes in Python 3 I think the
patch is a correct implementation of the PEP from the Python 3's point
of view. To avoid confusion note about the meaning of the term *string*
can be added to the PEP later.

New version of the patch:

• Now only Unicode strings are allowed as status and headers because
  allowing bytes leads to big changes in wsgiref.validate and
  wsgiref.handlers;

At 03:37 PM 12/22/2008 +0000, Antoine Pitrou wrote:

So, not accepting bytes in py3k is clearly a violation of the PEP!

On the contrary. Please read the two paragraphs *after* the one you quoted.

To be quite clear: this change requires discussion on the Web-SIG and an
appropriate revision of the PEP. Ideally, the patch should include the
necessary PEP revision.

The Web-SIG discussion regarding a switch to bytes should also take into
consideration the effects of running 2to3 on existing WSGI applications
and/or servers. Will their code be converted to use bytes, or Unicode?

The previous choice to use Unicode was based on source compatibility
across Python implementations, so this shouldn't be thrown out on the
basis of simple expediency.

OK, I've attached PEP-333 compatible fixes for wsgiref. I think there is
only one problem remains:

• wsgiref expects io.BytesIO as input and output streams because of
  http.server module. I didn't find any restrictions on data returned by
  read() method of the "wsgi.input" stream in the PEP. Maybe I've missed
  some details?

Please read the two paragraphs *after* the one you quoted.

I don't see anything forbidding bytes objects in those two paragraphs.

Le mardi 23 décembre 2008 à 11:15 +0000, Dmitry Vasiliev a écrit :

Dmitry Vasiliev <dima@hlabs.spb.ru> added the comment:

OK, I've attached PEP-333 compatible fixes for wsgiref.

I may be mistaken, but it seems that your patch forces iso-8859-1
encoding of http bodies.

Antoine Pitrou wrote:

Le mardi 23 décembre 2008 à 11:15 +0000, Dmitry Vasiliev a écrit :
> OK, I've attached PEP-333 compatible fixes for wsgiref.

I may be mistaken, but it seems that your patch forces iso-8859-1
encoding of http bodies.

No, just as PEP said str used as a container for binary data. For
example to return UTF-8 encoded data you can use the following code:

     def app(environ, start_response):
         ...
         return [data.encode("utf-8").decode("iso-8859-1")]

I don't like it but I guess it's strictly follow the PEP (actually I
didn't notice this sections before):

"""
On Python platforms where the str or StringType type is in fact
Unicode-based (e.g. Jython, IronPython, Python 3000, etc.), all
"strings" referred to in this specification must contain only code
points representable in ISO-8859-1 encoding (\u0000 through \u00FF,
inclusive). It is a fatal error for an application to supply strings
containing any other Unicode character or code point. Similarly, servers
and gateways must not supply strings to an application containing any
other Unicode characters.

Again, all strings referred to in this specification must be of type str
or StringType, and must not be of type unicode or UnicodeType. And, even
if a given platform allows for more than 8 bits per character in
str/StringType objects, only the lower 8 bits may be used, for any value
referred to in this specification as a "string".
"""

We definitely need to use bytes in the future but it requires PEP update
and some migration guide.

No, just as PEP said str used as a container for binary data.

This is clearly the wrong thing to do. The only (immutable) string-like
object adequate for binary data in py3k is bytes, not str.

I understand the desire to stick to the PEP, but the PEP was devised
even before the first py3k alphas, and it clearly wasn't written with
py3k in mind. For example the following sentence becomes nonsensical:

    Again, all strings referred to in this specification must be of
    type str or StringType, and must not be of type unicode or
    UnicodeType.

since "str" objects *are* of type UnicodeType in py3k (and the C API is
still named PyUnicode_*)...

When a legal text becomes nonsensical wrt. reality, one has to adapt his
interpretation of the text to reality, not adapt reality to match the
nonsense.

In other words, wsgiref should accept/expose HTTP bodies as bytes, not
str. Confusing binary data with iso-8859-1 text is the kind of mess py3k
was designed to avoid.

Antoine, you have three choices here:

1. Follow the PEP,
2. Take it to the Web-SIG and get the appropriate discussion,
   consensus, and PEP revision, or
3. Drop wsgiref from the current release of Py3K until Rename README to README.rst and enhance formatting #2 can be done.

Which would you prefer?

Please note that your arguments regarding what revision should take
place are irrelevant here; the correct forum for that discussion is
the Web-SIG. Personally, I think they are valid arguments; WSGI
simply did not have the benefit of having a sane (and standard!)
"bytes" type available, and were we writing the spec today, I would
absolutely have specified it in a bytes-oriented way, and treated
older Pythons as the special case.

However, we have to take into consideration how applications will be
*migrated* to Py3K. I am not an expert in this, nor do I personally
have huge volumes of code that will be migrated. Therefore, the
correct forum for discussing migration impact and how best to write
the spec is the Web-SIG.

Making the change to bytes is not something to be undertaken on a
whim: the spec must include how to handle inadvertent mixing of bytes
and unicode, in order to allow unambiguous error handling and
migration support. It will not be solved by the fiat of one
individual: certainly not by you or I. And it has absolutely nothing
to do with what is "right" in the technical sense, because it is not
a technical problem. A specification is a social construct, not a
technical one, so changing wsgiref by itself solves nothing.

And that's why those three choices are the only available options, so
far as I am aware.

Note that the page:

http://www.wsgi.org/wsgi/Amendments_1.0

contains clarifications for WSGI PEP in respect of Python 3.0. This list
was previously come up with on WEB-SIG list.

As another reference implementation for Python 3.0, you might look at
mod_wsgi (source code from subversion trunk), as that has been updated to
support Python 3.0 in line with those list of proposed clarifications for
WSGI PEP.

Attached new WSGI 1.0+ version of the patch.

Graham: thanks for pointing that out; I completely forgot we already
*had* the migration discussion on the Web-SIG! It just slipped my
mind because I didn't have any 3.0 work on the horizon.

Dmitry: A question about the new patch. Are bytearray and memoryview
objects in 3.0 really the same as bytestrings? It seems to me that
allowing mutable bytes objects is a mistake from a bug-finding
standpoint, even if it could be a win from a performance
standpoint. I think it might be better to be more restrictive to
start out, and then let people lobby for supporting other types,
rather than the other way around, where we'll never get to narrow the
list. Apart from that, the patch looks pretty good. Thank you!

Phillip J. Eby wrote:

Graham: thanks for pointing that out; I completely forgot we already
*had* the migration discussion on the Web-SIG! It just slipped my
mind because I didn't have any 3.0 work on the horizon.

Good to see we came to conclusion. Actually my first patch went in the
right direction. :-)

Dmitry: A question about the new patch. Are bytearray and memoryview
objects in 3.0 really the same as bytestrings? It seems to me that
allowing mutable bytes objects is a mistake from a bug-finding
standpoint, even if it could be a win from a performance
standpoint. I think it might be better to be more restrictive to
start out, and then let people lobby for supporting other types,
rather than the other way around, where we'll never get to narrow the
list. Apart from that, the patch looks pretty good. Thank you!

Actually I thought about functionality, not performance but I think
you're right and mutable bytes objects also can open doors for
unexpected side effects. I'll update the patch today.

Attached updated version of the patch.

If making changes in wsgireg.validate, may be worthwhile also fixing up one area where it isn't strictly correct
according to WSGI PEP.

As per discussion:

http://groups.google.com/group/python-web-sig/browse_frm/thread/b14b862ec4c620c0

the check for number of arguments supplied to wsgi.input.read() is wrong as it allows for an optional argument,
when argument is supposed to mandatory.

Graham Dumpleton wrote:

the check for number of arguments supplied to wsgi.input.read() is wrong as it allows for an optional argument,
when argument is supposed to mandatory.

I think it's a good idea. I'll update the patch soon.

Added check for wsgi.input.read() argument.

Hi,

The patch looks ok to me, although the tests against mutable byte-like
types are probably useless.

Antoine Pitrou wrote:

The patch looks ok to me, although the tests against mutable byte-like
types are probably useless.

Hmm, it's strange because such tests was removed two versions ago (per
discussion with Phillip). But at the time they really was needed.

Hmm, it's strange because such tests was removed two versions ago (per
discussion with Phillip). But at the time they really was needed.

Not a big deal anyway, let's keep them and we'll see.

Antoine Pitrou wrote:

> Hmm, it's strange because such tests was removed two versions ago (per
> discussion with Phillip). But at the time they really was needed.

Not a big deal anyway, let's keep them and we'll see.

I'm afraid I've lost your point here. Are you proposing to add back
tests for mutable bytes-like objects?

Why do you say they were removed? I see code like "assert
isinstance(value, bytes)".

Antoine Pitrou wrote:

Why do you say they were removed? I see code like "assert
isinstance(value, bytes)".

Support and tests for mutable "bytearray" and "memoryview" was removed.
All subclasses of "bytes" must be immutable so isinstance() should be OK
here.

Ok, sorry for the misunderstanding.

Philip, Graham, do you have any objections to the current patch?
Otherwise I think I'm gonna commit it soon.

One interesting thing of note that has occurred to me looking at the patch
is that although with Python <3.0 you technically could return a str as
iterable from application, ie., because iteration over str returns str for
each character, the same doesn't really apply to bytes in Python 3.0. This
is because iterating over bytes yields an int fdor each item.

Thus have odd situation where with Python 3.0, one could technically return
str as iterable, with rule that would apply would be that each str returned
would then be converted to bytes by way of latin-1 conversion, but for
bytes returned as iterable, should fail.

Not sure how this plays out in wsgiref server yet as haven't looked.
Anyway, make the validator code:

@@ -426,6 +436,6 @@
     # Technically a string is legal, which is why it's a really bad
     # idea, because it may cause the response to be returned
     # character-by-character
-    assert_(not isinstance(iterator, str),
+    assert_(not isinstance(iterator, (str, bytes)),
         "You should not return a string as your application iterator, "
         "instead return a single-item list containing that string.")

quite a good thing to have.

Graham Dumpleton wrote:

Thus have odd situation where with Python 3.0, one could technically return
str as iterable, with rule that would apply would be that each str returned
would then be converted to bytes by way of latin-1 conversion, but for
bytes returned as iterable, should fail.

Not sure how this plays out in wsgiref server yet as haven't looked.

If application return bytes instead of an iterable AssertionError will
be raised in handlers.BaseHandler.write().

It is committed now in py3k and the 3.0 maintenance branch. Thanks all
for your participation!

Reopening, the patch actually produces failures when run with "python
-bb", that is there are comparisons between str and bytes.

See the errors at the end of
http://www.python.org/dev/buildbot/3.x.stable/ppc%20Debian%20unstable%203.0/builds/26/step-test/0

People, does this patch look ok to you?

Antoine Pitrou wrote:

People, does this patch look ok to you?

Oh, didn't know about -bb.
The patch looks OK for me.

Le samedi 03 janvier 2009 à 20:24 +0000, Dmitry Vasiliev a écrit :

Dmitry Vasiliev <dima@hlabs.spb.ru> added the comment:

Antoine Pitrou wrote:
> People, does this patch look ok to you?

Oh, didn't know about -bb.

Well, it's meant to catch potential bugs. str and bytes always compare
unequal in py3k (i.e. "a" != b"a"), so it's good to use -bb when porting
some stuff from 2.x.

There's another problem in that buildbot failure with the environment
variable "NO_PROXY". We'll see if it's still there after the patch.

Antoine Pitrou wrote:

There's another problem in that buildbot failure with the environment
variable "NO_PROXY". We'll see if it's still there after the patch.

Strange error and it seems there is only part of the traceback. I've
already seen such "partially displayed" Python 3 traceback and error
actually can be in very different place.

Strange error and it seems there is only part of the traceback. I've
already seen such "partially displayed" Python 3 traceback and error
actually can be in very different place.

If you can reproduce such a problem, please open a bug.

Antoine Pitrou wrote:

> Strange error and it seems there is only part of the traceback. I've
> already seen such "partially displayed" Python 3 traceback and error
> actually can be in very different place.

If you can reproduce such a problem, please open a bug.

OK, I'll try to reproduce.

Attached patch for test_urllib, possible source of the "NO_PROXY" problem.

Nice catch! I've committed the two patches and we'll see whether it
makes the buildbots feel better.

I assume the buildbots were placated?