Thinking in Elm (and not cutting corners)

So, I’m processing this list of js data which, for this evening, I know
contains exactly one thingummy. I’m I want to wire up some Elm and get things
to type check before I start to worry about handling lists of different sizes.
Calling upon my conventional (python) wisdom, I set about processing the
singleton in the list.

  case msg of
    UpdateThing msg' ->
    (thing, cmd) = Thingummy.update msg'  ... singleton element from model.things ...

How do I get that sucker? I’m thinking python: `model.things[0]` but, of
course, both `model.things !! 1`, and `List.head` return `Maybe a`. So now I’m
going:

    ... (thing, cmd) ... = Maybe.map (Thingummy.update msg') (List.head model.things) ...

But now the lhs pattern is wrong, I’m using `map` and, ultimately, I need to
turn that updated singleton back into a list anyways.

  case msg of
    UpdateThing msg' ->
    (things, cmds) = List.unzip (List.map (Thingummy.update msg') model.things) 

Job done. (I should qualify that the job here is just the quick hack for
my singleton list. It’s applying the same change to all elements)

Moral: it isn’t any less complex to deal with a list of things than
with a thing taken from a list (Maybe). I can’t ask Elm to accept my
assumption that a list will contain exactly one item. Elm wants me to model
it. If there’s exactly one, just use a variable. If there’s zero or one,
use Maybe. Or just use a list, it’s just as simple!

Waiting on the platform

A recent conversation with Lauren Serota reminded me of a challenge we often face building small-to-medium sized digital solutions for our clients: the right platform.

We were talking about building custom phone apps in the context of Myanmar’s smartphone revolution, and the challenge of finding the right partners to implement innovative mobile service delivery systems. She told a story of one such development which looked great but cost too much to make small changes to. That entailed going through another development cycle with the software company; for changes that felt like they should be under the service team’s control.
“They wrote a custom native app.”, she said, “instead of building a platform.”
This started me revisiting similar stories from Aptivate, where we were the software company.

The IT Specialists

I’m thinking about AuthorAID, which helps support researchers from developing countries preparing academic articles for publication in peer-reviewed journals. It is a fairly complex web application which functions not entirely unlike a dating site (matching mentors with mentees) with a way to track collaboration through research publishing tasks. We built it with Django, a powerful framework rapid web development.

Django was a great choice, it enabled us to build AuthorAID at a cost within INASP‘s budget; an impressive achievement because AuthorAID is essentially a fairly complex, bespoke web application serving a relatively small audience. That’s why it’s our web platform of choice. But neither AuthorAID, nor Django are the platform I think Lauren was talking about.

As we maintain and develop AuthorAID for INASP, some of the changes we make, like adding faceted search, feel like they belong squarely with the “IT specialists”. Others, like adding fields to the form that users fill when they register, feel like they belong with the team running the site.

A lot of Django’s power for rapidly building web applications comes from its focus on meta-data. To add those new fields to the sign-up form, we edit the Model file which is written in Python, the programing language Django is written in. So, you have to be a programmer to change them. But once that’s done, changes to web-forms can be generated automatically from the models. The new field values will need to be stored in Django’s relational database, but the necessary changes to the database schema can also be generated from the models, and theres a whole system to manage applying those database changes. Django is a great tool for the “IT specialists”. Which makes sense: it was created by web developers at a Newspaper.

This works out well for INASP who retain a relationship with Aptivate that enables us to act as their “IT specialists”, for the website stuff. But some of the web applications we build aren’t commissioned by organisations like INASP, but as part of change programmes with a limited lifetime and budget. They can’t afford to keep us on as IT specialists, in case their web application needs to be tweaked or extended. If they want to change the data they gather or report on, they need to be able to make those changes themselves. They don’t want a bespoke application built with a rapid-development framework; I think they need the “Platform”.

The General Solution

I’m thinking now of another project that, in the end, we didn’t do. We were written into the budget proposal for a three year programme with institutional funding, to provide a database for monitoring and evaluation (M&E) data. Budget was small and the requirements weren’t very clear until the programme started and a team assembled. We did our discovery process soon found out that our clients had high hopes for their M&E database, but they could predict neither the structure of the data they would be storing, nor the work-flows they would want it to support.

Trying to manage the risk of Building The Wrong Thing, we pushed for specifics specific examples.
“It doesn’t have to be an all singing, all dancing system.”, they told us in one meeting.
“Okay”, I replied, “but you have to tell us exactly which songs, and which dances you will want it to do.
Eventually they told us not to keep trying to understand their specific requirements, and “just deliver the generic version”.

It’s much harder (read more expensive) to build a generic tool than a specific one. It’s very hard to get the abstractions right, especially without passing through specific examples on the way. But our client had had something specific in mind: a low-cost, generic solution. I think I understand where they were coming from: Excel. The generic solution, for fifty bucks, built on billions of dollars of R&D and decades of getting the abstractions right.

Most people have never bought bespoke software, and don’t know how to. If we were one of those people, the closest thing we can imagine to bespoke software might be a customised spreadsheet; with some protected cells, nice colouring, some predefined formulae and some VBA macros behind the scenes. I’m actually working on one of these at the moment, for another client. If I don’t do all the customisation work, what will be left? The full awesome power of Excel. Users would have to work harder to use it but, with the right skills, they could do so. Excel is a cheap product and all that customization effort is expensive IT specialist consultancy. I understand it must feel frustrating When the IT specialists start asking ‘which songs and which dances?’. Just give us The Thing. The thing we can live in and work it out for ourselves. The web-based Excel* for data-driven applications.

* I know this sounds like Google sheets, but they weren’t actually looking for an on-line spreadsheet.

Compared to evolving a customised spreadsheet, building a bespoke web application in, say, Django is a completely different proposition. It takes significant work to get to an MVP and when you get there the only data and workflow it supports are those the product designers have chosen to include. There are no blank workbooks to fall back on.

I’m wondering if there is something awesome we should be using as a platform for web applications. And if there isn’t shouldn’t we be designing it?

Deadlock: BDD testing a python Tornado app with py.test and Splinter

I’m only a day into my new asynchronous web development project in Tornado, and I already have a deadlock. And this is remarkable, I think, because all I have done so far is a copy-and-paste the Hello World example app from the Tornado website. An example which “does not use any of Tornado’s asynchronous features”. So, how did I manage to get deadlock?

Well, on top of Hello World, I have a test. A functional test, written with pytest-bdd that uses pytest-splinter makes sure here is some sort of web server up and running. Now pytest-splinter provides a browser object as a py.test fixture. Tests can use this to access the site, so
long as a server is running. Since I’m shopping for py.test plugins, I added pytest-tornado whose http_server fixture starts a tornado HTTP server, serving the application I provide by defining an app fixture.

All set to go? I thought I was. But I got the big crunch. My splinter browser starts but no page loads. What’s up? Turns out the tests aren’t asynchronous. The splinter driver is blocking and the asynchronous web server never gets a look-in to serve the content. Pytest-tornado wants me
to write tests using Tornado’s asynchronous http client, so it and the server can play nicely together.

That’s all fair, but I want to use Splinter on this project so I’m going to want to start my server in a thread. I created an apprunner module that can start my Tornado app in a thread (and return the url to access it). And also provides a function to instruct the server to stop, and exit the
thread:

import threading
import tornado
from .application import Application

ioloop = tornado.ioloop.IOLoop.instance()

def start(TESTING_PORT=8192):
""" Start Application runing on given port. Retrn the listening url"""
global thread
app = Application()
app.listen(TESTING_PORT)
thread = threading.Thread(target=ioloop.start)
thread.start()
return "http://localhost:{}".format(TESTING_PORT)

def stop():
""" Ask IOLoop to stop running, and wait for its thread to complete"""
global thread
ioloop.add_callback(ioloop.stop)
thread.join()

Then I define my own fixture to start and stop the server using pytest
yield_fixture:

@pytest.yield_fixture
def app():
yield start()
stop()

And now I can write lovely clean BDD steps using splinter browser and my
own app running in a thread, like this:

@when("I visit the site")
def visit_site(browser, app):
browser.visit(app)