Justus Perlwitz

Designing a Better Web Analytics Platform

Not many things a business owner has to go through are as scrutinous and unpleasant as a data privacy audit. In the case of Germany, this will include a state commissioner for data protection putting your companies data processing and storage practices under the microscope. Luckily, as it turns out, most of the time you’ll already be doing the right thing. Salting user passwords, allowing customers to delete their account, and allowing users to request all personal data that is being stored about them should be business as usual for most of us.

But there is one thing that will always be a problem in a data privacy audit: The common reliance on US-American service providers for cloud services, data storage, and web analytics. This includes Google’s famous product Google Analytics.

What does Google Analytics do? It allows you to pervasively track your website and app users, understand what they like, how they convert, and what their business value to you is.

How is this problematic? Using Google Analytics in Germany and other EU countries is legally questionable. I was personally told by a city official last year that in order to comply with German privacy laws, you need to close a legal agreement with Google Ireland that will guarantee you processing and storage of user data in conformity with the law. There exist countless templates for this and you can usually just mail this over to Google Ireland and get it counter-signed. But that’s not the end of the story.

At the end of the day, when you really think about it, you are handing over your user’s precious trust to a faceless multinational corporation that has a well-documented track record of not caring too much about their user’s privacy or third party information that they have been entrusted with. And if you ever want to migrate away from their platform, there is no easy way to retrieve all your customer data and move it someplace else. So clearly, choosing Google Analytics is a convenient thing to do, but it will really only get you to your goal half-way.

Another big limitation is of course the quality of results when deploying Google Analytics on high-traffic pages. Sampling will start at around 500,000 records, meaning that when you are investigating small user segments, you will not always get the full picture. More about their approach to sampling is described in this article by Moz.

I have spent the last few weeks trying to understand all the alternative solutions to web analytics out there on the market, both proprietary as well as open source. One of the most prominent alternatives is of course good old Matomo (formerly Piwik) that you can self-host. It is very unfortunate though that Matomo is hard to deploy outside of LAMP stack environments, meaning it runs on PHP, assumes you have a normal HTTP server like Apache running, and really does not like to be dockerized.

It is also completely tied to a MySQL database, which surprises me. Given the size of the project I would have expected Matomo to have a flexible storage backend that can use any available SQL database. Since I am a self-proclaimed PostgreSQL fanboy, I would of course like for that database to be PostgreSQL. Furthermore, the fact that while attempts at dockerizing Matomo exist out there, none of them really strike me as easy to deploy solutions, as they grow outdated quite quickly. Matomo itself does not officially maintain a Docker contain ready to be deployed.

While 10 years ago a reliance on LAMP would have been totally acceptable, I feel like in 2018 I just can’t deal with it anymore. We live in an age of containers, reproducible deployments and many more luxuries that many of us have gotten accustomed to.

So I was thinking what every programmer at some point has thought: What if I implemented it myself?

A question of course that can either lead you to the conclusion to never think about this question again, since creating something like Matomo took a decade, and how will you ever replicate that effort as a sole developer? Isn’t this just classic Not Invented Here Syndrome?

Or, another take on this could be the following: Sure, Matomo or Google Analytics can do a lot of things, but really what I need is only feature X, and perhaps feature Y. And as long as I choose my solution to be modular and flexible enough, I can easily add the rest when I need it. So, I will reinvent the wheel, but only to the point where it starts spinning.

1. Anatomy of an Analytics Platform

So what exactly is it, that Google Analytics does?

I would claim that there are 3 major architecture components, that almost every analytics platform is offering:

  • Beacon: A small pluggable piece of JavaScript or tracking pixel that tracks page views and other interactions with a page.
  • Storage: A program that stores each tracked interaction in a database.
  • Dashboard: A web or desktop application that allows you to look at user interactions, aggregate and filter them, segment by user characteristics, and generate reports that can be shared easily.

If we really reduce our project to a simplified combination of those 3 components, we can get the job done quite quickly. For the sake of example, we will implement it in the micro-framework Flask as it allows us to create the beacon and storage component without using a lot of code.

2. The Beacon

For a beacon, we have to provide an HTTP endpoint that can accept requests coming from a user’s browser that contain detailed information on what site the user visited just now and other data about this interaction like browser language, screen dimensions, and so on.

We can either choose to add a small piece of JavaScript that contacts this HTTP endpoint, or we can just add a small image tag to the end of a page. For the image that we serve we can either serve a real picture, like a 1x1 pixel GIF, or just serve up empty content with a 204 No Content HTTP Status code.

So we get out the good old Flask and start hacking away. All we’re really looking to do is to create a beacon endpoint that can

  • Serve an empty response, and
  • Track the request to a database, in this case using SQLite.

It’s important to keep in mind that we have to use the HTTP Referrer URL here, since the HTTP path used to retrieve the beacon will not tell us what page the user visited, but in the case of a Referrer URL, we can see on which page the tracking beacon was requested. This page is exactly the page that the user’s browser is displaying. To illustrate this, I will show a simple curl request to the handy site https://www.whatismyreferer.com, once with referrer, and once without a referrer. The referrer will be http://www.google.com/.

Fun fact: the HTTP protocol specifies the referrer Header tag to be spelled Referer.

Output:

		Your HTTP referer:
	</h2>
	<p class="lead" itemprop="description">
		<div class="alert alert-info" role="alert">
		<strong>
		http://www.google.com/		</strong>

Output:

		Your HTTP referer:
	</h2>
	<p class="lead" itemprop="description">
		<div class="alert alert-info" role="alert">
		<strong>
		No referer / hidden		</strong>

We can see that by adding the referrer header, What is my Referer knows what our referrer is.

3. The Storage

We will also have to create the database used to track page views. The only table we need will be called beacon_hit and it stores individual hits of our beacon, which in turn are page views. It contains the following columns:

  • id: Primary key,
  • timestamp: Time and date of page view, and
  • url: URL visited.

Of course in the real world we would track a lot of additional information, but for the sake of example we will keep it brief.

When the beacon is then hit, the following information will be added as a new row to our database in our Flask application.

id timestamp url
1 2018-03-09T18:28:10.357074 http://localhost:8000/

4. The Beacon, Continued

We can now create the tracking endpoint that stores page views to the database and serves an empty response. As described above, it will need to respond to every request with 204 No Content. We can easily do that in Flask by just serving a response tuple containing

and we can call it a day. So let’s combine those two requirements in the following track() function. First, we need to import Flask and create an application object. Then, we can define the API endpoint.

We can use Flask’s test client to then call the beacon and trigger a row to be created in our database.

Output:

<Response streamed [204 NO CONTENT]>

As expected, the result is a happy 204 No Content. We can then print out what the database has stored:

Output:

(1, '2018-03-09T20:16:17.045070', 'http://localhost:8000/')

Since we want to show some dashboard data later on, we will fire off another 200 test requests. We visit two pages, page_a and page_b, 100 times each.

5. The Dashboard

Here, we will try to get things done with minimum necessary effort. We will use Pandas to read in the data and create some plots. Our goal is to

  • create a histogram showing which page by URL is the most popular, and
  • how requests are distributed over time.

We import Pandas and get started:

To read in our data with Pandas, we will read in the whole beacon_hit table as a Pandas DataFrame. To achieve that we use pandas.read_sql .

Let’s take a quick look at the data contained in the table and now read in as a DataFrame:

Output:

id timestamp url
1 2018-03-09 20:16:17.045070 http://localhost:8000/
2 2018-03-09 20:16:17.089820 http://localhost:8000/page_a
3 2018-03-09 20:16:17.091435 http://localhost:8000/page_b
4 2018-03-09 20:16:17.092620 http://localhost:8000/page_a
5 2018-03-09 20:16:17.093510 http://localhost:8000/page_b

That looks excellent! Let us now plot page popularity using a simple Pandas value count on the url column.

Output:

matplotlib.figure.Figure at 0x1125e0ef0

We can clearly see that /page_a and /page_b are equally popular with our Flask test client.

In order to show request distribution over time, we have to be a bit clever since our test data has timestamps that are very close. We therefore use a Pandas Grouper to group the rows by milliseconds using pd.Grouper(freq='ms').

Output:

timestamp url
2018-03-09 20:16:17.283 0
2018-03-09 20:16:17.284 1
2018-03-09 20:16:17.285 1
2018-03-09 20:16:17.286 1
2018-03-09 20:16:17.287 1

That’s good enough for our purposes and we can simply plot the data:

Output:

matplotlib.figure.Figure at 0x10f3ebc18

6. Wrap Up

What was our initial goal? To find out what the minimum amount of moving parts required to host your own analytics solution is. We were able to condense this to three parts. Two of those I have made Open Source and available as Francis Beacon. I am very happy that it is possible to implement your own analytics solution. Especially in Data Science, the more power you have over the data source, the better your results will be.

Date created:
March 9, 2018

Back to Index

You are more than welcome to share your thoughts via email