[Python-Dev] python-dev Summary for 2004-08-01 through 2004-08-15 [draft]

Brett C. bac at OCF.Berkeley.EDU
Thu Aug 19 09:02:07 CEST 2004


Well, after reading 1,289 emails for this summary I finally finished it. 
  As you will notice in the Summary Announcements I have my yearly stats 
on the list since this summary marks the end of year two for me doing 
this.  As I did last year, the position is available to anyone who wants 
it and can do a comparable job.

Also notice that in the Summary Announcements I am asking for people to 
vote on one of two ways I am going to cut back on the amount of time the 
Summaries take up of my time.  If you actually read the summaries then 
do let me know since votes from regular python-dev'ers will count for 
more than from the general Python community.

I don't plan to send this summary out any sooner than Sunday night so 
you have at least until then to send me corrections.  I am burned out 
enough at the moment that I doubt I will hit that date, though.

Also, I doubt I covered anything new in my decorators summary, but if I 
did let me know and I will make sure to plug it into the PEP or wiki if 
needed.

---------

=====================
Summary Announcements
=====================
Well, I have now been doing the summaries for two years.  As has become 
a yearly tradition, I am offering to pass on the writing of the 
summaries to someone else.  My only requirement is you do a comparable 
job.  You do learn *tons* about Python's internals when you have to 
research a topic to be able to summarize it.  I know it greatly 
accelerated my understanding of both Python as a language and of its 
implementation.

With that said, it is also time for stats on the list to scare away 
anyone considering taking over this job  =) .  According to my probably 
buggy script that I wrote last year, I have read 10,190 emails over the 
past year (this month has already been the busiest month with 1,289 
emails and it is only half over; could set the record for busiest month 
ever).
The people with over 300 emails posted over the year are:

9. Michael Hudson (303)
8. Martin v. Lowis (307)
7. Barry Warsaw (341)
6. Phillip J. Eby (341) # not a typo; same number as Barry
5. Greg Ewing (354)
4. Raymond Hettinger (372)
3. Skip Montanaro (399)
2. Tim Peters (629)
1. Guido van Rossum (1031)

These 9 people make up over 40% of all the emails from the past year.

Longest threads were:

1. decorate-sort-undecorate (694 emails)
2. Call for defense of @decorators (195 emails)
3. PEP 318: Decorators last before colon (181 emails)
4. redefining is (162 emails)
5. Christmas Wishlist (162 emails)

These stats along with the insane amount of email has showed me 
something; the Summaries have detracted from my participation on 
python-dev this past year.  I have a bigger mouth and more opinions than 
the number of emails I sent to the list show.  This means something has 
to change, and it isn't going to be my mouth.  The Summaries need to 
change in one of two ways.

Option one is that I focus in on certain areas of interest and skip 
other ones.  I have labeled all of the summaries below with what their 
type is right under the titles.  Please email me your top ares of 
interest.  I realize that this month may not be typical but pretty much 
all areas are covered at least once so at least there is a good taste 
for the different areas.  So, choose from:

1. improvement stuff
       Pretty much anything that is not a major change to the 
language/stdlib.  Usually something you find out from the Misc/NEWS or 
the "What's New" doc.
2. how python-dev works
       Stuff like how python-dev handles things such as the PEP process, 
coding style, etc.
3. nuggets of info
       Cool tricks and ideas that come up on the list that the greater 
world probably does not know.
4. minor language change
       Stuff that deals with the language changing, but is not 
significant; such as discussions of existing PEPs.

Language evolution stuff (such as decorators) will always be covered so 
you don't need to vote on that.  If people like this first option then I 
will make sure I cover the area with most votes and everything else is 
just considered icing.

Option two out of all of this is people just say, "summarize what you 
want, Brett."  Then I just cover what I find interesting and just don't 
worry about covering a specific area.  I obviously prefer this option 
but if people really care about a specific area I want to make sure to 
cover it.  What will most likely happen is I will still cover almost 
everything but the thoroughness will be lower.  I will go more off of 
memory for example.

But something will change.  Being on the sidelines for the decorators 
discussion because I dreaded having to start reading all of those emails 
in terms of summarizing is not acceptable (and no, I can't just ignore 
some of them since that is just not how I work).  Having to read 1,289 
emails for just the first two weeks of this month finally caused me to 
lose my sanity.

========
Summary
========
--------------------------------------------------------------------------------
Multiple interpreters at a time in a C program kind of broken
--------------------------------------------------------------------------------
3. nuggets of info

Philip Eby thought he might have a fix for a known limitation of running 
multiple interpeters at once (using PyInterpreter_New() ) and having 
imports not being clearly separated between interpreters.  But Martin v. 
Löwis popped Philip's bubble somewhat by saying that he and some other 
people viewed multiple interpreter support as inherently broken.

Contributing threads:
   - `Safe to change a thread's interpreter? 
<http://mail.python.org/pipermail/python-dev/2004-August/046593.html>`__


--------------------------------------------
Another sucker becomes an official developer
--------------------------------------------
2. how python-dev works

Sean Reifschneider took the bait to become another one of Guido's minions.

Sean has been maintaining the RPM spec files for quite a while.

Contributing threads:
   - `New developer 
<http://mail.python.org/pipermail/python-dev/2004-August/046595.html>`__


----------------------------------------
Discovering leaks with a regrtest wrench
----------------------------------------
1. minor language change

Michael Hudson, who found and fixed a bunch of refcount leaks shortly 
after 2.3 was released, used his magical regrtest patch (which has 
subsequently been checked in) to hunt down leaks for 2.4 .  A bunch of 
tests were cleaned up to make them not raise false-positives along with 
fixing some true leaks.

Contributing threads:
   - `refleak hunting season 
<http://mail.python.org/pipermail/python-dev/2004-August/046611.html>`__


---------------------------------
Another Bug Day has come and gone
---------------------------------
2. how python-dev works

The Bug Day held on August 7 led to 19 bugs and 12 patches being closed. 
  You can see the results at 
http://www.python.org/cgi-bin/moinmoin/PythonBugDayStatus .

Feel guilty for not being able to make it?  Still want to help?  Go to 
that page to see bugs that could use some squashing that shouldn't be 
too difficult to deal with.

Contributing threads:
   - `Bug day coming up this Saturday 
<http://mail.python.org/pipermail/python-dev/2004-August/046645.html>`__


--------------------------------------------------
How to shut a compiler up about an unused variable
--------------------------------------------------
3. nuggets of info

Tim Peters had come up with some optimizations for list.pop() that 
avoided unneeded test+branches.  The problem is that it led to a 
variable possibly being unused.  C compilers tend to complain about that 
and so an initial solution was used.  Unfortunately gcc complained about 
it, and so yours truly came up with one.  But that solution was labeled 
as "perverse" (and it was; total hack solution), so another solution was 
found thanks to Sjoerd Mullender by just taking the variable and casting 
it to void.

Contributing threads:
   - `RE: [Python-checkins] python/dist/src/Objects listobject.c, ... 
<http://mail.python.org/pipermail/python-dev/2004-August/046678.html>`__


------------------------------------------------------
Variable sequence unpacking assignment shot down again
------------------------------------------------------
language evolution

David Cole suggested adding the ability to have sequence unpacking 
assignment just like *args for parameters; ``a, b, *c = (0, 1, 2, 3, 4, 
5)  # a == 0, b == 2, c == (3, 4, 5)``.  This idea got a -1 from Guido 
as not really needed.  That reaction makes Guido consistent; the idea 
was originally shot down back in Nov. 2002; 
http://www.python.org/dev/summary/2002-11-16_2002-11-30.html#half-baked-proposal-and-in-assignments 
.  If you really like that idea the previous summary contains a function 
that helps you do this in a way.

Contributing threads:
   - `Tuple/list assignment question 
<http://mail.python.org/pipermail/python-dev/2004-August/046684.html>`__


-------------------------------------------------------------------------------------
Changing the Big-O complexity for something in the language is now a 
language feature
-------------------------------------------------------------------------------------
language evolution

Armin Rigo came up with a way to have string concatenation in a loop 
(think ``for thing in iter_of_strings: concat_str += thing``) not be a 
quadratic algorithm thanks to some trickery for ``a = a + b`` and ``a += 
b`` conditions for strings.  The hope was to remove the (commonly 
considered) wart of having ``"".join(iter_of_strings)`` be the suggested 
way to concatenate a bunch of strings.

But Guido didn't like the patch.  His reasoning was that changing 
something that led to a change in the Big-O complexity of certain 
algorithms would inherently hurt other implementations of Python when 
people would start to code specifically for that performance gain.  For 
instance, having Jython be able to pull this trick off is, I believe, 
near impossible.  So, in order to make sure changes like this are 
considered before applying them, Guido instated a new rule that 
"implementation features that affect not just the running speed but the 
O() rate for certain algorithms should be considered language features, 
because any implementation will be required to implement them in order 
to ensure code portability" between implementations of Python.

In the end, though, this went in with a warning that the speed 
performance is not portable.  It is not to be used in the stdlib ever.

Contributing threads:
   - `Optimized string concatenation 
<http://mail.python.org/pipermail/python-dev/2004-August/046686.html>`__
   - `PEP 0008 confusion - here it is, but don't use it? 
<http://mail.python.org/pipermail/python-dev/2004-August/047194.html>`__


-------------------------------------------------------------------------------
Bet you didn't think about string interning and how that affects .pyc, 
did you?
-------------------------------------------------------------------------------
1. minor language change

Luckily Martin v. Löwis did.  A patch was originally applied to not 
intern strings representing filenames.  Problem is that every code 
object stores that string, so it increases .pyc files since it has to 
store multiple copies of that string instead of just one.

Contributing threads:
   - `Re: [Python-checkins] python/dist/src/Python compile.c ... 
<http://mail.python.org/pipermail/python-dev/2004-August/046703.html>`__


------------------------------------------------------------------------------
`PEP 324`_ (process - New POSIX process module) != process.py from 
ActiveState
------------------------------------------------------------------------------
1. minor language change

Guido asked if the APIs between the module proposed in `PEP 324`_ and 
the process.py module by Trent Mick of ActiveState were compatible. 
Turns out they are not.  Then the discussion went into select(), broken 
pipes, and other stuff not relevant to the OP.  =)

.. _PEP 324: http://www.python.org/peps/pep-0324.html

Contributing threads:
   - `PEP 324 (process module) 
<http://mail.python.org/pipermail/python-dev/2004-August/046707.html>`__


--------------------------------------------------------------------------------------------------------------------------------------------------
Getting it so string literals accept those nutty non-ASCII characters 
more easily (and getting files to be more upfront about their Unicode-ness)
--------------------------------------------------------------------------------------------------------------------------------------------------
language evolution

François Pinard asked what people thought about two things related to 
Unicode.  First, he thought having a __coding__ attribute for files that 
contained the encoding of the text would be nice.  Martin v. Löwis said 
it would be possible.  MA Lemburg added his vote of support.  It has not 
been implemented to my knowledge yet, though.

The second idea was to have a directive of some sort on a per-module 
basis so that all string literals could be considered in something other 
than ASCII.  Once again Martin said it was doable with MA saying he 
liked the idea.  But it was pointed out that you might as well just use 
the 'u' in front of strings now.

This led to a slight discussion on good i18n practices.  Both Martin and 
MA seemed to suggest that if you are going to be doing most of your work 
in a single encoding then just use that and declare everything Unicode. 
  ANd if you are going to support a lot of different languages, use 
gettext and such.  Martin also stated it is best to get text into 
Unicode ASAP and then only convert to the final encoding at the last moment.

Contributing threads:
   - `Python in Unicode context 
<http://mail.python.org/pipermail/python-dev/2004-August/046727.html>`__


----------------------------------------------------------------------
An exception is an exception, unless it doesn't inherit from Exception
----------------------------------------------------------------------
language evolution

Coming up in a discussion on Unicode of all things, a discussion on 
moving exceptions over to new-style classes came up.  Guido pointed out 
that moving over to new-style classes would seem to suddenly require 
that anything passed to 'raise' be a new-style class and that goes 
against the current acceptance.

But in comes Paul Prescod with his argument that exceptions are 
inherently organized based on inheritance and thus requiring things 
being passed to 'raise' subclass Exception somewhere is not that big of 
a thing.  Guido really liked that point.  So if it goes this way and you 
don't like it blame Paul (who is a nice guy, so go easy on him  =).

And then along come Holger Krekel bringing up the point that using 
exceptions to do non-standard flow control is handy.  But the question 
was asked as to why he couldn't still just subclass Exception?  Was it 
that big of a deal?  He admitted it wasn't and said when the name 
Raisable was suggested that would make it easier.

Michael Hudson then came in and wrote up a hacked-up patch to turn 
exceptions into new-style classes.  Most stuff seemed to still work.

Python 3 was already going to switch to new-style classes for exceptions 
and string exceptions have already been deprecated.  Now add to the mix 
the possible requirement that anything passed to 'raise' require a 
common base class.

Contributing threads (note that the thread starts part way in a thread 
on Unicode and end before the end of the full thread):
   - `Python in Unicode context 
<http://mail.python.org/pipermail/python-dev/2004-August/046757.html>`__


-------------------------
Python 2.4a2 out the door
-------------------------
language evolution

Python 2.4a2 has been released.  As usual, please download it and test 
it with your code along with the regression test suite.

Contributing threads:
   - `trunk frozen for 2.4a2 
<http://mail.python.org/pipermail/python-dev/2004-August/046822.html>`__
   - `RELEASED Python 2.4, alpha 2 
<http://mail.python.org/pipermail/python-dev/2004-August/046864.html>`__


-------------------------------------------------------
Amendment to how to compile with the free .NET compiler
-------------------------------------------------------
3. nuggets of info

Nick Coghlan expanded upon my summary on how to compile with the free 
.NET compiler under Windows at 
http://www.python.org/dev/summary/2004-07-16_2004-07-31.html#how-to-get-python-to-compile-with-microsoft-s-free-compiler-that-should-just-come-with-the-os-standard 
.  See his email at 
http://mail.python.org/pipermail/python-dev/2004-August/047215.html on 
the extra steps.

Or just move to another OS.

Contributing threads:
   - `python-dev Summary for 2004-07-16 through 2004-07-31 [draft] 
<http://mail.python.org/pipermail/python-dev/2004-August/047215.html>`__


--------------------------------------
Darned Solaris, g++, and _XOPEN_SOURCE
--------------------------------------
1. improvement stuff

Skip Montanaro discovered that building under Solaris with g++ raised a 
warning about redefining _XOPEN_SOURCE.  pyconfig.h defines it, but 
apparently so does g++ in order for Solaris' toolchain to expose some 
code that is only available if it is defined.  Martin v. Löwis came up 
with a couple of suggestions on how to handle this.  Skip ended up going 
with the idea of setting _XOPEN_SOURCE to the same value as it is 
defined by g++.

Contributing threads:
   - `use of #pragma GCC system_header to suppress _XOPEN_SOURCE 
warnings 
<http://mail.python.org/pipermail/python-dev/2004-August/046955.html>`__


-------------------------------------------------------------------------------------
pre-PEP on a function for re that "captures matches for groups that 
match repeatedly"
-------------------------------------------------------------------------------------
1. minor language change

Mike Coleman presented a pre-PEP on adding a function to re that would 
create a tree (in the form of a list) containing all group matches in a 
string.  It got a little support, but the discussion quickly moved over 
to dreamings of a full-on scanner or parser package for the stdlib.

But if the idea of the PEP works for you then speak up on c.l.py .

Contributing threads:
   - `pre-PEP: Complete, Structured Regular Expression Group Matching 
<http://mail.python.org/pipermail/python-dev/2004-August/047042.html>`__


---------------------------------------
Making ourselves follow the PEP process
---------------------------------------
2. how python-dev works

It was noticed early on that the PEP process had broken down for `PEP 
318`__ (decorators).  What should happen is a PEP gets written (and the 
author becomes its champion), there is a public discussion, the PEP is 
updated, that's repeated until it is deemed done, gets get BDFL 
pronouncement, if Guido okays it the code goes in.  Unfortunately the 
part about  updating the PEP didn't really happen.

This led to two major things being stated and generally agreed upon. 
One is that PEPs should not be checked in if the PEP has not been 
updated.  While a minor nit is not a reason to hold back code, not 
updating after a major public debate is not acceptable.

This directly segued into the other point of a PEP needs to have a 
champion, period.  Since the developers on python-dev do not have the 
time to keep PEPs updated it is up to the PEP champion to make sure it 
is kept current.  If it isn't it is take to mean the champion no longer 
cares, which means python-dev no longer cares, which means the PEP gets 
rejected outright.  This will also make sure that there is a focus to 
the design of what the PEP wants and does not lead to a 
design-by-committee problem.

.. _PEP 318: http://www.python.org/peps/pep-0318.html

Contributing threads:
   - `PEP 318, and the PEP process 
<http://mail.python.org/pipermail/python-dev/2004-August/047020.html>`__


----------------------------------------------------------------
How to tell Windows from Linux without lifting up anyone's skirt
----------------------------------------------------------------
3. nuggets of info

The question came up on what the best way was to tell what platform you 
are running on.  The answer came down to why you were cared.  If it was 
whether or not you had a specific functionality (or lack thereof), then 
just test for the functionality.  If you had other needs, though, using 
sys.platform seemed to be the most reliable way (at least for Windows 
since all of them use 'win32').

Contributing threads:
   - `Asyncore .set_reuse_addr() on Windows 
<http://mail.python.org/pipermail/python-dev/2004-August/047269.html>`__


-----------------------------------------
func_name can now have an identity crisis
-----------------------------------------
1. improvement stuff

Thanks to the rampant decorator discussion, Skip Montanaro came up with 
the idea of wanting func_name to be writable.  This would allow 
decorators to wrap a function with another function and yet reset 
func_name to its original value, thus not having the wrapper function 
become the prevailing name.

Guido liked and Michael Hudson subsequently wrote up `patch #1004703`_.

.. _patch #1004703: http://www.python.org/sf/1004703

Contributing threads:
   - `PEP 318 - writing to func_name attribute 
<http://mail.python.org/pipermail/python-dev/2004-August/047097.html>`__


-----------------------------------------
statistics.py ain't goin' into the stdlib
-----------------------------------------
1. improvement stuff

Raymond Hettinger (at the last Python Bug Day) made the statement that 
he didn't think 'statistic's should be taken out of the sandbox in CVS 
and put into the stdlib.  He felt that most of the important algorithms 
were available elsewhere (such as nsmallest and nlargest in the heapq 
module) and the remaining functions that were not found anywhere were 
not that critical.

He did say, though, he would like to do a calculator module where those 
remaining functions could go.

Contributing threads:
   - `status of statistics.py? 
<http://mail.python.org/pipermail/python-dev/2004-August/047129.html>`__


------------------------------------
Making re.split care about emptiness
------------------------------------
1. improvement stuff

Mike Coleman wrote up a patch for re.split() that AM Kuchling presented 
to the list.  It adds an argument to the function to allow an empty 
string to be put into the resulting list when a match occurs, even if 
the match consumes no characters.

No decision on the exact semantics of the function, how to handle 
turning on the new functionality (some say an extra argument, some say 
adding another bit flag like traditional re arguments), or even if it 
will be accepted.

Contributing threads:
   - `re.split on empty patterns 
<http://mail.python.org/pipermail/python-dev/2004-August/047272.html>`__


----------------------------------------
Pickler's 'bin' argument on its last leg
----------------------------------------
1. improvement stuff

As per `PEP 307`_ (Extensions to the pickle protocol), the 'bin' 
argument is being upgraded from PendingDeprecationWarning to 
DeprecationWarning; so it's gone plaid and the only way to stop it is 
with some emergency change.

Contributing threads:
   - `Pickler()'s bin argument 
<http://mail.python.org/pipermail/python-dev/2004-August/047274.html>`__


---------------------------------
Some modules are getting the boot
---------------------------------
1. improvement stuff

TERMIOS, mpz, statcache, xreadlines and rotor are all being removed. 
mimify, MimeWriter, and whrandom will raise a DeprecationWarning. 
rfc822 and mimetools will not be raising DeprecationWarning as specified 
by `PEP 4`_ since some code in the stdlib still uses it.

.. _PEP 4: http://www.python.org/peps/pep-0004.html

Contributing threads:
   - `Removing stuff from 2.4 
<http://mail.python.org/pipermail/python-dev/2004-August/047308.html>`__


-------------------------------------------------
Should decimal.Context.copy() be deep or shallow?
-------------------------------------------------
1. improvement stuff

Raymond Hettinger brought up the question of whether 
decimal.Context.copy() should be a deep or shallow copy.  While 
tradition dictates it be shallow based on name, it seems like the 
functionality should be deep.  No one wants context to be shared between 
number naturally since numbers tend to be viewed as unique things.

Staring in 2.4a3 it will be deep unless people come up with reasons to 
switch it to shallow.

Contributing threads:
   - `decimal.Context.copy() shallow or deep? 
<http://mail.python.org/pipermail/python-dev/2004-August/047416.html>`__


----------------------------------------------------------
The dark corners of Python allow you to skip return values
----------------------------------------------------------
3. nuggets of info

Christian Tismer discovered that you can actually stop a return 
statement from returning if you encase it in a 'finally' block and tweak 
it slightly (see the OP to get what I mean by this).  Turns out this is 
a reliable feature of Python if you really want to use it.

Contributing threads:
   - `capturing RETURN_VALUE 
<http://mail.python.org/pipermail/python-dev/2004-August/047360.html>`__


--------------------------------------
Is an int/long combined type worth it?
--------------------------------------
language evolution

Dmitry Vasiliev pointed out that `PEP 237`_ (Unifying Long Integers and 
Integers) mentioned that a a new type named 'integer' might be 
introduced that subclassed both int and long.  The discussion waivered 
between whether it was at all needed, and if it was if it should be a 
true type or just a tuple containing both types for use with isinstance() .

No conclusion was reached in the end.

.. _PEP 237: http://www.python.org/peps/pep-0237.html

Contributing threads:
   - `Unifying Long Integers and Integers: baseint 
<http://mail.python.org/pipermail/python-dev/2004-August/047404.html>`__


--------------------------------------
Should byte arrays be a built-in type?
--------------------------------------
language evolution

Through the discussion bout having a common type combining int and long 
a discussion on whether a byte array type should be introduced.  The 
initial suggestion was for it to be practically synonymous with str 
since str itself stores everything as an array of 8-bit values.  The use 
cases would be for streams and such that just want a stream of bytes 
with no care for any encoding.

Syntactically, having a 'b' and 'B' cookie before literals was 
suggested.  The issue with this, though, is that byte arrays should be 
mutable, which would make literals that can be mutated, and that is 
frowned upon by Guido.

Semantically, aliasing bytes to the str type was suggested.  That was 
not loved since that could create confusion.  Returning an object from 
'array' was suggested and seemed okay.

In general it seemed this could be useful and could go in 2.5, but 
nothing for 2.4 .

Contributing threads:
   - `Unifying Long Integers and Integers: baseint 
<http://mail.python.org/pipermail/python-dev/2004-August/047665.html>`__


-----------------------------
Thar the Windows stack blows!
-----------------------------
1. minor language change

Running the new test_compiler test (if you run it with ``-uall`` for 
regrtest it will recompile the entire stdlib) was leading to an odd 
failure: the "process on Windows "just vanishes" without a trace, and 
without an error message of any kind, but with an exit code of 128". 
After a lot of work put in by a bunch of people (led by Tim Peters) the 
problem was tracked down to a blown C stack.

Turned out that the stack-checking code was not being called frequently 
enough to pick up the error.  The problem with this was that it was 
leading to odd errors that should have been MemoryError but were 
manifesting themselves as KeyError.  This was because PyDict_GetItem() 
was failing and return NULL which is the same as signaling the key 
doesn't exist in the dict.

Trick was trying to come up with a good way to deal with this.  Adding 
more calls would be very expensive (reliable way of catching it was 
sticking a check in pymalloc code) and so that was ruled out. 
PyDict_GetItem() couldn't change its return value since that would break 
more code than could be imagined.  So, in the end, the stack was 
increased to 2 MB on Windows.

Contributing threads:
   - `Another test_compiler mystery 
<http://mail.python.org/pipermail/python-dev/2004-August/047468.html>`__


------------------------------------------
Someone else falls for the dangling carrot
------------------------------------------
2. how python-dev works

Johannes Gijsbers now has checkin rights.  May he close many bugs.

Contributing threads:
   - `New Developer 
<http://mail.python.org/pipermail/python-dev/2004-August/047815.html>`__


----------------------------------------------------------
Lying about being in __builtin__ is not always a bad thing
----------------------------------------------------------
1. improvement stuff

James Knight noticed that some objects (mostly C types such as iterators 
for the built-in types and such) claim in their __module__ attribute 
that they come from __builtin__ which is not technically correct since 
you can't access them from there.  The desire to fix this came from 
allowing for proper introspection.  The leading idea is to put these 
types in __builtin__ properly so that they are no longer lying about 
where they come from.

Contributing threads:
   - `Classes that claim to be defined in __builtin__ but aren't 
<http://mail.python.org/pipermail/python-dev/2004-August/047477.html>`__


---------------------------------------------------------
Bringing order to the order of application for decorators
---------------------------------------------------------
1. minor language change

Turns out that the order of application for decorators was implemented 
in the reverse order of what it was supposed to be in 2.4a2 .  Luckily 
James Knight spotted this and let the list know.  It has been fixed, 
though, in CVS and now follows `PEP 318`_ to apply in bottom-up order::

   @first
   @second
   def blah(): pass

is equivalent to::

   blah = first(second(blah))

The arguments for this ordering beyond it just making more sense to 
Guido and others is that people will typically put decorators such as 
staticmethod and classmethod first.  Going with the original order would 
have led to errors in most situations if people were not expecting to be 
receiving a class or static method to be passed to the other decorators.

There was a short discussion on the order of instantiation for the 
decorators, but in the end the order chosen was the order listed; first 
would be first() instantiated followed by second().

Contributing threads:
   - `Decorator order implemented backwards? 
<http://mail.python.org/pipermail/python-dev/2004-August/047512.html>`__


-----------------------------------------------------------------------------------
PEP 292 (Simpler String Substitutions) getting gussied up for the Python 
2.4a3 ball
-----------------------------------------------------------------------------------
1. minor language change

`PEP 292`_'s implementation got fixed up.  The names of the class names 
were changed to Template and SafeTemplate, the whole process was made 
lazy, and just cleaned up in general (solution is small, simple, and 
rather cool; all subclassable and works with Unicode to boot).

But then questions over the merits of $-based string interpolation 
popped up.  People wondered if going from ``%(blah)s`` to ``${blah}`` 
was worth it.  The answer is "yes".  And that is just worst case; when 
you just want ``$blah`` you get an even bigger win.

The other question was over whether the string module should become a 
package.  The original idea was to stick the original string module in 
as a submodule of the package and the Template code in another module. 
This would allow easy removal of the 'string' module code that has been 
deprecated for eons.  Barry Warsaw (author of the PEP) asked Guido to 
make the call on this, but he hasn't had the time yet to get to this.

.. _PEP 292: http://www.python.org./peps/pep-0292.html

Contributing threads:
   - `PEP 292 
<http://mail.python.org/pipermail/python-dev/2004-August/047580.html>`__


--------------------------------
Multi-line imports get clarified
--------------------------------
1. minor language change

Multi-line imports will only accept parentheses around what is being 
explicitly imported and only if there is more than one item being 
imported.  Trailing commas are also to be accepted.

Contributing threads:
   - `Multi-line import implementation (first part) 
<http://mail.python.org/pipermail/python-dev/2004-August/047393.html>`__
   - `Multi-line import implementation (second part) 
<http://mail.python.org/pipermail/python-dev/2004-August/047614.html>`__
   - `Naming nit 
<http://mail.python.org/pipermail/python-dev/2004-August/047641.html>`__


------------------------------------------------
For those of you who need Python to run on Win64
------------------------------------------------
3. nuggets of info

Nick Bastin asked if anyone has gotten Python 2.3 to work under Win64. 
Martin v. Löwis said "yes" for Python 2.4, but not for 2.3 .  He 
suggested to Nick that he run vsextcomp_ to generate the targets on the 
2.4 VC 7 build files and then move that over to 2.3 .

.. _vsextcomp: http://www.sf.net/projects/vsextcomp

Contributing threads:
   - `2.3.4 on Win64? 
<http://mail.python.org/pipermail/python-dev/2004-August/047631.html>`__


----------------------------------------------
Sometimes concealing the truth is a good thing
----------------------------------------------
1. improvement stuff

Nick Coghlan discovered that some of the function in the 'commands' 
module did actually work under Windows and he wanted to make sure it was 
okay to fix another module to work under Windows and to document the 
fact.  But the whole idea was shot down by both Tim Peters and Guido in 
order to keep the module simple.  Keeping the whole thing UNIX-only is 
much easier than having an API that is only partly compatible with 
Windows (and with only certain versions of Windows at that).  Guido also 
said that the module would not even be accepted today since it doesn't 
add enough functionality.

Contributing threads:
   - `'commands' module on Win32 
<http://mail.python.org/pipermail/python-dev/2004-August/047774.html>`__


----------------------------------------------
atexit module, good; sys.exitfunc, not so good
----------------------------------------------
1. improvement stuff

Raymond Hettinger pointed out that the docs for the atexit module state 
that sys.exitfunc was to be deprecated.  Well, atexit has been in the 
stdlib since 2.0 so the deprecation is long overdue.  Looks like it will 
get its deprecation finally.

Contributing threads:
   - `Deprecate sys.exitfunc? 
<http://mail.python.org/pipermail/python-dev/2004-August/047842.html>`__


----------------------------------------------------------------------------------------------
My Personal Hell: decorators and the people who suggest new syntax for 
them...  next, on Oprah
----------------------------------------------------------------------------------------------
language evolution

What led to a record-setting flurry of emails to python-dev was set up 
when Guido gave the green light to checking in code to implement 
decorators using the '@' syntax (now known at the pie syntax thanks to 
its checkin coming very shortly after the OSCON Pie-thon competition and 
'@' sort of looking like a pie).  People didn't like it.  People were 
screaming that there had to be a better syntax than just settling for 
the least offensive one.  Others started to question whether decorators 
were really needed.  Others wanted to extend them even more and what 
their role truly was.  Either way this was all coming a little late.

But then Guido decided to make my life difficult by saying that if the 
community could come up with an agreed-upon alternative syntax to 
propose to him he would consider ripping out the '@' syntax; decorators 
have always been experimental and '@' was checked in so people had 
*something* to play with.  This meant everyone and their mother started 
to propose both new and old syntaxes for decorators.  This led to a 
record amount of email on python-dev (at least compared to what we have 
archives for; back to April 1999).

In order of my own personal sanity I am not going to cover any specific 
syntax.  There are just too many and I actually like the '@' syntax 
(just *try* using it; it grows on you quickly).  Instead this summary 
will cover any important design considerations for decorators (it should 
all be in the PEP but so much happened there is no way it all got in 
there) along with any other important considerations about them.  I will 
discuss anything specific to the '@' syntax since it is the currently 
winning syntax.

With that said, `PEP 318`_ and the PythonDecorators_ wiki page are the 
de-facto place for info on this whole situation.  I am probably going to 
be more biased than normal in this summary just out of time restraints 
to get this done by not over-analyzing people's suggestions and just 
assuming what I summarize here is not covered elsewhere; checking if it 
is would take too long.

First off, going with a syntax just so you could add backward-compatible 
support was not considered worth it.  It's a new feature so there is no 
need to hobble it at the get-go just so people who don't want to upgrade 
can still use it.  Plus the implementations to make this even possible 
were playing some tricks with the interpreter and not considered a good 
thing.  Besides you can, always will be able to, still use the way to do 
it in 2.2 and beyond.

Another design point that needed to be taken into consideration was 
ambiguity.  There could not be a question about what is a decorator and 
what isn't (list-before-def is an example of a syntax that can be 
ambiguous).

Syntactic support that allowed for possible future compiler 
optimizations was also considered important.  The on-going hope is to 
eventually get a JIT compiler for Python and if the decorators are not 
placed before the def somehow you make it much harder to optimize since 
you learn about the decorators after the fact of the function starting 
to be defined.

An argument against '@' was that people thought it got rather unwieldy 
quickly (and I am sure the new decision that each decorator must be on 
its own line is not making these objectors happy).  But one thing that 
must be remembered is that chances are most functions will not need more 
than one decorator if the even need one.  Looking through the stdlib 
sans the test suite you will notice there are not a lot of uses for any 
of the built-in decorators.  Yes, usage will probably grow with 
syntactic support for decorators, but not exponentially in terms of 
applying multiple decorators to a single function.

Allowing decorators to take arguments is a requirement.  Not being able 
to pass arguments to a decorator at instantiation time would severely 
cripple their usefulness.

Being prefix instead of postfix was considered important.  Knowing about 
what is coming up was something Guido thought is better than finding out 
about some special tweak to the use of a function based on a decorator 
listed after the definition line.  And in the method body had been 
completely ruled out.

In terms of implementation details, decorators can only be functions 
directly (and not lambda).  This means no ``@a or b`` tricks and such. 
It also means ``@foo().bar()`` is not allowed.  This is all gut feeling 
from Guido for why the restriction should be there.  It could change in 
the future since it is easier to loosen restrictions that make something 
more strict.

A problem with '@' in terms of portability is that Leo_ and IPython_ 
both use '@' for special syntax.  The authors of both tools seem willing 
to change their tools (not necessarily overly joyous, though  =) .

Guido wanted a syntax "that ... [is] easy to remember once explained". 
It did not need to be miraculously obvious what it did just by looking 
at the first time.

Some people didn't like '@' because they keep wanting to say "at" when 
they saw it.  To that, I give you Tim Peters' fix for this: "Pick the 
keyword you'd like.  Then whenever you see "@", pronounce
that word instead <wink>".

A single decorator that applied to multiple code blocks is completely 
out.  Guido gave a couple of reasons why, but one that he had mentioned 
multiple times is indenting the code block again just for the decorator 
is out.

Some people suggested that things be kept simple by just throwing out 
the entire idea of syntactic support.  Could happen if everyone agreed 
to that, but I wouldn't count on it.

A supposed knock against '@' was that it wasn't Pythonic.  Well, it is 
Pythonic if Guido says it is so not really a valid argument.


OK, tirade time.  If you don't want to listen to me get up on my 
soapbox, then just skip the rest of this...

"In the old days, Guido would Pronounce, and we'd all bite our tongues 
(although not necessarily each his own).  The less time Guido can make 
for Python, the more important becomes graceful capitulation."  Tim said 
this and it makes me wish for the old days.  People had *months* to 
comment on decorators and no one spoke up until something went into the 
language.  Procrastination is not a virtue when it comes to major 
language evolution discussions.  What was worse was when the emails 
started repeating themselves (which was pretty much from the get-go when 
this exploded).  Seemed like people decided to start talking without 
doing some research.  Granted the PEP was outdated and the wiki page was 
not up yet, but this stuff was covered in the Summaries before and you 
could have just Googled for the previous threads.

Personally, if I was Guido, I would have said that the community had 
their chance to speak up and they just didn't take it.  But Guido is a 
nicer guy than that so people are getting a second chance with this. 
Personally this came off a great case of the tyranny of the majority in 
my eyes.  There is a reason why Python is a dictatorship.

At this point people should be hashing out which syntax alternative they 
want to present to Guido on comp.lang.python_.  No more talking on 
python-dev, no more syntax proposals.  The community should set a goal 
date (Sept. 1 seems good) and just choose a bloody alternative.  Then 
when Guido makes his choice people accept it or just go to another 
language.  No one had better voice their disappoint once Guido chooses 
his syntax or I will personally come beat you over with a stick for 
being a whiner.

OK, that's out of my system.  I feel better now.

.. _PEP 318: http://www.python.org./peps/pep-0318.html
.. _PythonDecorators: 
http://www.python.org/cgi-bin/moinmoin/PythonDecorators
.. _Leo: http://leo.sourceforge.net/
.. _IPython: http://ipython.scipy.org/

Contributing threads (sans emails that were warnocked):
   - `2.4a2, and @decorators 
<http://mail.python.org/pipermail/python-dev/2004-August/046599.html>`__
   - `Plea for simpler decorator syntax,	in addition to pie-shaped 
syntax 
<http://mail.python.org/pipermail/python-dev/2004-August/046798.html>`__
   - `Call for defense of @decorators 
<http://mail.python.org/pipermail/python-dev/2004-August/046878.html>`__
   - `@decorators, the PEP and the "options" out there? 
<http://mail.python.org/pipermail/python-dev/2004-August/046905.html>`__
   - `About pep 318--Intro 
<http://mail.python.org/pipermail/python-dev/2004-August/046918.html>`__
   - `Questions about '@' in pep 318 
<http://mail.python.org/pipermail/python-dev/2004-August/046934.html>`__
   - `A usability argument for list-after-def 
<http://mail.python.org/pipermail/python-dev/2004-August/046935.html>`__
   - `The impact of '@' on Leo 
<http://mail.python.org/pipermail/python-dev/2004-August/046937.html>`__
   - `Similar syntax 
<http://mail.python.org/pipermail/python-dev/2004-August/046969.html>`__
   - `def fn (args) [dec,dec]: 
<http://mail.python.org/pipermail/python-dev/2004-August/047112.html>`__
   - `pep 318, Decorators for Functions, Methods and Classes 
<http://mail.python.org/pipermail/python-dev/2004-August/047132.html>`__
   - `Density of pie-decorator syntax 
<http://mail.python.org/pipermail/python-dev/2004-August/047154.html>`__
   - `elements of decorator syntax suggestions 
<http://mail.python.org/pipermail/python-dev/2004-August/047185.html>`__
   - `318: Annotation background 
<http://mail.python.org/pipermail/python-dev/2004-August/047218.html>`__
   - `IPython, @, and option E from the wiki 
<http://mail.python.org/pipermail/python-dev/2004-August/047223.html>`__
   - `Decorators: vertical bar syntax 
<http://mail.python.org/pipermail/python-dev/2004-August/047231.html>`__
   - `Suggesting '.' decorators (PEP318) 
<http://mail.python.org/pipermail/python-dev/2004-August/047234.html>`__
   - `Attribute strings? 
<http://mail.python.org/pipermail/python-dev/2004-August/047362.html>`__
   - `request: add keywords for static and class methods only 
<http://mail.python.org/pipermail/python-dev/2004-August/047498.html>`__
   - `@decorator syntax is sugar, but for what exactly? 
<http://mail.python.org/pipermail/python-dev/2004-August/047500.html>`__
   - `Semantics of decorators? 
<http://mail.python.org/pipermail/python-dev/2004-August/047623.html>`__
   - `A decorator syntax not yet mentioned (I think!) 
<http://mail.python.org/pipermail/python-dev/2004-August/047625.html>`__
   - `Another approach to decorators. 
<http://mail.python.org/pipermail/python-dev/2004-August/047647.html>`__
   - `Decorators after 'def' should be reconsidered 
<http://mail.python.org/pipermail/python-dev/2004-August/047637.html>`__
   - `def ... decorate 
<http://mail.python.org/pipermail/python-dev/2004-August/047781.html>`__
   - `Decorator syntax J2 (decorate..def) 
<http://mail.python.org/pipermail/python-dev/2004-August/047807.html>`__
   - `Decorators and docstrings 
<http://mail.python.org/pipermail/python-dev/2004-August/047827.html>`__
   - `block-based decorators and other block issues 
<http://mail.python.org/pipermail/python-dev/2004-August/047839.html>`__
   - `More concerns about decorators 
<http://mail.python.org/pipermail/python-dev/2004-August/047863.html>`__


More information about the Python-Dev mailing list