main: build in a virtual env by default #18
Conversation
Codecov Report
@@ Coverage Diff @@
## master #18 +/- ##
==========================================
Coverage 100.00% 100.00%
==========================================
Files 6 9 +3
Lines 256 382 +126
==========================================
+ Hits 256 382 +126 |
|
@pganssle @gaborbernat could you review? |
|
I choose |
FFY00
force-pushed
the
envbuild
branch
2 times, most recently
from
90e0401 to
ee9a9fd
Compare
|
I refactored the patch based on the comments, should be better now. |
FFY00
force-pushed
the
envbuild
branch
4 times, most recently
from
7b5c426 to
8db9c11
Compare
|
I think overall this PR is good, but I think it's a bit under-tested considering the heavy use of mocks. I've never found it terribly easy to test CLI applications, but I think I would get rid of I also think that it would be better to use fake packages designed to test a single thing, rather than real packages that you pull from PyPI (though presumably — at least to start with — you can still hit PyPI during the tests anyway). You can see we do something like that in setuptools; in that case, we define the files in code and then construct the package in a temporary location as part of the test, but you could just as easily have a "tests/packages" directory or something with a bunch of stub packages in it. |
|
I think adding proper tests with fake packages is great, it's something that I have in my TODO. |
FFY00
force-pushed
the
envbuild
branch
2 times, most recently
from
9515d4a to
7ed64ba
Compare
IMHO they should be part of the test suite. They are integration tests, and as such should be marked as such in the tests, giving you the option to not run them if you want a fast test. |
|
Okay, I can add a |
|
What I tend to do is to make it auto on by default for nox, but disabled if you just invoke pytest. If you check tox/virtualenv you'll find exact how to of this. |
|
I agree with this :). On by default in nox but off in pytest. |
build/__main__.py
Outdated
|
|
||
| builder.build(dist, outdir) | ||
| if isolation: | ||
| with pep517.envbuild.BuildEnvironment() as env: |
There was a problem hiding this comment.
I just had a conversation that made me realize that this may actually not be sufficient to achieve proper isolation. I was under the impression that BuildEnvironment() would create an isolated build environment, but it actually adds a layer on top of your existing environment. For example, if I create a project with the following setup.py:
from setuptools import setup
try:
import attr
raise ValueError("Attrs should not be available!")
except ImportError:
pass
setup()If I run python -m build on the project without installing attrs it works fine, but if I pip install attrs into the same environment that build is installed in, it raises the exception.
That's especially troubling because build itself has runtime dependencies, so any undeclared dependencies on packaging, pep517, pyparsing, toml, etc, would go unnoticed.
I don't know if we should try and fix this here or in pep517. I assume @gaborbernat will have a better understanding of the options than I do.
There was a problem hiding this comment.
AFAIK (@takluyver would know for sure) it's the caller responsibility to ensure that on the host python only the standard library is available. So I'd say the responsibility would be on this lib to fix this. In practice, this would imply that you would need to remove from sys.path all paths that are not stdlib (you can get this via https://docs.python.org/3/library/sysconfig.html#sysconfig.get_path and passing in stdlib).
There was a problem hiding this comment.
"The caller" as in the user of python-build, or "The caller" as in python-build?
I think the way it is now we're basically doing the equivalent of building a virtualenv with --system-site-packages, which is not a good default, so I feel strongly that we should not be doing that. How does virtualenv achieve this?
I suppose this explains why pip has their own logic for build isolation and doesn't directly use pep517's. We may want to copy them (though their approach does bring in at least one problem).
There was a problem hiding this comment.
How does
virtualenvachieve this?
It knows the interpreter deeply to be able to tell how to achieve. There's no single global answer, various interpreters do it differently. Here's a list https://github.com/pypa/virtualenv/blob/master/src/virtualenv/discovery/py_info.py#L35-L103 of options that can influence things. Either pick pips way or pull in virtualenv package to create a virtual environment the way I see it. I agree we should not pull in global site package.
There was a problem hiding this comment.
We're thinking of removing BuildEnvironment from pep517 (along with all the higher-level code that uses it) - it's not very good, as you've noticed, and it distracts people from the useful part of the pep517 package. See pypa/pyproject-hooks#83
I wrote BuildEnvironment and the code using it as a kind of example of the simplest possible PEP 517 frontend that puts all the pieces together (installing dependencies, calling the hooks with those dependencies available). The spec only says that the build environment SHOULD be isolated, so that 'overlay' model is technically compliant, but it's not best practice.
There was a problem hiding this comment.
I think having this high level code would be in-scope for this project, we just need to be careful with what we choose to export as public API.
So I propose for now to move BuildEnvironment here as private API, and then in the future think about exporting it.
I wrote
BuildEnvironmentand the code using it as a kind of example of the simplest possible PEP 517 frontend that puts all the pieces together (installing dependencies, calling the hooks with those dependencies available).
I have expressed this opinion previously, but I'll reiterate as I think it's relevant. Nothing you consider an "example" should be ever exported publicly, being this API, CLI, etc. Once you make something public, people will use it. Providing code as an example forces people to copy it, tailor it to their needs, and most importantly, lets you off the hook 😄.
There was a problem hiding this comment.
Yeah, I'm right with you there. I must have been feeling rather gung-ho when I decided to make it part of the library. ;-)
There was a problem hiding this comment.
I don't think the issue here was if it's private or not. The issue is that it's not working, so we should use something else, which can be private. I'd say perhaps virtualenv, but I'm happy with venv too.
There was a problem hiding this comment.
I don't think the current pep517 code should be hard to fix up, we just need to adjust sys.path, no? I think virtualenv would be the direction forward if we needed to deal with custom interpreters, but we don't. I would like to avoid adding a new dependency.
There was a problem hiding this comment.
I don't think the current
pep517code should be hard to fix up, we just need to adjustsys.path, no?
Well sure, in theory, you just* need to fix the sys.path; however, in practice, it's non-trivial how to fix it up. Things differ per python implementation what needs to stay and what needs to go. virtualenv contains this complex logic. Hence why suggesting we use the slower venv on python 3, as that does not pull in any dependency 🤷
|
The integration tests already run in the CI, and they are passing. This PR is already big, I don't think adding more stuff is a good idea. I see no reason to block it due to the integration tests. IMO we should merge it and then work on the integration tests. I would definitely block a release though. |
I have a distinct feeling that they are passing by chance, not by design; because how they are set up (e.g. the reason why a sdist is unusable on the master, but the CI still passes; or how even without using |
|
|
||
|
|
||
| def test_isolated_environment(mocker): | ||
| with build.env.IsolatedEnvironment() as env: |
There was a problem hiding this comment.
I think this test should run in an isolated python environment (one that has no pip/setuptools/etc installed) so there's really no chance things are being pulled from some other path on the python path. E.g. is this past failing if you change from sys. get_scheme_names to that manual thing you defined before? It should if it's an accurate test.
There was a problem hiding this comment.
That "manual thing" I had before was hardcoding the scheme names, it should work the same. Why would this be failing?
really no chance things are being pulled from some other path on the python path
What things being pulled? There is nothing being pulled. In this test I check if the environment is correct, after that I check if pip is called correctly, but we never actually call pip. I look at the path and make sure it is correct.
For eg. I import sysconfig before entering the isolated environment and make sure purelib or platlib are not there.
for path in ('purelib', 'platlib'):
assert sysconfig.get_path(path) not in os.environ['PYTHONPATH'].split(os.pathsep)There was a problem hiding this comment.
I had before was hardcoding the scheme names, it should work the same.
It should have been failing on pypy, as pypy uses different scheme names then what you hardcoded. The fact that it did not, shows something is not totally correct in this sense.
For eg. I import
sysconfigbefore entering the isolated environment and make surepureliborplatlibare not there.
To be fair I'm not sure this is enough. I'd expect we'd isolate from any potential PYTHONPATH the user might have set, or additional paths the user injects via user/sitecustomize.
There was a problem hiding this comment.
It shouldn't.
>>>> import sysconfig
>>>> sysconfig.get_path('purelib')
'/opt/pypy/site-packages'
>>>> sysconfig.get_path('platlib')
'/opt/pypy/site-packages'If you read the code you'll notice I remove both the default paths and the different schemes.
The test is correct and working as expected.
|
The integration tests are definitely not passing by chance. I agree that we should have more of them but I keep my position this should not block this PR. We have unit tests which should fail when someone changes the behavior, and we have integration tests that build both this project and pip. To me, this should be enough to merge this PR. |
|
@pganssle can you have a look? |
There was a problem hiding this comment.
I'm inclined to say that @gaborbernat's review is enough, because he's really the expert here — considering he maintains multiple virtualenv-related CLI tools and I primarily maintain libraries — and I doubt that I'll be able to grasp the implications of all of these changes just reading through the code in the limited time I have available for this sort of thing.
Is there anything particularly controversial or any major decisions that I'm glossing over here? I can probably form opinions on individual questions a lot easier than I can grok the entire change.
|
|
||
| builder = build.ProjectBuilder() | ||
|
|
||
| assert builder._build_system == build._DEFAULT_BACKEND |
There was a problem hiding this comment.
I am not really an expert in testing CLIs, so maybe this is the most expedient way to do this, but my general rule for tests is that I only test the public API, so anything that's making assertions about the internal private structure is usually off-limits (or I mark it with a comment or something to indicate why this is the only / best way to test it).
Put another way, if you did an otherwise safe refactoring of the ProjectBuilder class, you would need to change this test, but that defeats one of the purposes of tests — allowing you to refactor with confidence, because you've made assertions about the behavior that should remain invariant.
Of course, I rarely have to test CLIs, so I think the rules may be a little different there (when there's no public API available from Python anyway).
There was a problem hiding this comment.
I am not really an expert in testing CLIs, so maybe this is the most expedient way to do this, but my general rule for tests is that I only test the public API, so anything that's making assertions about the internal private structure is usually off-limits (or I mark it with a comment or something to indicate why this is the only / best way to test it).
We are not testing CLI here, we are testing ProjectBuilder, which is public API. python-build provides a CLI but also a minimal API to build projects, which is ProjectBuilder.
Put another way, if you did an otherwise safe refactoring of the
ProjectBuilderclass, you would need to change this test, but that defeats one of the purposes of tests — allowing you to refactor with confidence, because you've made assertions about the behavior that should remain invariant.
If I do a safe refactoring of ProjectBuilder these tests should not fail. They might break if I change ProjectBuilder's public API.
There was a problem hiding this comment.
If I do a safe refactoring of
ProjectBuilderthese tests should not fail. They might break if I changeProjectBuilder's public API.
Well, if you removed or renamed build._DEFAULT_BACKEND or ProjectBuilder._build_backend, these tests would fail, but presumably those are not part of the public interface.
It seems like our public interface exposes a hook attribute, which would be the PEP517HookCaller object, and that does expose a .build_backend attribute, publicly.
I think you'd need to ease off the mocking a bit, but you can define a top-level DEFAULT_BACKEND in this test file, and then do:
assert builder.hook.build_backend == DEFAULT_BACKENDUnless it is unintentional that builder.hook is part of the public API, or that it is guaranteed to be a Pep517HookCaller object, in which case probably mocking out the default backend, calling .build and then using Mock.assert_called_once or Mock.assert_called_once_with would be more approriate.
IMHO it's a matter of taste at the end of the day. Looking through the code I'm not confident that we have a good enough test suite to validate the approach here. So I was advocating for more tests; ideally on top of moving our current integration tests from CI to pytest. @FFY00 wants to move ahead as is now. The current state at a glance looks ok, so I'll not block it, though the tests we have does not give me enough confidence to approve it either. |
| def test_missing_backend(mocker, pyproject_no_backend_mock): | ||
| mocker.patch('importlib.import_module') | ||
| mocker.patch('{}.open'.format(build_open_owner), pyproject_no_backend_mock) | ||
| mocker.patch('pep517.wrappers.Pep517HookCaller') |
There was a problem hiding this comment.
Why is this mocked out? Because the constructor will try and import the build backend? If so, can we try a more targeted mock that would simply provide the build backend?
There was a problem hiding this comment.
Yes, this can be removed when we refactor the test suite to use proper packages.
I tend to agree that the testing I see here is not really adequate given that this is the core functionality of this module. I am very dubious about the heavy use of mocks. If it were me, I'd take the time to do the testing beforehand, because I know that I very rarely go back to add more tests, and if I don't add the tests stuff will be fundamentally broken, but that's my workflow. I don't think we should do a production release of this software without adding the tests, but if there's some compelling reason to merge first and refactor the tests later (like avoiding annoying merge conflicts if/when we move to the |
| def test_default_backend(mocker, empty_file_mock): | ||
| mocker.patch('importlib.import_module') | ||
| mocker.patch('{}.open'.format(build_open_owner), empty_file_mock) | ||
| mocker.patch('pep517.wrappers.Pep517HookCaller') |
There was a problem hiding this comment.
I'm not super familiar with this mocker fixture. Is there any way to use something equivalent to autospec with it? If we really need to mock this stuff out, we should be as surgical with the mocking as possible.
There was a problem hiding this comment.
This plugin provides a mocker fixture which is a thin-wrapper around the patching API provided by the mock package
https://github.com/pytest-dev/pytest-mock/
Everything supported by unittest.mock.path is supported by mocker.patch.
Yes, I want to avoid having to keep this in sync with new changes, especially since it is such a big PR. The current test suite is not optimal, but it does its job. I want to refactor it and remove a lot of the mocks, but that will have a lot of conflicts with this. |
You would not need to if you switch to improving the test suite first and then coming back to this 🤔 |
There was a problem hiding this comment.
https://github.com/FFY00/python-build/issues/52 is follow-up ticket for this, I feel like we'll need to fix this PR soon 👍 but seems @FFY00 wants to do no more with it.
|
I was gonna open a PR to this branch but would be a bit of a weird review process, I prefer it this way. We can always revert the commits if needed. |
Not really once we do the refactor needed, e.g. src layout, the revert would cause merge conflicts. |
|
Which can be solved. Only if you build on top of the current code it might become an issue. There are no plans for that right now, what we have in hand is nox > tox (in.c src layout change) and refactoring the tests. |
|
You forgot about https://github.com/FFY00/python-build/issues/19 which is bound to make changes everywhere |
|
I would stall it until all this is sorted out. |
In the end no need for that. Reverting it has no value at the end of the day because it's not like we'd go ahead without a similar feature. If it's not working we can just fix-it in a follow-up PR. In the end, the decision is with you as primary maintainer, but it would likely prevent me from doing PRs. Changing code that's not auto-formatted raises too many potential style arguments I'm not willing to engage as part of my free time. |
Hence the "if needed".
You can make your PRs and after approving I can push a commit fixing the style and squash merge. You don't need to worry about this, feel free to send code PRs without worrying about the style. This would be temporary until we adopt black. Anyway, this conversation is not going anywhere, we should focus on actually writing the code 😛. |
|
No thanks I'll wait until black is in place 👍will take a look at tox/nox though. |
As discussed in [1] python-build should now build inside a virtual
environment, as recommended in PEP517. This patch implements that
functionality and sets it as the default behavior. It also introduces a
--no-isolation flag to disable this.
[1] https://discuss.python.org/t/moving-python-build-to-pypa/4390
Signed-off-by: Filipe Laíns lains@archlinux.org