Developing a RESTful micro service in Python


This is the tale of why and how I re-wrote an ageing Java-based customer & order management system in Python and how the same feature set was implemented using only one 10th of the code base.

If it ain't broke, don't fix it

During Christmas in the year of our Lord 2007, I wrote a simple customer and order management system using what I found to be cool technology at the time: Java 6, iBatis (persistence framework), Tomcat 6 (application server), MySQL (database), Debian Linux (operating system) and (deep breath) JSP (templating),

After 1-2 weeks of coding, I had something that was usable and, after some initial bug fixing after going live in January, it ran happily without any maintenance or supervision of any kind for 7 years. The only time it wasn't available, was when there was a power outage in the server room.

The customer was happy and the story could have ended here. But something made the system to be completely re-written after 7 years of blissful existence. What was it?

Change. The customer wanted new features. Which meant picking up the development and, lo and behold, I discovered how outdated the system had gotten when looking at it again through the goggles of 2014. The architecture was pure RPC over HTTP, JSPs was such an old technology it's wasn't even fun thinking about it, MySQL had gone from being a hip and cool Open Source database to being a crippled step child of Oracle, but worst of all: some of the technology, most notably iBATIS, wasn't supported anymore. Not supported meant no security updates, outdated documentation, a shrinking user base and last but not least: the sad feeling of missing out on all the fun stuff happening in the world.

The obvious choice

There was a fork of iBATIS available called MyBatis but using it meant making changes to both my code and build setup, so I started looking somewhere else. Having used Java as my main programming language at work for more than 10 years, the natural choice seemed to be the now nicely standardised and usable JPA for persistence with the rest of the JEE7 stack for the other building blocks: Java 7, JAX-RS (REST framework) and JSF, Rich Faces or Wicket (templating).

After a month or two of hacking away at this evenings and weekends, I had something that worked and was Java wise nicely layered, object oriented, de-coupled, unit testable and so on. It was just not fun.


And programming should be fun. Especially when working for free on open source projects like this system. So I asked myself, what would I prefer programming in? In which language and on what platform would I have the most fun programming in while at the same time being able to quickly add new features? Although I'm a huge fan of BASH, I did admit that it wasn't ideal for developing a rich web based order management system. My choice was then easy, it had to be Python.

I've always been very fond of Python ever since learning it in university but I quickly moved away from it in favour of BASH for writing command line programs as BASH is always available on all kinds of UNIX and Linux variants (exception being HP-UX). For larger applications, my work has always been Java related, so Python has continued to sit quietly in my toolbox, waiting for the task at hand being right for it. And now it finally was.

What a joy

And what a wonderful relief it was to start programming in Python! Just like coming home. I like programming in Python so much because its programming model maps so well with the way I think. There's so many thing that just feels right. Natural.


Another nice welcome was to find how excellent support Emacs has for Python. There's many Python plugins to choose from. I settled for anaconda-mode, which together with pyflake and other great plugins I already use, like flymake, auto-complete-mode and projectile, give me everything I could ever wish for when coding: auto completion, on the fly syntax checking, code navigation, interactive shell for prototyping and documentation lookup.

And there was more to be rejoice about. Lots. Like Flask. It's a lightweight web framework which is just wonderful to work with. It's blissfully free of bloat or half thought out ideas, incomplete documentation or patchy libraries that haunt so many other frameworks. You know, the frameworks which let's you easily do mundane tasks, but don't scale up to world applications. Your applications. I cannot recommend you enough to try out Flask. It's strikes an impressive balance between simplicity and feature richness.

Another affable acquaintance was that of Jinja. It's templating done right. Again striking a good balance between being easy to use and having all the features you need to cover all your project needs. As an example of how well behaved Jinja is, take a look at this error message after I did something illegal in Jinja:

{% block head-title %}

And Jinja2 told me in this super useful way: TemplateSyntaxError: Block names in Jinja have to be valid Python identifiers and may not contain hyphens, use an underscore instead.

Man, I wish other API and framwork authors would take note.

Another amazing piece of the puzzle was the Werkzeug web server which Flask bundles. It's a lightweight server, making Flask a perfect micro service platform. If you think DropWizard for Java is nice and easy, try out Flask. It's a ten times easier to use, configure and start. This is all you need to do:

  # apt-get install python-pip
  # pip install flask

Then, add a REST endpoint, instantiate and run Flask:

#! /usr/bin/env python
from flask import Flask
app = Flask(__name__)

def get_search():
    return render_template("search.html")

if __name__ == '__main__':

With this in place, you can start your micro service in pure UNIX fashion with:

  $ ./

And the server was up! That's it. Nothing more. Now, go back to your Maven based, Java and DropWizard project:

  1. Add DropWizard dependency/ies to your POM
  2. Create you DropWizard YAML configuration file
  3. Create a Java application class to bootstrap it all.
  4. Add whatever REST end points you want (the actual fun bit).
  5. Run Maven to compile your code and package it all up with mvn package.
  6. Start your application, e.g. with: java -jar target/application-0.1-SNAPSHOT.jar server path/to/app.yml

And this is something of the easiest you get in the java world! Now, tell me you're not already feeling homesick and want to go back to using Flask. I for sure am!

Keep it really simple, stupid

Whenever I'm discussing with fellow Java developers, we all agree that we should make things as simple as possible. However, the moment we get back to our IDEs after the coffee break, we continue creating our super intricate, object oriented, multi layered solutions with an indefinite number of indirections.

This time, getting a clean slate from a new language and software stack, I set out at the other end: what's the minimum number of indirections, layers and objects I can get away with while keeping the code reasonably loosely coupled, easy to read, maintain and extend?

One decision I made, was to work directly on JSON structures instead of transforming HTML forms to domain objects and then translating these into database tables and rows again. Since the MySQL Python driver has a cursor which delivers a Python dictionary, which is practically an immutable JSON structure, I was almost home free. And when discovering that the HTML form data from the web client also came in JSON wrapping, I was good to go.

"But what about type safety?" I hear you cry. Well, I did a few issues with this, but surprisingly few. Much of this is because Python is really good at doing "what you mean". It tries to be smart and understanding and most of the time it gets it right. If you have a date time field in the database and you instert a string with '2015-04-11', it'll happily convert it to a date time object. And so on. So far, I've spent about two hours on type (conversion) issues. Not bad, and definitely no more than I would have had if I'd enforced entity objects with types. Because that's the funny thing: even with Java, JPA, Hibernate, Joda, Jadira and all that jazz, you still venture into type problems. There's no getting away from that you need to be in command of your application, your stack, no matter the technology you use.

So I ditched the domain objects and just used JSON, very closely resembling the database tables. Another thing I did, was to just have two layers on the Python side: the data layer and the REST layer. Nothing in between. I actually wanted to add a middle layer, but I stalled at the last minute because I saw just how easy it was to understand and debug the code when having just two layers.

My architecture thus became:

view:  <Jinja HTML templates with Python objects>
model: <Python code with Flask REST routing>
data:  <Python code with MySQL connectivity>

Less than 1000 lines of Python code. It's easy to understand (yes, I know I wrote it, but still), it's easy to debug and it's easy to extend.

If the system is to grow, however, I do see some challenges with the current structure. I would need to introduce some delegation in the two Python layers as the files shouldn't grow significantly more now (main Flask file is ~400 lines, main data file ~500 lines). Of course, I've crated own Python modules for the Jinja filters, data handling, SQL generation and configuration file parsing. Still cramming everything in ~900 lines of codes.

To be clear: My point is not that complexity is a bad thing in itself, it' just that very often in the Java world (and I'm sure in other sub cultures as well) we tend to introduce great complexity up front, something which we seldom need. Be a great master of complexity, but be equally afraid of introducing it.

So how did you do...?

Unicode support all through the stack

Database: Set the character set on the database connection:

import MySQLdb as mdb
con = mdb.connect(self.db_host, self.db_user, self.db_password, self.db)

Date, currency formats and the like: Set the encoding together with the desired locale using the standard locale module, just as you would do on the UNIX command line:

from locale import setlocale
setlocale(locale.LC_ALL, "nb_NO.utf8")

Make sure the HTML render characters using UTF-8: Add this to the HTML head. Typically this will be in a Jinja HTML template that all templates inherits:

<meta http-equiv="Content-Type" content="text/html; charset=utf-8"/>

Create a web service accepting multiple HTTP methods

Here, I have a REST service accepting two HTTP methods and takes a parameter from the URI template defined, passing it to the method, picking up the values from the HTML form, saves it to the DB and performs a redirect to another method performing a GET of the now updated customer.

@app.route("/customer/<id>", methods=["POST", "PUT"])
def update_customer(id):
    return redirect(url_for("get_customer", id = id, updated = True))

Database transactions

Making several database queries run in the same transactions and rollback is so easy and elegantly done in Python, I thought it wasn't there! As often before, Python just works the way you want it to, taking care of you and cleaning up the mess without whining.

Here, I delete all related order items before deleting the order itself. All in the same transaction and using prepared statements to avoid SQL injection:

def delete_order(self, id):
    con = self.get_db_connection()
    with con:
        cur = con.cursor()
        cur.execute("delete from order_item where order_id = %s", (id))
        cur.execute("delete from customer_order where id = %s", (id))

It's so neat and tidy, you'd be forgiven to think you'd missed something.

Input validation

HTML5 has matured so much now that I delegated most of the client input validation to it (<input type="email"/> is great!).

I also let the browser's HTML5 capabilities provide the data pickers. At the time of writing (2015-04-11), this means that only users of Google Chrome and Opera get the date pickers, the rest must type in the dates using good old ASCII. In both cases, HTML5 takes care of the validation. Wonderful!

Consistent, professional layout of web pages

I've used Twitter's Bootstrap. Once I got my head around its grid system and form layout, it was nice to have some pre-defined, well tested layout system that works on resolutions and devices without too much hassle.

Was it all bad, the old stuff?

No, not at all. When re-writing all of the Java and JSPs in Python and Jinja, I did keep the database as it was. It's not just out of convenience, I really liked the database modelling and kept it the way it was. It's proof, if you needed one, that it's good to keep away from mixing application code with the database, like PL/SQL.

The operating system is also something I kept. Debian is rock solid and the most wonderful Linux distribution as far as I'm concerned.

As I mentioned in the introduction, MySQL has become somewhat of a step child in Oracle's helm. I don't know what to make of it. What does Oracle want with MySQL? Hence, I always pick one of the MySQL forks or patch sets if you will. For many years now, my favourite has been Percona. Apart from being a different distribution, it's basically MySQL, so all the SQL scripts and data could be kept without modification.

And of course, having implemented this system once before was a huge benefit. Domain knowledge is crucial for developing good software and all the user feedback from the old system helped me creating a good new version.


Together, Python, Flask, Jinja and Emacs made this such a smooth experience, bringing back the fun in programming while at the same time implementing all the features of the old system in one tenth of the code: ~900 lines of Python versus ~9000 lines of Java code, excluding templating, HTML, JS and CSS.

I publish this in the hope that more people will get inspired to try out a new stack when diving into their next big project as well as taking a step back and reconsidering all the industry standard, best practise and modelling principles when tackling "enterprise" challenges . Do you really need all the levels of abstractions? All the objects? All the type safety? All the frameworks? What does it really give you? How many bugs do you actually avoid by adding all these layers of complexity and indirections? And also the other way around: how many bugs do you think this complexity have introduced over the last 5 years in your system? I'm just asking 😊

Useful articles

Licensed under CC BY Creative Commons License ~ ✉ torstein.k.johansen @ gmail ~ 🐘 ~ 🐦 @torsteinkrause