Quick question if anyone is around - should I expect cucumber/aruba to work with Windows? I have ruby 2.2.3p173 and aruba 0.13.0. When I try to have it run a command, it says that cmd is undefined in the Processes module. I see that he windows build hasn't worked in a while. Should I not be using this?
Great! I think at best would be to have a look at the build logs on AppVeyor first and make sure all existing feature tests are green. A short list of issues I'm aware of.
- File System Modes: Windows only supports something like 644, 666 and so on. It's not that flexible like UNIX
- Runner: Make sure, that all commands can run on windows via cmd.exe (this is partially implemented) + bonus make sure, we can test powershell scripts on windows
- Environment variables: Windows is case in-sensitive regarding environment variables, that's "a problem" because we convert it via with
#to_h. This is partially implemented and needs to be rebased. whichcommand: Make sure command paths can be resolved - this is related to 2. and should be fixed first.- Add some more documentation about how to use
arubaon Windows if needed
I also added a platform tag to our issued tracker: https://github.com/cucumber/aruba/issues?q=is%3Aissue+is%3Aopen+label%3Aplatform%2Fwindows
@dg-ratiodata Had a quick look at the failing Windows specs.
- File System Modes: what is your thought on that? Should we fail if specifying 666 or should it be automagically converted to 644? (BTW, the Windows FS has much better ACLs but they don't work like that)
- Runner: Why do they need to run with cmd.exe? The problem I see is that the tests use
- echo: built-in command
- true: AFAIK it does not exist, but it could be changed
- env: built-in command called set
- cat: AFAIK it does not exist
- shell script to redirect the output of echo to stderr: cmd can do the same
- No comment
- Same as 2., the problem is that the commands do not exist or are built-in
- No comment
- NEW spec/aruba/api/environment/restore_env_spec.rb - wrong number of parameters for delete (2 specs out of 45): fixed in paoloambrosio/aruba@fix-windows-platform.
- NEW spec/aruba/aruba_path_spec.rb:101 - Windows filesystem does not support counting blocks
To fix 2. and 4. I was thinking of re-implementing those commands as Ruby scripts (they are 2 or 3 lines each) to be executed by the specs so that they work in the same way on all platforms.
I believe that 7. should return what Ruby does, so Nil on Windows, and it should be tested.
For 7. I believe that stat on Windows does not expose the block size. From the macro hell that is the filesystem API on the MRI, this might be the data structure used: https://msdn.microsoft.com/en-us/library/14h5k7ff.aspx
I would design the test suite assuming that it is run on the standard filesystem for the platform (any "standard" filesystem on Linux and OSX, and NTFS on Windows). For example I wouldn't do anything that might fail in the case-insensitive filesystem in OSX. We should be able to find a representative superset that works on all common combinations of FS and OS.
@paoloambrosio Wow! :smile: Thanks a lot about getting into this. First one quick question: Where do 1., 2. ... refer to? Build errors? About 7: Normally I add things I need myself and think they're useful for others as well, but I can't remember for what project I added the block size thing. But I think I hat in mind, that windows does not support this kind of functionality and should raise "NoImplmented"-error or so.
If you like, send me a PR for your changes. We then can pair on this. Just push the branch to the
aruba/repo.
I'm really glad that you found some time to have a look into this.
First, I think Aruba tests shouldn't run any apps it doesn't own. This includes using bash, zsh, echo and related for shell. It would be much better to have ruby bash/zsh "shim" that just supports the given "OS feature". The other extreme would be to just run all tests in Docker - which would guarantee consistency without endless workarounds and issues related to "local setup". This is because Travis is not a good place for "development".
For blocks counting on Windows, I'd just either approximate it (cluster size * (1+file_size/cluster + file_count + dir_count)) or expect a tool like du.exe to be installed. (But, for the latter, there's a sever limit to the argument size for a command on Windows, so that complicates things). I'm sure users would prefer an approximation rather than a perfectly accurate implementation no one will ever implement in Ruby...
@dg-ratiodata - it's mostly bugfixes and new features (like Docker), so there shouldn't be any conflicts. If that's an issue, it's best to switch master to 1.x, branch off 0.14 for backporting, and just focus on releasing 1.x ASAP. It's no problem to then work on 2.x a month from now - no one will complain because of "too many major releases". Aruba is complex, so learning is expected. And if learning is reflected in "multiple major releases", that's just an indication of how much learning (or rework) was needed, nothing else. I'm guessing it's actually more useful to just skip to 2.x (master - without releasing), branch off 1.x as unreleased as "work in progress" (too keep the branches). E.g. I'm (as a contributor) only interested in master and not backward compatibility or deprecations - so it makes no sense to add an artificial "burden" on master. You can always work on a "transition release" - but only if there's a genuine need in the community. Master branch should be optimized for quickly accepting PRs. (If PRs don't have priority over other work, contributing feels very discouraging - as if there's a ton of "bureaucracy"). It doesn't make sense to expect contributors to make backports of their own fixes (that they don't need themselves).
@dg-ratiodata - if there's a lot of scrunity for PRs on master (like now), there's no need for a release/1.0.0 branch. Release branches don't fix a discipline/quality problem, they only create a branch managing problem. (Especially if plans aren't communicated to contributors effectively).
cucumber/aruba#363
@paoloambrosio Good point. From my understanding - current state of discussion - It will only solve problem for developers developing on Linux. We will still have issues on Windows. And we need a fix for that. I cannot think of a container solution for windows which has no licensing issues beside the fact that there is none so far - from my knowledge.
Since I'm not developing on Windows we need developers from that area to describe their need...
As for the introduction of docker: I thought about that for a different internal project myself and implemented that. This worked quite well and reduced the setup "costs" for external developers, so I think adding Docker to the repo still has some value
I hope to find some time during the weekend to wrap my head around the dev xp things - on a higher level - to form a "vision".
IMO given a choice, I would prefer to slim down Aruba and try to reduce the dependencies rather than making the build more complex. Having said that, if there are some dependencies we’re agreed we do really need that are fiddly to install then I think having a Docker build is a good idea - anything that makes it easy to get a developer up and running is great. Thanks for the suggestions @e2!
@mattwynne - right now, if I want all tests to be green on my desktop, I have no other choice than to use Docker. There are pending PRs, but juggling them while rebasing over new stuff would mean "unrelated" breakages from other issues. Nothing is worse than failing tests. Also, the build isn't complex - it's just Docker was added as an option, beside the current build. It would also allow Windows users to run Linux tests on their machines - directly using their checked-out sources. (So they can rework Windows-platform specific stuff and immediately test on Ubuntu/Docker if they broke anything on the other side. (All before sending a PR to trigger a Travis build). Technically this allows allows testing things on specific patch versions of a given Ruby. And you can run multiple containers with isolated tests, with different sets of gems. Also, if anyone reproduces a weird issue in a docker image, it's not a problem to upload/download ... and roll up your sleeves to start investigating. And with Docker Compose I'd argue it's actually easier than using Rake tasks for it...
@mattwynne - as for "slimming down", I'd focus on the codebase. It's not like Aruba is suffering for an overwhelming amount of dev tools inside. Meanwhile, Ruby 1.8.7 support is still there - along with lots of other deprecations. There's no use putting effort into slimming anything else down that what's needed to release 1.x. And then e.g. increasing test coverage and other improvements.
@e2 sure - I think Docker gives us some great options, I just wouldn’t want it to become an excuse for absorbing more and more dependencies into the project. For example, I don’t like the fact that there are steps / helper methods that assume the user will have RVM or Bundler installed. I think those should move out intro a
contrib gem.
By the way, one of our biggest users is the RSpec project, who need 1.8.7 support since many of their users still use 1.8.7 and they want to keep supporting them.
@mattwynne I agree that having a docker base image is quite a good idea. I've had some troubles pinning down bugs in aruba with jruby - since I don't use rvm + only the latest mri ruby - so having a Docker image available is very welcome.
@e2 I like the idea to have docker-compose, but is python required to run it?
excuse for absorbing more and more dependencies into the project.
I agree! Besides that, I think we need to find a balance between documentation and dependencies. All problematic dependencies are
- introduced by me (while improving our documentation)
- are part of our documentation (aka features) to demonstrate that aruba can run bash-scripts, zsh-scripts, java-programs, elf-binaries, ...
- are no dependencies of the aruba library itself
I think we can handle this by flagging all scenarios with external deps as @documentation-only or so. Then exclude them in a default cucumber run (cucumber.yml) and run them only in container/CI. I like the idea of having living documentation and don't want to "reduce" - sorry no better idea how to name it - our feature files' value for the project.
I think excluding them is a OK compromise.