Handling deprecations in the face of PEP 384
As I clean up Python/import.c and move much of its functionality into Lib/imp.py, I am about to run into some stuff that was not kept private to the file. Specifically, I have PyImport_GetMagicTag() and NullImporter_Type which I would like to chop out and move to Lib/imp.py.
On Fri, Apr 20, 2012 at 6:59 PM, Brett Cannon <brett@python.org> wrote:
Yeah, PyImporter_GetMagicTag() looks like a public API, parallel with PyImporter_GetMagicNumber(). Maybe it was accidentally not documented? I'm not sure when it was introduced. Should we even deprecate it? I'd say do the same thing you're doing for GetMagicNumber(). NullImporter_Type looks like it was accidentally not made static, so don't fret about that. -- --Guido van Rossum (python.org/~guido)
On Sat, Apr 21, 2012 at 12:16 PM, Guido van Rossum <guido@python.org> wrote:
I'd keep it and just make it a convenience wrapper for the call back into the Python code.
NullImporter_Type looks like it was accidentally not made static, so don't fret about that.
Yeah, the lack of the Py_ prefix suggests this one being visible is just an accident of the implementation, and the name is unusual enough that it never caused a symbol collision for any third parties. Cheers, Nick. -- Nick Coghlan | ncoghlan@gmail.com | Brisbane, Australia
On Apr 20, 2012, at 09:59 PM, Brett Cannon wrote:
I'd have to go back into my archives for the discussions about the PEP, but my recollection is that we intentionally made PyImport_GetMagicTag() a public API method. Thus no leading underscore. It's a bug that it's not documented, but OTOH, it's unlikely there are, or would be, many consumers for it. Strictly speaking, I do think you need to deprecate the APIs. I like Nick's suggestion to make them C wrappers which just call back into Python. -Barry
On Sat, Apr 21, 2012 at 12:10, Barry Warsaw <barry@python.org> wrote:
That was my plan, but the amount of code it will take to wrap them is making me not care. =) For PyImport_GetMagicTag() I would need to expose a new attribute on sys or somewhere which specifies the VM name. For PyImport_GetMagicNumber() I have to do a bunch of bit twiddling to convert a bytes object into a long which I am just flat-out not in the mood to figure out how to do. And all of this will lead to the same amount of C code as there currently is for what is already implemented, so I just don't care anymore. =) But I'm glad the clarifications are there about the stable ABI and how we are handling it.
On Sat, Apr 21, 2012 at 4:17 PM, Brett Cannon <brett@python.org> wrote:
I thought I already (mostly) worked it all out in that patch on issue13959. I felt really good about the approach for the magic tag and magic bytes. Once find_module() and reload() are done in imp.py, I'm hoping to follow up on a few things. That includes the unresolved mailing list thread about sys.implementation (or whatever it was), which will help with the magic tag. Anyway, I don't want to curtail the gutting of import.c quite yet (as he hears cries of "bring out your dead!"). -eric p.s. I understand your sentiment here, considering that mothers are often exhausted by childbirth and the importlib bootstrap was a big baby. You were in labor for, what, 6 years. <wink> [There's an analogy that could keep on giving. :) ]
On Sat, Apr 21, 2012 at 20:54, Eric Snow <ericsnowcurrently@gmail.com>wrote:
You didn't update Python/import.c in your patches so that the public C API continued to function. That's what is going to take a bunch of C code to continue to maintain, not the Python side of it.
Even w/ all of that gutted, a decent chunk of coding is holding on to dear life thanks to PyImport_ExecCodeModuleObject() (and those that call it). IOW the C API as it is currently exposed is going to end up being the limiting factor of how many lines get deleted in the very end.
It's also about maintainability. It isn't worth upping complexity just to shift some stuff into Python code, especially when it is such simple stuff as the magic number and tag which places practically zero burden on other VMs to implement.
All that PEP 384 gives you is that you MAY deprecate certain API (namely, all API not guaranteed as stable). If an API is not in the restricted set, this doesn't mean that it SHOULD be deprecated at some point. So there is no need to deprecate anything. OTOH, if the new implementation cannot readily support the API anymore, it can certainly go away. If it was truly private (i.e. _Py_*), it can go away immediately. Otherwise, it should be deprecated-then-removed. Regards, Martin
On Fri, Apr 20, 2012 at 6:59 PM, Brett Cannon <brett@python.org> wrote:
Yeah, PyImporter_GetMagicTag() looks like a public API, parallel with PyImporter_GetMagicNumber(). Maybe it was accidentally not documented? I'm not sure when it was introduced. Should we even deprecate it? I'd say do the same thing you're doing for GetMagicNumber(). NullImporter_Type looks like it was accidentally not made static, so don't fret about that. -- --Guido van Rossum (python.org/~guido)
On Sat, Apr 21, 2012 at 12:16 PM, Guido van Rossum <guido@python.org> wrote:
I'd keep it and just make it a convenience wrapper for the call back into the Python code.
NullImporter_Type looks like it was accidentally not made static, so don't fret about that.
Yeah, the lack of the Py_ prefix suggests this one being visible is just an accident of the implementation, and the name is unusual enough that it never caused a symbol collision for any third parties. Cheers, Nick. -- Nick Coghlan | ncoghlan@gmail.com | Brisbane, Australia
On Apr 20, 2012, at 09:59 PM, Brett Cannon wrote:
I'd have to go back into my archives for the discussions about the PEP, but my recollection is that we intentionally made PyImport_GetMagicTag() a public API method. Thus no leading underscore. It's a bug that it's not documented, but OTOH, it's unlikely there are, or would be, many consumers for it. Strictly speaking, I do think you need to deprecate the APIs. I like Nick's suggestion to make them C wrappers which just call back into Python. -Barry
On Sat, Apr 21, 2012 at 12:10, Barry Warsaw <barry@python.org> wrote:
That was my plan, but the amount of code it will take to wrap them is making me not care. =) For PyImport_GetMagicTag() I would need to expose a new attribute on sys or somewhere which specifies the VM name. For PyImport_GetMagicNumber() I have to do a bunch of bit twiddling to convert a bytes object into a long which I am just flat-out not in the mood to figure out how to do. And all of this will lead to the same amount of C code as there currently is for what is already implemented, so I just don't care anymore. =) But I'm glad the clarifications are there about the stable ABI and how we are handling it.
On Sat, Apr 21, 2012 at 4:17 PM, Brett Cannon <brett@python.org> wrote:
I thought I already (mostly) worked it all out in that patch on issue13959. I felt really good about the approach for the magic tag and magic bytes. Once find_module() and reload() are done in imp.py, I'm hoping to follow up on a few things. That includes the unresolved mailing list thread about sys.implementation (or whatever it was), which will help with the magic tag. Anyway, I don't want to curtail the gutting of import.c quite yet (as he hears cries of "bring out your dead!"). -eric p.s. I understand your sentiment here, considering that mothers are often exhausted by childbirth and the importlib bootstrap was a big baby. You were in labor for, what, 6 years. <wink> [There's an analogy that could keep on giving. :) ]
On Sat, Apr 21, 2012 at 20:54, Eric Snow <ericsnowcurrently@gmail.com>wrote:
You didn't update Python/import.c in your patches so that the public C API continued to function. That's what is going to take a bunch of C code to continue to maintain, not the Python side of it.
Even w/ all of that gutted, a decent chunk of coding is holding on to dear life thanks to PyImport_ExecCodeModuleObject() (and those that call it). IOW the C API as it is currently exposed is going to end up being the limiting factor of how many lines get deleted in the very end.
It's also about maintainability. It isn't worth upping complexity just to shift some stuff into Python code, especially when it is such simple stuff as the magic number and tag which places practically zero burden on other VMs to implement.
All that PEP 384 gives you is that you MAY deprecate certain API (namely, all API not guaranteed as stable). If an API is not in the restricted set, this doesn't mean that it SHOULD be deprecated at some point. So there is no need to deprecate anything. OTOH, if the new implementation cannot readily support the API anymore, it can certainly go away. If it was truly private (i.e. _Py_*), it can go away immediately. Otherwise, it should be deprecated-then-removed. Regards, Martin
participants (6)
-
"Martin v. Löwis" -
Barry Warsaw -
Brett Cannon -
Eric Snow -
Guido van Rossum -
Nick Coghlan