Monday, October 31, 2016

Carousel Cotton Candy

Version 0.3

I'm super excited to announce the Cotton Candy release of Carousel, version 0.3 on PyPI and GitHub. There were a few more issues I really wanted to close with this release, but I decided to push it forward anyway. So the remaining milestones for v0.3 will get pushed to v0.3.1.

  • issue #62 use Meta class for all layers - currently usage is spotty and inconsistent. I wanted to keep the number of commits to close PR #68 to a minimum (following the winning workflow) so I only implemented Meta classes where I had to. In fact it's not even implemented in the DataSource example below.
  • issue #25 move all folders into project package - this is already how I have set up the PVPower demo. It just makes more sense with new style models to have them all in the same package.
  • issue #63 and issue #22 split calculations into separate parameters - I knew this couldn't be done by v0.3, it was a stretch goal, but I'm super excited about this. By moving dependencies to an attribute of each parameter, the DAG shows which calculations are orthogonal, so we can run them simultaneously. According to issue #22, I had this idea already, but when I saw this presentation at PyData SF 2016 on Airflow by Matt Davis I was even more motivated to make it happen.
  • issue #59 and issue #73 which don't seem like they're relevant, but they both have to do with implementing a Calculator class whose job it is to crunch through the calculations, somewhat similar to what DataReaders and FormulaImporter do for their layers. Then the boiler plate uncertainty propagation code in the static class could be applied to any calculator such as a dynamic calculator, a linear system solver for an acyclic DAG of linear equations or a non-linear system solver for an acyclic DAG of non-linear equations.

The Parameter class

What's new in Carousel-0.3 (Cotton Candy)? The biggest difference is the introduction of the Parameter class which is now used to specify parameters for data, formulas, outputs, calculations and simulation settings. For example, previously data parameters would be entered as a dictionary of attributes.

Bicycle Bears

class PVPowerData(DataSource):
    """
    Data sources for PV Power demo.
    """
    data_reader = ArgumentReader
    latitude = {"units": "degrees", "uncertainty": 1.0}
    longitude = {"units": "degrees", "uncertainty": 1.0}
    elevation = {"units": "meters", "uncertainty": 1.0}

Cotton Candy

class PVPowerData(DataSource):
    """
    Data sources for PV Power demo.
    """
    latitude = DataParameter(units="degrees", uncertainty=1.0)
    longitude = DataParameter(units="degrees", uncertainty=1.0)
    elevation = DataParameter(units="meters", uncertainty=1.0)

    class Meta:
        data_reader = ArgumentReader

Why the change? The Bicycle Bear version did not have any way to distinguish parameters, like latitude from attributes of the DataSource like data_reader. This had two unfortunate side-effects:

  • each layer attribute had to be hardcoded in the base metaclass so that they wouldn't be misinterpreted as parameters
  • and users could not define any custom class attributes, because they would be misinterpreted as parameters and stripped from the class by the metaclass.

The Cotton Candy version makes it easy for the metaclass to determine which class attributes are parameters, which are attributes of the layer and then leaves everything else alone. Every layer now has a corresponding Parameter subclass which also defines some base attributes corresponding to that layer. Any extra parameter attributes are saved in extras. Attributes that apply to the entire layer are now specified in the `Meta` class attribute, similar to Django, Marshmallow and DRF. The similarities are completely intentional as I have been strongly inspired by those project code bases. Unfortunately, the Meta class is only partially implemented, but will be the major focus of v0.3.1.

Separated Formulas

The Formula class is also improved. Now each formulas is a Parameter with attributes, rather than a giant dictionary. This improvement is still on the roadmap for the Calculation class. As I said above, it was a stretch goal for this release.

Wednesday, October 12, 2016

Winning Workflow

Intro

There are many blog posts on the topic of effective Git workflows, SO questions and answers, BitBucket tutorials and GitHub guides and an article in the BBC. So why another post on git workflow? None of these workflows seemed right for us, but recently it's just clicked, and I feel like we've finally found the process that works for us. The key was finding the simplest workflow that included the most valuable best practices. In particular, we found that complicated multi-branch strategies were unnecessary, but test driven development (TDD) and continuous integration (CI) were a must.

Winning Workflow


Setting up Remotes

We start with the assumption that all of collaborators fork the upstream repository to their personal profile. Then each person clones their profile to their laptop as origin and adds another remote pointing to the upstream repository. For convenience, they may also create remotes to the forks of their most frequent collaborators.

[myusername@mycomputer ~/Projects]
$ git clone git@github.com:myusername/myrepo.git
[myusername@mycomputer ~/Projects]
$ cd myrepo
[myusername@mycomputer ~/Projects/myrepo]
$ git remote add upstream git@github.com:mycompany/myrepo.git
[myusername@mycomputer ~/Projects/myrepo]
$ git remote add mycollaborator git@github.com:mycollaborator/myrepo.git
[myusername@mycomputer ~/Projects/myrepo]
$ git remote show
  origin
  upstream
  mycollaborator

Ground Rules

The next assumption is that we all keep our version of master synchronized with upstream master. And we never work out of our own master branch! Basically this means at the start of any new work we do the following:

  1. I like to do git fetch --all to get the lay of the land. This combined with
    git log --all --graph --date=short --pretty=format:"%ad %h %s%d [%an]"
    let's me know what everyone is working on, assuming that I've made remotes to their forks.
  2. Then I pull from upstream master to get the latest nightly or release,
  3. and push to origin master to keep my fork current.

Recommended Project Layout

I'm also going to assume that everyone is following the recommended project layout. This means that their project has all dependencies listed in requirements.txt, is developed and deployed in its own virtual environment, includes testing and documentation that aims for >80% coverage, has a boilerplate design that allows testing, documentation and package data to be bundled into a distribution and enables use with a test runner with self discovery, and is written with docstrings for autodocumentation. Nothing is ever perfect, so being diligent of path clashes, aware of the arcana of Mac OS X1 or Windows2 and able to use Stack Overflow to find answers is still important.

Branching, Testing, Pull Requests and Collaboration

  1. Now I switch to a new feature branch with a meaningful name - I'll delete this branch everywhere later so it can be verbose.
  2. The very first code I write is a test or two that demonstrates more or less exactly what we want the feature or bug fix to do. This is one of the most valuable steps because it clearly defines the acceptance criteria. Although it's also important to be thoughtful and flexible - just because your tests pass doesn't necessarily mean the feature is implemented as intended. Some new tests or adjustments may be needed along the way.
  3. Now, before I write any more code, is when I submit a pull request (PR) from my fork's feature branch to upstream/master. So many people are surprised by this. Many collaborators have told me they thought that PR's should be submitted after their work is complete and passing all tests. But in my opinion that defeats the entire point of collaborating on a short iteration cycle.
    • If you wait until the end to submit your work you risk diverging from the feature's intended goals especially if the feature's requirements shift or you've misinterpreted the goals even slightly.
    • Waiting also means you're missing out on collaborating with your teammates and soliciting their feedback mid-project.
    On the other hand, by submitting your PR right after you write your tests means:
    • Every push to your fork will trigger a build that runs your tests.
    • Your teammates will get continuous updates so they can monitor your progress in real-time but also on their time so you won't have to hold a formal review, since collaborators can review your work anytime as the commits will all be queued in the PR.
    I think the reason people wait until the end to submit PR's is the same reason they like to write tests at the end. I used to hate seeing my tests fail because it made me feel like I was failing. I think people delay submitting their PR's because they're nervous about having incomplete work reviewed out of context and receiving unfair criticism or harsh judgment. IMO, punitive behavior is dysfunctional and a collaboration killer and should be rooted out with a frank discussion about what mutual success looks like. I also think some people aren't natural collaborators and don't want other's interfering with their work. Again, a constructive discussion can help promote new habits, although don't expect people to change overnight. You can take a hard stance on punitive behavior but you can't expect an introvert to feel comfortable sharing themselves freely without some accommodations.
  4. Now comes the really fun part. We hack and collaborate until the tests all pass. But we don't have too much fun - there should be at most 10 commits before we realize we've embarked on an epic that needs to be re-organized, otherwise the PR will become difficult to merge. That will sap moral and waste time. So keep it simple.
  5. The final bit of collaboration is the code review and merging the PR into upstream master. This is fairly easy, since there are
    • already tests that demonstrate what the code should do,
    • only a few commits,
    • and all of the collaborators have been following the commits as they've been queuing in the PR.
    So really the review and merge is a sanity check. Do these tests really demonstrate the feature as intended? Anything else major would have stood out already.
  6. Whoever the repository owner or maintainer is should add the tag and push it to upstream. This triggers the CI to test, build and deploy a new release.

Continuous Integration

This is key. Set up Travis, Circle, AppVeyor or Jenkins on upstream master to test and build every commit, every commit to an open PR and to deploy on every tag. Easy!

Wrapping Up

There are some features of this style that stand out:

  • There is only one master branch. Using CI to deploy only on tags eliminates our need for a dev or staging branch because any commits on master not tagged are the equivalent of the bleeding edge.
  • This method depends heavily on an online hosted Git repo like GitHub or BitBucket, use of TDD, strong collaboration and a CI server like Travis.

Happy Coding!


footnotes

  1. On Mac OS X matplotlib will not work in a virtual environment unless a framework interpreter is used. The easiest way to do this is to run python as PYTHONHOME=/home/you/path/to/project/venv/ python instead of using source venv/bin/activate.
  2. On Windows pip often creates an executable for scripts that is bound to the Python interpreter it was installed with. If the virtual environments was created with system site packages or if the package is not installed in the virtual environment then you may get a confusing path clash. For example running the nosetests script will use your system Python and therefore the Python path will not include your virtual environment. The solution is to never use system site packages and install all dependencies directly in your virtual environment.
Fork me on GitHub