Remove fig.packages replace with real deps.

[dnephin]

This change will allow users to install the latest docker-py which adds .dockerignore support.

It is also much easier to manage dependencies with a version, instead of trying to merge in patches from another repo.

If someone needs to run the latest docker-py (as I do), pip supports git urls (http://pip.readthedocs.org/en/latest/reference/pip_install.html#git), so you can set a requirement of the master branch. This also lets you run your against your own fork if the upstream is slow to merge in PRs.

I've verified that script/build-linux, and script/build-osx still works (as well as the test suite). Let me know if there are any other tests I should run.

[dnephin]

I noticed this was wrong. It needs to install requirements.txt first because that has pinned versions, after it can install fig itself and all deps should already be installed.

[aanand]

I'm afraid we've looked at this before and it didn't work out: #264

[dnephin]

Ok, I read over #264, but I don't see any blockers.

The two issues were pyinstaller version of six and docker-py missing a tag.

I've verified that pyinstaller works fine now by running both of the build scripts. Can you still reproduce the error you saw before?

docker-py is well past the versions that was required, and that was never an issue in the first place because you can pin to any git sha.

Managing docker-py this way means that fig users will always be behind in features and will require a lot more work from you to merge in updates and fixes. At the very least you could pin docker-py to your forked version if you have concerns with that project.

[dnephin]

It looks like the travis failure was just a connection failure downloading pyinstaller (the python 2.7 build passed). Could I bother you to re-run that build? Thanks!

[bfirsh]

We don't have a forked version – fig.packages.docker is exactly the same as upstream docker-py. We vendor it for annoying python packaging reasons.

Thanks for the contribution though! Maybe we should add a readme to fig.packages explaining why it exists. /cc @aanand

[dnephin]

@bfirsh I read that argument but I don't understand it. Python packaging has come a long way in the last year, and I work with it daily without any real problems. Is there anything specific that prevents this? Also why would docker-py be such a special case? Why not vendor every dependency? It seems to be working well enough for other dependencies (and basically every other python project).

I understand that docker-py is not a fork, but it's also way behind master. It's currently not possible to get .dockerignore support with fig because of this. Although the short term fix for this is just to upgrade, it doesn't resolve the bigger issue of being stuck with the vendored version. It will always be behind the official release. Vendoring prevents others from using their own fork of docker-py without also maintaining a fork of fig.

Thank you for re-running the travis build.

[bfirsh]

1. We are very closely tied to docker-py so we're always going to pin to a specific version.

2. We often make upstream changes to docker-py and we want to be able to make Fig releases without waiting for docker-py releases

3. We need to use a vendored copy of six for pyinstaller which also means vendoring docker-py

[bfirsh]

Besides – .dockerignore support hasn't even been released yet. If we wanted to include it, we'd have to vendor it!

[dnephin]

@bfirsh No! That's exactly what I'm trying to say, you never need to vendor it, because python packaging is awesome! It looks like I forgot to link to the docs: http://pip.readthedocs.org/en/latest/reference/pip_install.html#git

I created this branch so that I could use .dockerignore

1. You can do this without vendoring! Just pin a git tag, or sha, or a previous release if you really want to
2. You can do this without vendoring! Just pin a git tag or sha
3. I haven't been able to reproduce this issue, but if it's still an issue, you can just pin to the previous version of six! Still don't need vendoring.

Pros for vendoring:

• None, you can do everything with a requirements.txt

Cons for vendoring:

• you're always behind master
• it's more work to upgrade (apply diffs instead of just changing a version number)
• fig is less useful to developers because they can't use a custom version of docker-py directly, and have to wait on upgrades

[bfirsh]

Sorry @dnephin, I appreciate your enthusiasm, but we have spent a lot of time struggling with Python packaging.

Referencing git tags only works with pip, and we can't guarantee that people will install using pip.

The problem is pyinstaller comes with a built-in six package that it injects into the start of PYTHONPATH, so the only way to use a more recent version of six is by vendoring it.

[dnephin]

Have you considered moving away from pyinstaller? From #102 it's pretty clear pyinstaller is causing more than one issue, and there are better tools out there (https://github.com/jordansissel/fpm)

Ok, if six is a blocker, I can add that back, but I still think the docker-py dependency is a win.

[bfirsh]

We have no plans to switch from a static binary, I'm afraid. We did that because we suffered so much pain from Python packaging.

[bfirsh]

We still need to vendor docker-py to be able to reference our vendored version of six.

[dnephin]

I must say that I am disappointed that you have given up on python packaging being a workable solution for managing dependencies in python. All I can say is that I have not had these problems, and that I guess I will have to maintain a fork with this support.

[dnephin]

Re-opening this issue. There is another issue it would solve (now it's both tls and .dockerignore).

I'd also like to re-iterate that I've tested this with scripts/build-linux and it builds correctly. The executable it produces also works just fine without a vendored six. I'm not sure what issue you were hitting before, but it's not an issue anymore.

[bfirsh]

@dnephin it only occurs on specific versions of OS X, if I recall correctly. We can look into removing the vendored package, but we don't have the bandwidth to battle Python packaging at the moment.

If you want to help vendor the current master of docker-py, we'll gladly accept that!

[bfirsh]

To be clear, the reason I'm being stubborn about this is because our primary distribution mechanism is through shipping a binary produced by PyInstaller. It's always going to have fixed dependencies.

[dnephin]

I understand that. But from someone who doesn't use that mechanism, it's really frustrating to work with.

I wish I had an OSX machine to try and reproduce the error, but I'll see if I can find anything about it in the pyinstaller source or issue tracker. This can't be a unique issue with fig.

I also think that, in the long term, real packages (osxpkg, deps, rpms) are going to be quite a bit more reliable (a tool like fpm makes this easier). #102 is a good example of why I think this is the case.

I know you can use dh-virtualenv for debs so that you don't have to worry about dependencies, you can just package up everything as a python virtualenv. I'm not sure if there are equivalents for osxpkg and rpms.

[dnephin]

I see pyinstaller/pyinstaller#114 is merged, but not in an official release yet. Using a git url as a requirement would still work (it is only a dev requirement).

[bfirsh]

Okay, yes, that sounds good.

Now we're part of Docker we actually have the power to ship new versions of
docker-py too. ;)

On Thu, Aug 14, 2014 at 10:57 AM, Daniel Nephin notifications@github.com
wrote:

I see pyinstaller/pyinstaller#114
pyinstaller/pyinstaller#114 is merged, but not
in an official release yet. Using a git url as a requirement would still
work (it is only a dev requirement).

—
Reply to this email directly or view it on GitHub
#375 (comment).

[dnephin]

Pin to current master of pyinstaller, fix setup.py:tests_require and Dockerfile to work with git pin

[bfirsh]

Does this mean we're still pinning versions for the build we make with PyInstaller? We definitely want to pin to specific versions for that.

[dnephin]

Yup, we install pinned requirements from requirements-dev.txt and requirements.txt first, then when it goes to install fig itself, it sees that all the dependencies are already satisfied, so it doesn't install anything else.

[bfirsh]

Hmm. I think in https://github.com/docker/fig/blob/master/script/build-osx venv/bin/pip install . should be venv/bin/pip install -r requirements.txt then. We should fix that.

[dnephin]

Ah yes, you're right. I'll fix that

[bfirsh]

Could we add git to the first apt-get install so the caching works right?

[bfirsh]

These are added separately so they are cached.

[dnephin]

Ah, I didn't realize that worked

[bfirsh]

Yep – a few versions ago, Docker started storing the hashsum of the tarball of an ADD to detect if the contents have changed. It's pretty sweet.

[prologic]

+1

[satyrius]

+1

[dnephin]

I believe I've addressed all open issues here. Let me know if there is anything else.

[bfirsh]

Great. Integration tests pass, Linux build works, OS X build works.

[bfirsh]

LGTM

[aanand]

LGTM

[dnephin]

rebased with master to resolve merge conflicts

[bfirsh]

@dnephin D'oh. Needs another rebase.

[bfirsh]

Tell you what I'll do it for you.

[bfirsh]

Thanks for your help with this @dnephin! Glad we got there in the end :D

[thaJeztah]

Happy to see this merged as well, and a big thank you @dnephin for being persistent 😺