port to win32

[mindw]

Really fast and brutal fix to some win32 issues

[mindw]

ping?

[bitzl]

Windows support would be great!

[pfmoore]

I would also like to see Windows support. Is there any particular issue with this patch - it looks OK to me from a quick glance?

I would be happy to contribute time to testing and working on Windows support if the project needs people with Windows expertise.

[Ivoz]

https://github.com/kennethreitz/envoy might be helpful for subprocess running

[pfmoore]

On Windows, a number of the tests fail. But this is likely to be because the tests are not written to be portable, and this PR does not update the tests. I can look at creating a PR to make the tests portable to Windows, but it's going to be a bit of work. One thing - can someone tell me how to run a subset of the tests (one test or one test file) under spec? The suggestions I've found for nose don't seem to work for me with spec - I am tending to get "0 tests run", for example with spec cli or spec cli.py or spec tests/cli.py or spec tests/cli (you get the idea...)

[bitprophet]

What actual exception (or exceptions, if there's >1 likely) is pexpect throwing under Windows? Please try to catch that instead :)

[pfmoore]

It'll be an import error - from the import of pty within pexpect - because pty doesn't exist on Windows. It might be better just to do a straight platform check round the pty stuff.

[bitprophet]

I don't think we're modifying pexpect at all, just vendoring it, so I'd prefer to avoid touching it if possible. If these changes work well enough as-is and we can just change this line to except ImportError:, that's fine too. Just don't want a bare except.

[pfmoore]

Sorry, what I meant was in runner.run, just do

if pty and sys.platform == 'win32': raise RuntimeError("ptys not supported")

Then there's no need to import pexpect at all, and we can skip it altogether on Windows.

I'm currently putting together an alternative patch that puts the platform-specific stuff in a single file. It might make for a cleaner structure for future work.

[bitprophet]

Gotcha, and yea given the constraints of ptys on Windows, that makes sense. And big +1 on corralling platform specific stuff; again my philosophy for Windows support on my projects is basically "POSIX is 1st class citizen; Windows is 2nd class and lives off to the side". I don't like having to do that but it's the nature of the platforms & their distributions amongst the userbase.

[pfmoore]

Sounds good. The only thought I have is that sometimes the "POSIX first class" philosophy can lead to non-portable code which could actually be trivially made portable but wasn't simply because nobody thought about portability. But it's up to us Windows users to keep you honest on that one :-)

[pfmoore]

By the way, pip has switched to using invoke for some of its release process scripts, and I can confirm that with this patch, invoke works on Windows for that usage.

[bitprophet]

any particular reason for the blank line?

[mindw]

None. It's added automatically by GitExt.

[bitprophet]

Thanks for the contributions! Left a bunch of line comments, will wait for responses/changes.

@pfmoore Use spec --tests=tests/whatever.py, spec is still technically a nose plugin and that's a nose flag; it's what inv test uses: https://github.com/pyinvoke/invocations/blob/master/invocations/testing.py#L14

[bitprophet]

@pfmoore Oh, also, I would support updating the tests to be portable, but hopefully in a way that doesn't make them significantly more verbose/boilerplatey. I want to support Windows but I also don't want to sacrifice significant code clarity for it :(

If it's easier to just have a quick high level integration test file that can be run by Windows-using devs (esp given Travis-CI doesn't do Windows, there's no good way to have strong regression testing for the platform) then that might be a better approach.

[pfmoore]

@bitprophet Thanks for the test invocation hint - not sure how I didn't try that variation!

Test portability would probably need to be a mix of skipping the ones that will never work on Windows (pty stuff), a bit of fiddling round things like line ending assumptions. Probably the biggest one would be the OS commands being run in the tests - they use bash syntax in a few places, and assume a standard Unix command set is available. That can be fixed, but it could be a bit intrusive - I'll see how it goes.

A short term approach might be to just have a high-level integration test for now, with phased changes to the main tests to improve coverage as we go along.

One other thing I spotted in the tests is the ANSI colour support. Windows doesn't use ANSI codes, so that'll show up as garbage. How unhappy would you be with changes to the way that's done? It could be a dependency on colorama (which does colour in a cross-platform way) or vendoring it, or some wrappers around an optional dependency, falling back to ANSI on Unix and black-and-white on Windows?

Maybe the best approach would be a "compatibility" module that encapsulated the stuff that differs between platforms: That would keep the impact on code clarity minimal.

[bitprophet]

@pfmoore:

• Portability: The phased test support sounds reasonable, I'd like to prioritize getting the tool functional for Windows users, and wouldn't want "make the tests perfect on Windows" to hold that up.
• Colors: I need to port more of these Invoke-labeled Fab tickets over, but see Fabric #101 for my previous thoughts on color - colorama is on that list IIRC so if it does this well that might be a +1 for it. EDIT: Looks like 'blessings' was the top-of-mind (lol business terms) last time I looked at that ticket; we should see whether it handles x-platform. But let's keep that discussion over there.
• Compatibility module: as per my sub-comment higher up, my default is to treat POSIX as the primary implementation and have code branch off into Windows specific overrides as needed. I am not 100% wedded to that but IMO it makes things easier to reason about for what I see as the primary case, especially in situations where Windows-specific code is only in a few spots.

[bitprophet]

Merging this discussion into #119 so it's not in two places. Thanks!