> this file was independently reproduced by <name of trusted party>
Sounds risky, unless paired with strong policy about what a verifier should do.
e.g. if the verifier just grants network access to the buildbot that then checks what the result should look like, then compromise could remain invisible while the label silently downgrade to "independently downloaded by".
And I do not expect there are many parties that would be willing to provide such service beyond their own needs, while simultaneously not already providing that service in places where the infrastructure and policies are already set (such as Debian). Or at least they would already contribute towards build dependencies that do not break reproducibility on every other version bump.
I think running this with network access disabled is a good idea, but it's also important to note "was independently reproduced" is about "we challenged the SBOM and were able to produce this output from the stated inputs".
Whether the code in the SBOM subvertly downloads additional code (during build time or later during runtime) is still for the code reviewer to be found.
These people are so far gone in their fantasy land...
> So we either have to shrug and say, "don't use sdists if you want reproducible builds,"
What is this, I can't even... How can anyone expect to have a reproducible build's artifact if they downloaded the source?
Now, and most importantly, all this talk is about "pure" (as in Python-only) packages. This is, practically, worthless for most real Python projects because those rely on native bindings. Why should anyone expect PyPI to be able to build those if no standard procedure exists for building them is beyond comprehension.
They use the word "Reproducible Builds" for linking the VCS commit to PyPI's repacked source code upload, that's why "reproduce the source code" sounds a little confusing.
For projects with native bindings you can still fairly trivially solve this with SBOMs, and some Linux distributions are doing this for many years now. The compiler is a dependency that needs documenting, and they explicitly write they need means to re-create the documented environment from that SBOM.
> For projects with native bindings you can still fairly trivially solve this with SBOMs
For bitwise identical results there are many small details that need to be taken care of, even in the pure-Python case: Ensuring that files appear in the wheel (zip) file in the same order (glob() doesn't guarantee this and you can definitely see different orders from run to run), timestamps are set to some fixed standard time. Any kind of build-time code generation or pulling in of external info like git commit hashes needs to be handled. Then once you include compiled build dependencies, there might be nondeterminism in the order that a multithreaded compiler writes object code, hidden timestamps, hidden absolute paths, and things like runtime detection of CPU version leading to different instructions being emitted on different machines.
I'm well aware of those issues (having worked on this for many years), but I still believe the majority of python packages won't be affected by this. :)
Most of these you would fix once on the relevant [build-system] and be done with it, no need to fix each individual python library.
Please try building with conda-build to understand the actual problems this entails. Your solution is a fantasy because of how library linkage works. Dynamically loaded libraries need to be able to find each other at runtime. In order to do that, they need to store the location (the filesystem path) to the library they want to load. How these links are resolved is beyond the scope of Python (on Linux, this is managed by ld and ldconf).
A lot of Python package maintainers don't understand this problem and don't understand how to make a portable library. Often times builds contain either an absolute path to another library, or a relative path... into nowhere, or a collection of paths... into questionable places. And there's no standard that tells the library authors how to do this the right way. As long as there's no standard, anyone can claim to have done it right and expect their users to accommodate them (which is what currently happens in Python).
About your other ideas: making a compiler a dependency is... an awful idea. Unless you can require the same compiler is used for every dependency in your project, at the minimum, you will have an uncontrolled (how are you going to make sure that the right compiler is used to build a package?) zoo of compilers and their versions in your project. In the worst case, compilers will embed their signatures in the generated binary code preventing other binaries from loading such code, if they are generated with a different compiler version (kind of like what happens if you try to load Linux drivers that weren't compiled with the same compiler that compiled the kernel).
* * *
Like I said: these people have next to no practical experience with the problem they are trying to solve. Why can't they just go some place else and apply themselves elsewhere is beyond my comprehension.
I don't think they are talking about getting the artifact at the same time as the source or any such thing. They are saying, the metadata required to reproduce a build is not present in that sdist format, and there's no place to attach it. So, if you got an artifact and separately got the source, you couldn't verify it without another source of information about how to do the build itself in exactly the same way. It goes beyond pure reproducibility as well. Without that metadata to set up your environment, you may see bugs or other differences in your build that are not in the distributed artifact.
OK. I see. But, even if this is the problem, the solution they are trying for is bad. What Python needs is a project definition (like what Ada has in GNAT Project Manager). PyPA repeatedly proved themselves incapable of coming up with something like this (first setup.cfg, then pyproject.toml), and I don't expect them to independently discover the solution to the problem that was discovered before Python was invented. They are not the kind of people that would be able to do that. They've been at it for some fifteen years and every time they roll out a new iteration it just gets worse.
Yeah I agree. The Python language sucks for stuff like this. Reproducibility is hardly a priority in a language where compatibility isn't even a priority. I like Python in spite of its many warts and wish it was possible to start a new language without the mistakes and have it be as popular as Python, but I don't see that ever happening. I think they've abandoned simple logic like the classic "There should be one-- and preferably only one --obvious way to do it." Nevertheless, I appreciate people trying to make the best of it.
For python, it seems very difficult to actually do reproducible builds in a way that is useful, because python wheels have very implicit assumptions about paths and environemtns that do not happen anywhere else.
"Please tell me where the masses are going so that I can lead them!"
Brett Cannon made his career on the fake CoC power games which led to the cancellation of Tim Peters and others.
The "build the PyPI server" issues were discussed here at length in the past month. Unsurprisingly, Brett Cannon picks up the issue, presents it as his own idea and runs for yet another position of power.
Nothing will get done if you elect him. He just gets titles on the back of other people.
17 comments
[ 3.0 ms ] story [ 28.1 ms ] threadSounds risky, unless paired with strong policy about what a verifier should do. e.g. if the verifier just grants network access to the buildbot that then checks what the result should look like, then compromise could remain invisible while the label silently downgrade to "independently downloaded by". And I do not expect there are many parties that would be willing to provide such service beyond their own needs, while simultaneously not already providing that service in places where the infrastructure and policies are already set (such as Debian). Or at least they would already contribute towards build dependencies that do not break reproducibility on every other version bump.
Whether the code in the SBOM subvertly downloads additional code (during build time or later during runtime) is still for the code reviewer to be found.
> So we either have to shrug and say, "don't use sdists if you want reproducible builds,"
What is this, I can't even... How can anyone expect to have a reproducible build's artifact if they downloaded the source?
Now, and most importantly, all this talk is about "pure" (as in Python-only) packages. This is, practically, worthless for most real Python projects because those rely on native bindings. Why should anyone expect PyPI to be able to build those if no standard procedure exists for building them is beyond comprehension.
For projects with native bindings you can still fairly trivially solve this with SBOMs, and some Linux distributions are doing this for many years now. The compiler is a dependency that needs documenting, and they explicitly write they need means to re-create the documented environment from that SBOM.
For bitwise identical results there are many small details that need to be taken care of, even in the pure-Python case: Ensuring that files appear in the wheel (zip) file in the same order (glob() doesn't guarantee this and you can definitely see different orders from run to run), timestamps are set to some fixed standard time. Any kind of build-time code generation or pulling in of external info like git commit hashes needs to be handled. Then once you include compiled build dependencies, there might be nondeterminism in the order that a multithreaded compiler writes object code, hidden timestamps, hidden absolute paths, and things like runtime detection of CPU version leading to different instructions being emitted on different machines.
Most of these you would fix once on the relevant [build-system] and be done with it, no need to fix each individual python library.
Please try building with conda-build to understand the actual problems this entails. Your solution is a fantasy because of how library linkage works. Dynamically loaded libraries need to be able to find each other at runtime. In order to do that, they need to store the location (the filesystem path) to the library they want to load. How these links are resolved is beyond the scope of Python (on Linux, this is managed by ld and ldconf).
A lot of Python package maintainers don't understand this problem and don't understand how to make a portable library. Often times builds contain either an absolute path to another library, or a relative path... into nowhere, or a collection of paths... into questionable places. And there's no standard that tells the library authors how to do this the right way. As long as there's no standard, anyone can claim to have done it right and expect their users to accommodate them (which is what currently happens in Python).
About your other ideas: making a compiler a dependency is... an awful idea. Unless you can require the same compiler is used for every dependency in your project, at the minimum, you will have an uncontrolled (how are you going to make sure that the right compiler is used to build a package?) zoo of compilers and their versions in your project. In the worst case, compilers will embed their signatures in the generated binary code preventing other binaries from loading such code, if they are generated with a different compiler version (kind of like what happens if you try to load Linux drivers that weren't compiled with the same compiler that compiled the kernel).
* * *
Like I said: these people have next to no practical experience with the problem they are trying to solve. Why can't they just go some place else and apply themselves elsewhere is beyond my comprehension.
Brett Cannon made his career on the fake CoC power games which led to the cancellation of Tim Peters and others.
The "build the PyPI server" issues were discussed here at length in the past month. Unsurprisingly, Brett Cannon picks up the issue, presents it as his own idea and runs for yet another position of power.
Nothing will get done if you elect him. He just gets titles on the back of other people.