[Python-Dev] difficulty of implementing phase 2 of PEP 302 in Python source

Phillip J. Eby pje at telecommunity.com
Thu Sep 28 01:41:18 CEST 2006

At 04:11 PM 9/27/2006 -0700, Brett Cannon wrote:

>On 9/27/06, Phillip J. Eby 
><<mailto:pje at telecommunity.com>pje at telecommunity.com> wrote:
>>At 02:11 PM 9/27/2006 -0700, Brett Cannon wrote:
>> >But it has been suggested here that the import machinery be rewritten in
>> >Python.  Now I have never touched the import code since it has always had
>> >the reputation of being less than friendly to work with.  I am asking for
>> >opinions from people who have worked with the import machinery before if
>> >it is so bad that it is worth trying to re-implement the import semantics
>> >in pure Python or if in the name of time to just work with the C
>> >code.  Basically I will end up breaking up built-in, .py, .pyc, and
>> >extension modules into individual importers and then have a chaining class
>> >to act as a combined .pyc/.py combination importer (this will also make
>> >writing out to .pyc files an optional step of the .py import).
>>The problem you would run into here would be supporting zip imports.
>I have not looked at zipimport so I don't know the exact issue in terms of 
>how it hooks into the import machinery.  But a C level API will most 
>likely be needed.

I was actually assuming you planned to reimplement that in Python as well, 
and hence the need for the storage/format separation.

>>   It
>>would probably be more useful to have a mapping of file types to "format
>>handlers", because then a filesystem importer or zip importer would then be
>>able to work with any .py/.pyc/.pyo/whatever formats, along with any new
>>ones that are invented, without reinventing the wheel.
>So you are saying the zipimporter would then pull out of the zip file the 
>individual file to import and pass that to the format-specific importer?

No, I'm saying that the zipimporter would simply call the format importers 
in sequence, as in your original concept.  However, these importers would 
call *back* to the zipimporter to ask if the file they are looking for is 

>>Thus, whether it's file import, zip import, web import, or whatever, the
>>same handlers would be reusable, and when people invent new extensions like
>>.ptl, .kid, etc., they can just register format handlers instead.
>So a sepration of data store from data interpretation for importation.  My 
>only worry is a possible explosion of checks for the various data 
>types.  If you are using the file data store and had .py, .pyc, .so, 
>module.so , .ptl, and .kid registered that might suck in terms of 
>performance hit.

Look at it this way: the parent importer can always pull a directory 
listing once and cache it for the duration of its calls to the child 
importers.  In practice, however, I suspect that the stat calls will be 
faster.  In the case of a zipimport parent, the zip directory is already 

Also, keep in mind that most imports will likely occur *before* any special 
additional types get registered, so the hits will be minimal.  And the more 
of sys.path is taken up by zip files, the less of a hit it will be for each 

>   And I am assuming for a web import that it would decide based on the 
> extension of the resulting web address?

No - you'd effectively end up doing a web hit for each possible 
extension.  Which would suck, but that's what caching is 
for.  Realistically, you wouldn't want to do web-based imports without some 
disk-based caching anyway.

>   And checking for the various types might not work well for other data 
> store types.  Guess you would need a way to register with the data store 
> exactly what types of data interpretation you might want to check.

No, you just need a method on the parent importer like get_data().

>Other option is to just have the data store do its magic and somehow know 
>what kind of data interpretation is needed for the string returned (e.g., 
>a database data store might implicitly only store .py code and thus know 
>that it will only return a string of source).  Then that string and the 
>supposed file extension is passed ot the next step of creating a module 
>from that data string.

Again, all that's way more complex than you need; you can do the same thing 
by just raising IOError from get_data() when asked for something that's not 
a .py.

>>Format handlers could of course be based on the PEP 302 protocol, and
>>simply accept a "parent importer" with a get_data() method.  So, let's say
>>you have a PyImporter:
>>      class PyImporter:
>>          def __init__(self, parent_importer):
>>              self.parent = parent_importer
>>          def find_module(self, fullname):
>>              path = fullname.split('.')[-1]+'.py'
>>              try:
>>                  source = self.parent.get_data(path)
>>              except IOError:
>>                  return None
>>              else:
>>                  return PySourceLoader(source)
>>See what I mean?  The importers and loaders thus don't have to do direct
>>filesystem operations.
>I think so.  Basically you want more of a way to stack imports so that the 
>basic importers are just passed the string of what it is supposed to load 
>from.  Other importers higher in the chain can handle getting that string.

No, they're full importers; they're not passed "a string".  The only 
difference between this and your original idea of an importer chain is that 
I'm saying the chained format-specific importers need to know who their 
"parent" importer (the data store) is, so they can be data-store 
independent.  Everything else can be done with that, and perhaps a few 
extra parent importer methods for stat, save, etc.

>>Of course, to fully support .pyc timestamp checking and writeback, you'd
>>need some sort of "stat" or "getmtime" feature on the parent importer, as
>>well as perhaps an optional "save_data" method.  These would be extensions
>>to PEP 302, but welcome ones.
>Could pass the string representing the location of where the string came 
>from.  That would allow for the required stat calls for .pyc files as 
>needed without having to implement methods just for this one use case.

Huh?  In order to know if a .pyc is up to date, you need the st_mtime of 
the .py file.  That can't be done in the parent importer without giving it 
format knowledge, which goes against the point of the exercise.  Thus, 
something like stat() and save() methods need to be available on the 
parent, if it can support them.

>>Anyway, based on my previous work with pkg_resource, pkgutil, zipimport,
>>import.c , etc. I would say this is how I'd want to structure a
>>reimplementation of the core system.  And if it were for Py3K, I'd probably
>>treat sys.path and all the import hooks associated with it as a single
>>meta-importer on sys.meta_path -- listed after a meta-importer for handling
>>frozen and built-in modules.  (I.e., the meta-importer that uses sys.path
>>and its path hooks would be last on sys.meta_path.)
>Ah, interesting idea!  Could even go as far as removing sys.path and just 
>making it an attribute of the base importer if you really wanted to make 
>it just meta_path for imports.

Perhaps, but then that just means you have to have a new variable for 
'sys.path_importer' or some such, just to get at it.  (i.e., code won't be 
able to assume it's always the last item in sys.meta_path).  So this seems 
wasteful and changing things just for the sake of change, vs. just keeping 
the other PEP 302 sys variables.  I just think the *implementation* of them 
can move to sys.meta_path, as that simplifies the main __import__ function 
down to just calling meta_path importers in sequence, modulo some package 

One other rather tricky matter is that the sys.path meta-importer has to 
deal with package __path__ management...  and actually, meta_path importers 
are supposed to receive a copy of sys.path...  ugh.  Well, it was a nice 
idea, but I guess you can't actually implement sys.path using a meta_path 
importer.  :(  For Py3K, we could drop the path argument to find_module() 
and manage it, but it can't be done and still allow current meta_path hooks 
to work right.

>>In other words, sys.meta_path is really the only critical import hook from
>>the raw interpreter's point of view.  sys.path, however, (along with
>>sys.path_hooks and sys.path_importer_cache) is critical from the
>>perspective of users, applications, etc., as there has to be some way to
>>get things onto Python's path in the first place.
>Yeah, I think I get it.  I don't know how much it simplifies things for 
>users but I think it might make it easier for alternative import writers.

That was the idea, yes.  :)

More information about the Python-Dev mailing list