This post is about automating the provisioning of your local machines via code. That could be your {work, personal} {laptop, desktop, VM}.

It requires an upfront investment but ultimately adds value by reducing time on subsequent configurations of new environments.

In this post I will show you how I achieve this via Ansible and the benefits of this approach.


For motivation, this is how I setup some core applications and preferences automatically when on a new environment:

Installing and personalising git, i3, direnv and pipenv via Ansible.

With a few lines I configured a window manager, git, direnv and pipenv. My actual setup does almost all my entire local provisioning from scratch. But before going further in this example, let’s revisit the problem and the solution in more detail.


As a developer I end up using a number of machines (at home and at work), with various OS/distributions. Examples:

  • personal laptop (Fedora)
  • personal desktop/server (Fedora and Windows)
  • work laptop (Ubuntu)
  • work servers (any)
  • temporary setups for each of the above whenever one breaks
  • temporary VMs (e.g. using Linux in Virtualbox in a Windows laptop)


  • Each of the environments above gets (re)provisioned periodically.
  • Each of these require a number of things set up so that I’m in my most productive environment.
    • git, ssh keys, gpg, direnv, pass, i3 (with my keychain, HiDPI support), Gnome, Chrome, Atom, PyCharm (with the last two needing extensions, personal config, keymap etc). My Python setup (pyenv with various Python versions, pip, pipenv, virtualenv). A virtualenv configured for each of my Python projects.
    • Most of the above benefit from:
      • integration with the bash shell for tab completion and an updated bash prompt
      • personal fields/preferences (e.g. my email address in git, favourite i3 setup).

OS package managers today are highly reliable for the installation part, but work isn’t finished just when a package gets installed. And the periodic manual provisioning of each of the machines/VMs above results in a substantial amount of time. If you have a similar situation, the initial investment in automating the provisioning (as well as the maintenance of it) requires less time than the manual provisioning over a few years time.

Furthermore, sometimes you want to start from a clean image or test big changes on your setup without being concerned with breaking the workstation you rely so much upon.

If having the best setup in your environments isn’t automated, you will compromise. In my tight work schedule I failed for some time to setup things that my personal laptop never failed to have, which accelerate or improve the way I work. You will fix something quickly in one environment but it won’t impact other environments, you might make notes but you’ll forget to update, and you’ll defer problems until they get just enough in your way. And the setup of your environments can get messy.


Backups solve part of this need for a single environment but aren’t ideal to reuse elsewhere (not to mention the difficulties in arranging for a decent full-system backup in Linux laptops).

Some people - like myself months ago - rely on tools like Dropbox for sharing a subset of data in the user’s home such as config folders and files via soft links to the shared Dropbox folders/files, while manually installing system packages. Some problems with this approach:

  • Critical config is not version controlled
    • Did you make a mistake in a config? Hopefully it was within the retention period of Dropbox deleted files and you won’t have to deal with customer support for files/directories not handled properly by their web app.
  • Generates a number of file conflicts between environments.
  • Requires setups and application versions to be very similar across different machines/VMs (generally not the case between work and home environments, or even in machines within each of these).
  • The setup of backups isn’t automated itself within the scope of this solution.
  • You still can’t test changes.

Either you will move almost everything to Dropbox (generating constant syncs for files you wouldn’t normally care about, and did I mention file conflicts?) or it will be a partial move still requiring a manual configuration.

It would be great to abstract what’s common and what’s specific to the OS and share the common bits, facilitating reusability across OS’s (either for switching or parallel use, like work and personal laptops). And then also share reusable common patterns within the community.

Enter Infrastructure as Code

Years ago a new movement started that brought development and infrastructure together to make Infrastructure as Code (or IaaC). Tools were developed to manage fleets of machines via a master machine with inventories and tasks defined in code to do things like install and configure databases, reverse proxies, LDAP, web servers, or any sysadmin task really - making infrastructure finally reproducible and automated.

In the current world of managed services and with a competitive cloud environment, there is rarely now the need for small to medium sized companies and individuals to manage so much infrastructure directly themselves, with tools like Terraform sufficing to just manage cloud services instead.

The improvement of online IDEs and platforms planning to manage the entire SDLC and a wider adoption of machines like Chromebooks might change this, but at least one machine remains on our hands to coordinate everything else - and that is the laptop or desktop in front of you.


One of these IaaC tools is Ansible. Simple yet powerful, written in Python and whose inventories and tasks are written in readable YAML with an extensive range of native modules to change the state of your applications, services, files, networks, etc.

A module for instance is apt, so you can request packages to be present in Ubuntu/Debian - or not, with ansible’s job being to make sure your desired list of packages is in the state you instruct via your code.

A task uses modules, in the example above a task would be:

# if the git package is already installed this action does nothing

# otherwise this installs git

name: git # name of package
state: present # instructs said package to be installed
become: true # ansible needs to run sudo

Ansible modules are generally idempotent. Ansible is faster and easier than doing the equivalent in say a bash script to handle all the setup (although this isn’t apparent from the example above alone).

Roles and Ansible Galaxy

One big promise of Ansible (albeit not matching expectations) was the advent of reusable sets of tasks - aka roles - that could be developed and shared within the community on Ansible Galaxy, not too differently from PyPI in the Python world. Needing only a few user defined variables, a role would take over all aspects related a certain outcome. You’d be able to say download and install a role for setting up say nginx, grafana or Apache in a few commands, without having to code much yourself.

I have made available I few roles I use for myself for git, i3 and others. They are no longer on Ansible Galaxy but you can still find them further below.

Ansible walkthrough

Let’s setup a few of my roles as an example, from which you can build your entire Ansible setup. Note that this will make changes to your system, review the roles actions before starting.

Because pipenv might not yet be installed, we just assume Python 3 is:

python3 -m venv ~/.virtualenvs/n-batalha-roles
. ~/.virtualenvs/n-batalha-roles/bin/activate
pip install -r requirements.txt

We install the roles (they go into ~/.ansible/content):

mazer install -f --namespace n-batalha

Now we define a playbook that uses these roles, defining a few user variables for git. Note that any local setup you have might be changed, be sure to adapt the below to your preferences before running:

- hosts: localhost
  roles: - role: n-batalha.ansible-roles.git
  tags: 'git'
  email: ""
  user_name_long: Your Name Here
  git_settings: - name: core.editor
  value: nano - name: color.ui
  value: auto - name:
  value: "" - name:
  value: "" - name:
  value: checkout - name:
  value: branch - name: alias.c
  value: commit - name: alias.s
  value: status - name: alias.unstage
  value: reset HEAD -- - name: alias.last
  value: log -1 HEAD - name: alias.visual
  value: "!gitg" - name: commit.gpgsign
  value: true - { role: n-batalha.ansible-roles.direnv, tags: 'direnv' } - { role: n-batalha.ansible-roles.i3, tags: 'i3' } - { role: n-batalha.ansible-roles.pipenv, tags: 'pipenv'}

Save it in playbook.yml and execute it:

ansible-playbook playbook.yml --ask-sudo-pass

Now git (with bash completion, git-prompt, git aliases, user and email defined), direnv, i3 (with HiDPI, shortcuts, status bar, and Gnome keyring) and pipenv are installed in your environment.

My personal Playbook includes the above in n-batalha/ansible-roles (reusable roles I’m sharing with the world), a number of external ones (see references below), as well as some less reusable roles.


An advantage of this setup which is easy to overlook is the ability to test system configurations before actually running them on your machine. Either via Docker or Vagrant. I won’t cover this here but you’ll be able to find how this works on my roles project.


If you have more than a single machine and you’re an advanced Linux user, it might be worth to automate its provisioning. Use the above as a starter point, and explore Ansible Galaxy for more reusable roles.

Perhaps I will share later my entire setup and the playbooks to invoke on a ready to use repo for newcomers. Meanwhile I hope this helps.



As with any role, I recommend reading any in full before running.

Mine, used above


Before I was even working in the fields of data and engineering, I have been intrigued as to how bots and automated personal assistants (the likes of Siri, Cortana and Google Now) work.

I now know enough to make a machine learning powered bot from scratch. But actually the best answer I have came later: does it matter?

In this post, I’ll dwelve into this answer and leave the interesting details of bots (the Machine Learning) for a future one.

Bots are not meeting expectations

First things first, let’s just acknowledge that the whole bots revolution has failed to live up to the high expectations many had.

The appeal was understandable for non-experts, on a superficial level. Unlike other machine learning applications, Siri and text bots are as close to everyone’s concept of AI (aka HAL 9000) as possible. Before using them, it was hard not to get excited, all skepticism aside. Even if covering only a fraction of what general AI can cover, it would still be quite useful.

WeChat seemed to be having enormous success in Asia and it relied on a conversational interface that was yet to be applied elsewhere, so it was easy to assume that it would be a matter of time before it did. So the bot gold rush began and by 2015 the hype was still at its highest. Dan Grover, then a Product Manager at WeChat, recalls in his brilliant article:

Conversations, writes WIRED, can do things traditional GUIs can’t. Matt Hartman equates the surge in text-driven apps as a kind of “hidden homescreen”. TechCrunch says “forget apps, now bots take over”. The creator of Fin thinks it’s a new paradigm all apps will move to. Dharmesh Shah wonders whether the rise of conversational UI will be the downfall of designers. (…)

Benedict Evans prophecized that the new lay of the land is “all messaging expands until it includes software.”

In practice, any interaction with most bots would quickly reveal just how far away we are from this reality. Alex Hern writes:

The problems with existing chatbots begin with how they actually work. Almost uniformly, the initial examples of Messenger bots are disastrous: unable to parse any instruction that doesn’t fit their (entirely undocumented) expectations, slow to respond when they are given the correct command, and ultimately useful only for tasks which are trivial to perform through the old apps or websites.

Mashable reports a bird’s eye view of the industry:

The Information reported last week that Facebook is rethinking its development efforts on all automation fronts after research showed bots failed to adequately address 70 percent of customer questions and requests. (…)

A recent Forrester survey found that only four percent of digital business professionals use any sort of chatbot. The research firm concluded in its most recent report on the state of the practice that its hype as a “world-changing technology” has led to a “peak of inflated expectations” that have gone largely unmet.

It seems indeed we have just left the peak and we are halfway down on the trough of disillusionment:

Bots, sliding into the trough of disillusionment.

Whilst one can attribute exagerated expectations alone, these likely resulted from not realizing that:

  • a conversational interface is only one of many interfaces, and it has a much greater friction than a GUI in many situations (regardless of the bots intelligence)
  • limitations in Natural Language Processing further constrain the number of successful conversational interfaces we can develop. For now, they need to be simple and limited in scope to be accurate (more on this on the next post).

To these fundamental issues, some temporary ones might have worsened the situation:

  • lack of experience in the field (design practices are/were not as established)
  • a low barrier of entry that encourages numerous poor submissions

Whilst improvements on NLP are hard to predict, an answer to the first fundamental issue is just reaching the market. Or better yet, we are back to the starting point.


A pure conversational interface doesn’t suffice, but combine it with a more complex GUI and this hybrid model might just be good enough.

While WeChat is popular and supports bots, it also supports web views mixed with a conversational model to provide a much better user experience. And that is claimed to be one of the key reasons of its success, not the conversational aspect.

Dan Grover compares the experience of ordering a pizza in Microsoft’s demo bot versus Pizza Hut’s account on WeChat. In WeChat the user starts in Pizza Hut’s account (a bot), but the interface really becomes a web like interface.

Microsoft's pizza ordering bot.
Pizza Hut ordering on WeChat.

73 vs 16 taps. And note how in WeChat’s version the user is provided a wealth of information on what’s available, whereas in the previous the user is left in the dark. Layer is taking one step further, where each message has the potential to be a mini application.

Perhaps more importantly, the most understated advantages of the WeChat model have less to do with chat than with the fact that users can immediately access a number of services in a central platform, not requiring installation of separate apps, accounts to be made and logged in, payment information to be entered, and more friends lists to manage.

Apple/Android Pay in mobile browsers, the evergrowing Facebook walled garden, if not improvements at OS level, might all evolve to provide the frictionless experience of WeChat, bots not required.

As Dan puts it:

This maybe a bit disheartening to hear since creating bots powered by AI sounds super cool and cutting edge, while making mobile optimized websites definitely does not.


Between starting this review a few years ago, and finally doing a post, there were enormous changes in the field. The most significant might just be the development ecosystem that sprung up and makes writing simple bots almost trivial.

From not requiring development skills, to doing only the machine learning bits (open-source or not), some service appears to cover it. See for instance Rasa (open-source), Wit.AI and Microsoft’s Bot Framework.

I won’t list all the good services I found just yet, but other people have covered most of them:


TL;DR: text interfaces are not enough, these serve a niche purpose but they can be combined with other interfaces providing more possibilities for developers to apply where appropriate. WeChat’s success was not likely due to the conversational aspect alone, but perhaps in great part due to their frictionless user experience to reach out to services (rather than needing separate apps with separate accounts and onboardings). Also, you don’t really need Machine Learning skills anymore to make bots.

Still, aren’t you curious as to how Siri finds the weather? Now that I’ve made clear it’s not that important, I’ll soon make a post about it, guilt free.


If introductions are needed, Flask is a great Python microframework (which can also be seen as a “framework” of frameworks). With it, one can create a simple API or web application in no time.

I most recently started using it for a personal project and at work, having been using Django exclusively for over 2 years. “Great” needs to be defined better of course, where it excels or would be better replaced by Django or others.

Since starting this path, I made similar findings (the hard way) to what was recently published by Tony Narlock in this comprehensive comparison of Flask and Django.

Soon after going through the introductory docs and writing a first simple app, I wanted to add tests to it. And this is where one finds one of the main differences reported above:

Flask developers will be forced to reinvent the wheel to catch up for things that’d be quick wins with Django.


One popular measure of the quality of a codebase is the test coverage of such a codebase. Some go as far as assuming that “no unit tests mean that the code does not work and cannot be expanded upon”. Not everyone likes to write tests, but not doing comprehensive tests is not an option. As such, tests should be easy to write. They should also be fast. In general, they should not introduce friction or get in your way.

As Patrick covers in his post:

I believe that you should create an environment for developing unit test that is both easy to use and fun to use. These may sound like very fluffy words, but they have meaning for me:

  • create helper methods to allow quick generation of unit tests
  • provide quick feedback on progress of unit testing
  • create an environment that allows for tests to be added quickly
  • make sure that the unit tests can be executed with a single command
  • provide visual feedback of path coverage

As with so many things in software development, if it’s easy to use, it will be used. If it’s tough to use, it won’t be used as much as it should be.

So how do we get this in Flask, if we want to use a relational database?

I won’t cover all of this, but only what concerns Flask and SQL-Alchemy, but the others are easier. Several articles and gists have been written regarding this section I’m covering, but of those that I found, many are outdated, have suggestions I disagree with (you shouldn’t use sqlite for tests if you run a different database in production) or cover specific subtopics separately, so I wrote yet another post.

Flask and SQLAlchemy

If you haven’t heard about SQL-Alchemy and you’re starting in Flask, read this. Django has a ORM, but in Flask it’s a separate framework.

In Django one might take for granted some of the nice features that run in the background. In testing, one is:

Tests that require a database (namely, model tests) will not use your “real” (production) database. Separate, blank databases are created for the tests. Regardless of whether the tests pass or fail, the test databases are destroyed when all the tests have been executed.

No such thing exists in the Flask side. There is of course a compromise between having very lightweight and specialized frameworks or “batteries included” frameworks, opiniated or not. But perhaps extensions here can be improved to save some of the boilerplate. In hindsight such boilerplate is small and I’ll easily port it to new projects, but it was the result of time consuming API reviewing, testing and tinkering.

App config

In Flask, we take care of the details of starting the app. A recommended path is to use a factory for apps. This factory takes as argument a config, for instance:

def create_app(config_object):
app = Flask(**name**.split('.')[0])

Then somewhere in your or wherever you start your app object, you detect the environment and load the respective config object. The tests would create a separate app, with the config that is best suited for tests only (although the less you deviate from production, the better).

One of the extensions you’ll want to register above is Flask-SQLAlchemy as it takes care of the boilerplate integration bit between Flask and SQLAlchemy.

With the Flask-SQLAlchemy extension registered, in your config, you’ll need to have specific variables such as SQLALCHEMY_DATABASE_URI, to point to the database each environment should be using.

To have tests running as Django though, we need to have a different database just for tests. This could be solved by having two databases per environment (one being for the tests only). But again in Django “separate, blank databases are created for the tests (…) [and] destroyed when all the tests have been executed..

For this I found testing.postgresql, which can be used to set SQLALCHEMY_DATABASE_URI at test runtime:

from testing.postgresql import Postgresql

\_app = create_app(TestConfig) # TestConfig is defined further down
with Postgresql() as postgresql:
\_app.config['SQLALCHEMY_DATABASE_URI'] = postgresql.url()

We’re not done yet though, as we’re missing an equivalent of TestCase, something to allow writing tests easily.


I made the jump to Flask and PyTest at the same time. I’m not yet sure whether it’s better than unittest.

Instead of classes, in PyTest we make extensive use of the advanced fixtures mechanism. Fixtures here take a more general meaning than usual, it’s anything consumed by a test. The test app will be a fixture, along with the database.

One thing we don’t want though is to generate a database per test, as that would render tests very slow. For this, PyTest allows us to define a scope per fixture. Scopes can be function and session for instance - meaning that the fixture will be used and run per function or session respectively.

However if we scope the database fixture per session, how to do we make sure that the database is clean and at the same state when the different tests are run? At the end of each test, we rollback any changes made to the database during the test.

Putting it all together in code, you want a similar to:

from testing.postgresql import Postgresql

from import create_app
from import db as \_db

class TestConfig(object):
DEBUG = True
ENV = 'test'

def app():
\_app = create_app(TestConfig)
with Postgresql() as postgresql:
\_app.config['SQLALCHEMY_DATABASE_URI'] = postgresql.url()
ctx = \_app.app_context()

        yield _app


def testapp(app):
return app.test_client()

def db(app):
\ = app

    yield _db


@pytest.fixture(scope='function', autouse=True)
def session(db):
connection = db.engine.connect()
transaction = connection.begin()

    options = dict(bind=connection, binds={})
    session_ = db.create_scoped_session(options=options)

    db.session = session_

    yield session_


where yourapp\app\ has

from flask_sqlalchemy import SQLAlchemy

db = SQLAlchemy()

and create_app is as above.

Now I went a little further and set the session fixture at autouse, but you might not want that. If True it’s automatically used per each test, otherwise you need to invoke it each test function by adding session as an argument. conftest needs to be set as a parent to all the test files that would use it, and PyTest takes care of discoverability.

Now let’s say you have a defining a simple Request table:

class Request(SurrogatePK, Model):
**tablename** = 'request'
request = Column(db.JSON)

Somewhere in your app you save in this table each POST request that is received. You can now test this with:

def test_save_request(testapp):
request_data = # some example request params
resp ='/', content_type='application/json', data=request_data)

    # the test should also filter by the request_data content, this is a simplified example
    assert Request.query.filter_by(method='POST').count() == 1

And what is SurrogatePK you ask? This is taken from sloria/cookiecutter-flask where you can take on more good ideas on how to structure your Flask project.

If I find more time, I’ll try and write an extension of this cookiecutter to include the setup above, more helpers and the goals Patrick covered above, meanwhile I hope this helps!

Amazon Echo

If you haven’t heard about Echo, it’s a new device that can be described as Siri in your living room, but actually getting used. And it does just that, listening to your voice and handling simple queries.

Amazon Echo, causally listening to your entire life.

It’s supposed to have a seven-piece microphone array (that your phone doesn’t have) making it more accurate in the speech recognition, it’s always on and listening (making the interaction more frictionless) and has a great speaker. And perhaps being in the comfort of your own home - where talking to machines is somewhat less awkward - is another big reason behind it’s popularity.

Infographic: What the Amazon Echo Is Actually Used For | Statista
Amazon Echo Usage - Statista.

Despite not having a great experience with it, I couldn’t resist the fun of developing for it. For now, I’m hoping that my unusual accent is the reason why we don’t get along that great, and next generations will improve.

Alexa skills

Also an advantage over Siri (although that’s changing), it that it has its own app store. Apps here are called “skills”. But you don’t need machine learning expertise to develop most skills, Amazon does the heavylifting for you.

The way it works is:

  1. the user will speak to an Alexa device (like Amazon Echo)
  2. Amazon does the speech recognition, intent classification and entity extraction, and calls our service while providing the processed speech (eg. the user says “what’s the weather in London”, and the service gets {'intent': 'KnowWeather', 'entities': {'city': 'London'}
  3. the skill will receive the nicely parsed data, and do its logic (eg. query a weather API for London) and send a response via an HTTP POST request to Amazon
  4. Echo converts the skill’s text response to speech.

Want to code an app for it in Python? Read:

After going through a number of issues getting an Alexa skill up and running properly in Python via Kubernetes, I thought I’d abstract the core setup in a cookie cutter. To get started:


And continue with these instructions. It should take you from nothing to a working app in a real life Echo device in ~30 minutes.

The included skill is barebones, if that. For a small number of hardcoded ingredients, it provides an example replacement if you ask for it. Alexa, ask cook bot what does replace lemon?


What is not included, as of January 2017, is:

  • proper logging (for exceptions and performance monitoring): hopefully someone will later do a Helm chart for Sentry and others to easily deploy these
  • autoscaling, for both pods and the nodes
  • some patterns of dialogues along with more examples of the API (with tests)
  • CI/CD
  • test environments
  • integrated analytics
  • local Kubernetes tests (currently it tests the app itself, but everything up from gunicorn is untested)