Intermittent failures in tests

[sirnewton01]

There are intermittent failures in TestStatus (status_test.go) that happen rarely on Mac, quite often on Travis CI (disabled using an environment variable for now), sometimes on Plan 9 and often on Dragonfly. UnidiffApply also appears to be failing only on Plan 9 or Dragonfly. These tests should be more robust and work on multiple platforms so that dgit can be verified in a variety of places.

[driusan]

UnidiffApply execs the system patch command, so it's likely that those are legitimate OS specific failures.

[sirnewton01]

If it invokes the system patch command how would it work on Windows?

[driusan]

It wouldn't for now

[sirnewton01]

I wonder if this library could replace the use of OS installed diff/patch.

https://github.com/sergi/go-diff

[driusan]

It might be worth looking into. Diffing isn't hard to do, any reasonable diff algorithm can be written in Go. It's patching that gets difficult, because you need to support any arbitrary patch for apply, not just the ones you generate yourself. Even APE uses a port of the gnu patch tool, so if you don't use the system patch you probably need bug for bug compatibility with it anyways.

[sirnewton01]

Base Plan 9 has the diff tool but I don't see anything that allows you to apply a diff to a file. Maybe I'm missing something obvious.

[driusan]

dgit uses ape/patch and ape/diff on Plan 9. It's defined here: https://github.com/driusan/dgit/blob/master/git/consts_plan9.go

[sirnewton01]

I was thinking that the whole git diff and apply system is just a convenience to standardize the way that git users can pass around patches. It's not a very Unix/Plan 9 way of doing things, but I suppose git is not following the traditional Unix way at all.

Thinking back on the way that Plan 9 development was originally managed before APE I wonder how they passed around patches and applied them. I see that the base system has a diff command (not APE), but I don't see any patch command that isn't APE. There is a patch system in man(1), but that seems to be a system for managing patches. How did one casually give an experimental patch to someone and allow them to apply it?

[driusan]

But using git as a manager for patches is a very git way of doing things, just not the GitHub way. The whole git am and git format-patch commands are dedicated to it, and some of the largest git projects (git itself, and the Linux kernel, for instance) are developed that way rather than using a GitHub PR model.