Past, Present, and Future of Armstrong

Most of you who know me have heard me talk about Armstrong, the open-source news platform that I helped create when I first joined the Texas Tribune. I have and continue to talk at length about Armstrong and its future, but I’ve never collected those thoughts into one cohesive document outlining how we got to where we are now, what the current state of the project is, and where I hope to see it go.

This post is my attempt to do that.


Armstrong is peculiar if you look at it from the outside. It might be hard to understand exactly what it is and how it got to the point it is. This section should help you understand that a bit more.

Not a CMS

Using common names can be unfortunate. Most people hear Armstrong CMS and think it’s a content management system akin to Eidos on the closed-source side or Wordpress and Drupal on the open-source side. I’ve always envisioned Armstrong as a platform to build on top of, not simply a CMS. The distinction is small, but important.

You work within a CMS

You build on top of a platform

A CMS is something you use. It provides the tools you need to manage content. A platform is something that provides a base to build upon. It’s my belief that more 95% (maybe more) of the pieces that make a news focused website a news website are all the same. Everyone needs a way to collect similar content into sections, a way to schedule content, or a way to control publication status on content. That last 5% of content, however, is unique and what makes a site interesting. Being able to reuse data about bills throughout your site without having to confine it to a big blob of text is what allows us to do interesting things with a site. This is where Armstrong comes in.

Armstrong isn’t meant to be used out-of-the-box any more than a Lego™ train set is meant to be a toy train set as soon as you unpack it. Both require assembly, but both allow you to exercise some creativity in how you assemble them. To do this, we’ve taken an unconventional approach to packaging everything in the project.

Everything is a package

Python is famous for it’s horrible packaging solutions. It’s gotten a lot better over the last few years, but people still package most software in one big bundle. I created the packaging schema for Armstrong based on the way I wished Python packages were handled, as if they followed an adopted version of the Unix Philosophy:

Packages should do one thing, and do them well. Packages should work together.

This means that there’s some 25 packages on PyPI for Armstrong. Many of these work in concert with other Armstrong packages to form a bigger whole. I broke it down along two main lines with a few others thrown in.


This section of Armstrong contains all of the pieces essential to nearly all websites. These packages either have no models, or are meant to be used almost exactly as-is with little or no modification.

All core packages contain an arm_ prefix in the last part of their name: armstrong.core.arm_content, armstrong.core.arm_wells, and so on. This is to avoid potential naming conflicts since Django still assumes that its apps are all flat without full module names.

Try as I might, armstrong.core.arm_content did end up pretty big. It contains most of the mixins used to build the larger models throughout the system. Anything that can loosely be considered content in an Armstrong project probably has some connection to this package.


These apps are meant to be usable out-of-the-box, but are most useful as example implementations. This is one area that I didn’t document as well as I should have. Almost no one would (or should, for that matter) use armstrong.apps.articles as their out-of-the-box article implementation. For very simple sites it will probably work, but my assumption has always been that most sites will take that as a guide and build something similar to it.

A great example of this is the armstrong.apps.donations project. We use that pretty extensively at the Texas Tribune since we’re a member-driven organization, but we don’t use it in its out-of-the-box configuration. We have a custom tt_donations app that extends the views to add extra functionality and we have custom backends for all of our payment processing and CRM syncing.

Hatband and Pops

Any news tool is only as good as it’s admin interface. Unfortunately, most of our time while funded by the Knight Foundation was spent on backend code, but we do have a solid start of a custom admin interface that’s broken down into two pieces.

Hatband is the collection of Armstrong-specific interfaces to the Django admin. It’s meant to be used as a drop-in replacement for django.contrib.admin that extends the behavior. It provides several custom inline interfaces and will hopefully contain even more. It exposes a JSON API for searching any type of model that’s registered with the admin and has search_fields turned on. The plan has always been to use this to create a rich API on top of the admin. Search is simply the first foray into that.

Where Hatband is behind-the-scenes, Pops is the user-interface side. It’s currently built using Twitter’s Bootstrap framework and was built on top of a fork of django-admintools-bootstrap. Pops is meant to be entirely standalone and have no ties to Armstrong at all since it’s simply the skin on the interface.

Current State

The original Knight Foundation grant to The Bay Citizen and The Texas Tribune ended in early summer of 2012. Since that time there hasn’t been any full-time development dedicated to Armstrong, but that’s not to say that development has stopped. Both the BC and TT, along with a handful of other organizations, use Armstrong internally and are continuing development on the project.

There are a few things I want to call out.

Timed Releases of Armstrong

The original idea, and one I’ve deviated from since last year, was to do timed releases, every three month. You might have noticed that the version numbers are pretty high. That’s because they follow the format vYY.MM. That way you know when the last major release of Armstrong was. Each one of those releases is just the latest stable code from all of the components of Armstrong that are considered production ready.

One key point, however, is that the main armstrong package (note the lowercase, that means it is code in its packaged state, not the project as a whole), is just a collection of other packages. You don’t have to install armstrong to be able to use various components. For example, you can pull armstrong.utils.backends into any project without using anything else from Armstrong if that’s all you need.

Release Components

All of the components of Armstrong are released independent of the major armstrong releases. There’s been a handful of component releases in the last year and more are being worked on right now. Each of these follows Semantic Versioning, or SemVer. That means that you can always upgrade within a major version, say from v1.2 to v1.6, and not worry about your code breaking. If anything breaks, a new major number is introduced. So far, we’ve only had to do that once: armstrong.apps.related_content.

All of the components in Armstrong that reach a v1.0.0 release, other than those grandfathered in for armstrong’s first stable release (v11.09), are being used in production. Following SemVer, production code goes to v1.0.0 as soon as it’s production. Part of the criteria for code making it to stable is that it’s being used on at least one site as production code. There’s one v0.x release that has running code so you can install it, then once that’s ready the v1.0.0 should be a simple version number bump with no code change.

The production requirement goes for point releases of code once it goes stable as well. Someone has to be using the code that’s in a pull release in order for it to be considered stable (and tested) enough for a release. Unit tests are a requirement, but code that is running on production is the final requirement for any component that’s released as stable.

This points the burden mainly on the Bay Citizen and Texas Tribune to make sure we’re running code that we’re trying to get released. Right now we’re the main ones that are effected by this, but it allows us to ensure the quality of the code. As more organizations start to contribute, they’ll have to play by the same rules.

Future Plans

There are a handful of areas where I would like to see Armstrong grow toward in the future. These coincide with the technical direction here at the Texas Tribune, so we’ll be driving many of these changes based on restructuring of our internal code.

Streams of Content

One of the biggest regrets in Armstrong was that I relented when arguing that we should structure content to exist independently with streams that any content could opt-in to. The argument against this route was that we had a tried solution—monolithic concrete-model dependencies—so why try something new until we’d replicated what we knew works. The old method does work, but it’s not scalable, whereas independent streams that you push content in to means you can decouple that relationship and scale it to many different types of content.

For example, say you have an Education Stream that contains content related to education. Stories can put themselves in that stream by providing the information the stream expects, but so can an update to data about a campus. All the data has to do is be able to render itself the way the stream expects and it can opt-in.

My initial plan is that an object would provide at least one rendered version of itself and its canonical URL. For a section stream, that rendered version would probably be the title, plus artwork, byline, and description.

This decentralization means that the stream display can be moved around to different services. You could also make it streams all the way down. Content notifies data streams that expect certain JSON documents and those data streams notify content destination streams (think: HTML, iOS, computers on the dashboard, TVs, and so on and so forth).

Those with a background in enterprise software might recognize this type of decentralization by another name: Service Oriented Architecture, or SOA. This type of architecture is not simply nice, it’s a requirement in a multi-device world. Building services that can only return HTML is shortsighted and going to cause problems as the number of devices our content is displayed on explodes in the coming years. Decoupling content from the various streams they’re consumed in is the first step in future proofing Armstrong.

Testbed for a new Django Admin

One area where I think a SOA allows greater freedom is the admin interface. The Django admin is great for what it gets you out-of-the-box, but you outgrow it very quickly especially when it’s laid next to modern web tools. You have to remember, the Django admin was designed in 2004/2005 when your main option for dealing with any type of data was phpMyAdmin and editing the database directly!

One thing I hope to do with the Hatband/Pops combo is create a testbed for experimenting on top of Django’s admin. These roles aren’t set in stone, but my thought is that Hatband will serve as the place for Armstrong-specific experimentation and Pops will be the place for generic Django experimentation.

Since starting Hatband and Pops, a few other tools have popped up in this space. Nobody has gotten significant traction, but I’m not opposed to joining forces with one of them if somebody does start to head down a solid path, but there are a few things that I see as a requirement.

  1. It needs to build on top of the existing django.contrib.admin code. The admin’s bones (with a few exceptions) are really solid. Rewriting it from the ground up isn’t a good use of time. It needs a lot of refactoring and many of the changes wouldn’t be backwards compatible, but it’s possible to make it happen by simply building on top of what already exists.

  2. It needs to focus on decoupling the actual interface from the discovery and registration of apps along with exposing an API to them. Right now, the biggest wart on the Admin is how tightly coupled display and discovery are. Any new admin needs to focus on providing a solid API (both Python and REST) for working with the apps that are registered. On top of that you can build a solid, default client interface. Having a dogfooded API ensures that others can build alternative interfaces on top of it. Think: native iOS apps for the admin!

  3. It needs to exist outside of core Django. I think the Admin is one of the reasons Django has gotten the traction it has. That’s great, but right now it moves too slowly for that to be useful for such a potentially rich web application. Having the admin exist as a separate project means that it can move more nimbly, release more often if it needs to, and gather support from those who might have no interest in working on a traditional web framework, but would love to work on a web application.

Separating Editing and Publishing

Currently, CMS tools assume that you’re editing and publishing content in the same tool. Those are two different workflows that need to have different tools: a collaborative authoring/editing tool and a publishing tool. They can exist on the same domain and even within the same major tool, but each workflow needs a different presentation and to be separated from the other.

The editing tool needs to have real-time feedback for the user when edits are happening. It should update in real-time showing who is editing what, the changes they’ve made, and so forth. It could include the ability to lock a field by focusing on it, but it should allow you to return control back to other users by removing focus, but ideally it’s smart enough to work with multiple users editing the same content.

This interface needs to focus the user on writing and editing. Things like sections, tags, locations, and so forth are all secondary content that should be tucked away, accessibly only when called on, to allow the user to focus on the task at hand. It would take a lot of user testing to design this system in a way that it could replace the existing tools (the number of emailed Word documents at the Tribune is still a source of embarrassment for me), but you’d end up with a solid workflow to take something from idea to finished product with the right focus.

Once the content has been written and edited, it needs to be published. Focusing explicitly on content takes all layout decisions away from the authoring experience and moves them to a place where you can make device-specific choices.

Responsive design should be the first choice for all content so it reaches the broadest audience, but there is also room for device-specific display where it make sense. Layout tools should enable this.

There is a start for this in armstrong.core.arm_layout with some code to help reuse model-specific rendering throughout the site, but those are baby steps toward a GUI-based layout tool.

Being able to control any aspect of the display of the site across multiple devices is the Holy Grail of a news platform, and one I hope we’re able to tackle as part of the Armstrong community.

Up Next

There are some very immediate plans, however. First, we need to roll another release of armstrong. I plan on creating armstrong v13.06 at the end of the month with the latest stable versions of all of the stable components.

v13.06 is going to be a maintenance release to get everything updated to the latest and remove the component-level requirement on Django and ensure full testing of Django from Django 1.3.x through Django 1.5.x. From here forward, the only dependency on Django is going to be specified at the armstrong level, leaving it up to developers to work through whether they can upgrade. We’ll continue to test components against all supported versions of Django.

This release will put us back on the timed release. The next release after this will be v13.09. I know the folks over at the old Bay Citizen have some new code they would like to see released and I’m hoping to have a solid admin interface for armstrong.apps.related_content by then as well.

Tools vs Materials

Last week I attended the Artifact Conference. It was my first design conference, and I was impressed. The talk lineup was amazing and speakers consistently delivered. The talk that stuck in my head (and not only because it was the last talk I attended) was Dan Rose’s Photoshop’s New Groove.

The talk was about using Photoshop while designing for the web. It was the counter balance to all of the random snarky comments on Twitter and so forth about Photoshop and how it should die in a fire (I’m paraphrasing, but just barely) and how you’re not a real web designer if you use Photoshop and so on and so forth. It was a great talk.

Dan spent a lot of time talking about using the tools that work for you. Dan Mall had a great line in his talk on Monday about fighting your tools as a way to kill creativity. I think he’s 100% right. Find the tools that work for you and use the hell out of them.

That said…

HTML and CSS aren’t the tools of web design, they’re the raw materials.

The difference between tools and materials means the world. It’s like a chef who doesn’t cook their new recipes because they know how the various ingredients are supposed to work together. A huge part of culinary school is learning how various ingredients interact with each other to produce different effects. It borders on a chemistry degree. The training doesn’t stop with the chemical reactions, though.

A chef wouldn’t produce a new dish by only thinking about the interaction of certain ingredients and putting together something she thinks would work well, she actually makes it. The designer who is creating for the web who doesn’t move to HTML and CSS with their design is like the chef who relies on the cooks to know exactly what to do. Until you get your hands dirty with code, you’re simply brainstorming ideas.

It’s taken me a long time to actually be able to articulate this. It wasn’t until I heard Dans refer to HTML and CSS as tools that it clicked: we’re looking at the world differently (shocking! I know!!). Hearing designers at Artifact talk about HTML and CSS as tools akin to Photoshop is what gave me the proper lens realize where the disconnect was.

Open Source Licenses

IANAL, but I like to pretend like I am on the Internets. This past week at NICAR, the discussion of open source licenses came up in one of the evening tracks over a few bourbons, or it might have been wine by that point, but I digress. The general theme: licenses are confusing.

I know a little bit about them I’m hoping to shed some light on them for fellow journalisty type developers who are thinking about releasing their code but aren’t sure which license they should use.

Caveats and such: I’m seriously not a lawyer, this isn’t legal advise, and so on, et cetera. Please talk to one if you have serious legal questions.

Range of Licenses

There are 69 official open source licenses in use. There are many, many more that are snowflake licenses—licenses that have provisions that are unique to them. Many companies, including ones that I’ve worked for in the past, have created custom licenses by modifying one of the main open source licenses. Many of these have been written by lawyers, but snowflake licenses are an unknown quantity until they’ve been tried in court.

You should avoid snowflake licenses for your open source code. Having a license that is unique to your project increases the barrier to entry. Each developer has to read and understand the license and try to tease out any differences you have with the more common licenses.

Instead of going the snowflake route, opt for one of the popular open source licenses that are commonly used. Each of the licenses have their place, but I’m going to touch on the three that are the most common and one additional license that I think journalists should be familiar with.

GPL: The Viral License

GPL, the Gnu Public License is possibly the most popular and familiar of the open-source licenses. It’s the license that the Linux Kernel and many of the tools that ship with the Linux operating system are released under as well as the wildly popular WordPress blogging platform. I can distribute GPL software any way I want. I can give it away, I can charge, I can do some hybrid of those two. One thing I can’t do is limit what you do with it after you acquire it.

The GPL is a copyleft license, sometimes referred to as a viral license. It’s viral because it forces your hand when it comes to licensing derivative works. Any derivative software must be distributed a compatible license like the GPL. In other words, if I came up with a way to modify Linux and wanted to distribute it, I would have to distribute it under the GPL license. That distribution could be paid, but anyone who pays for it could then redistribute it at will.

GPLv3 has some interesting provisions to. Namely, the Additional Terms. These are optional things that the author can add. For example, 7b requires “preservation of… author attributions” in a project. This is useful for businesses who want to release their software, but want make sure that their competitors can’t do a find-and-replace for their competitor’s name and repackage the software as their own and have to fully credit them, including displaying logos in the user-interface and such.

New BSD and MIT: Do what you will

On the other end of the spectrum are the New BSD (more commonly referred to simply as BSD) and the MIT licenses. These two licenses are much more permissive, allowing redistribution with only minor restrictions.

The MIT simply requires that the copyright notice be transmitted with “all copies or substantial portions of the Software.” Essentially, you have to tell the outside world that the software you’re distributing contains the MIT licensed software. Both Backbone.js and Underscore.js, two JavaScript projects that originated in the DocumentCloud project, are licensed as MIT.

The New BSD license says the same thing, plus one other clause that says you can’t reuse the original package’s name nor the names of any of the contributors to “endorse or promote products derived from this software without specific prior written permission.” FreeBSD and OpenBSD use the BSD license as does Django.

Licenses and Communities

My thoughts on licenses have evolved over the years. Jacob Kaplan-Moss introduced me to the idea of thinking of licenses as a community identifiers (Side note: he was introduced to this thought process by Van Lindberg, the current PSF Chairman and author of the book Intellectual Property and Open Source). All communities have certain things that they use to identify those who they have a common interest with. Rockabillies have fashion sense and a music that’s unique to them. Gangs have the color of their clothes. Developers have their languages and their licenses.

Each sub-community in the open-source community have their preferred license. For example, jQuery is dual-licensed as GPL/MIT, so most developers releasing software for jQuery use a similar license. The JavaScript and Ruby community tend to use the MIT license, as is evidenced by the amount of MIT code on and Rails. The Python community, and particularly the Django have a bias toward BSD.

Releasing software meant to be a part of those communities without following the cultural norms within those communities is a sure way to stick out. It’s like walking into a rockabilly bar dressed in a suit. You should always have a good reason for bucking the norm within a community that you want to be a part of. Trying to release GPL licensed code that builds on top of Django means that you’re not part of the community—you’ve set yourself up as an outsider.

Releasing your software with a more restrictive license than is common in a community that you’re trying to participate in also means you’re placing further restrictions on those in the community. You can use their BSD or MIT licensed code, but they can’t use your GPL code in their projects. That’s essentially telling the other developers that you love their contribution, but not enough to let them use what you’ve built under an equally permission license.

So what to use?

This is where I should mention discussions of being in Rome and so on, however, I think you should use another license: Apache License 2.0. Apache is essentially a BSD license with two very distinct modifications.

  1. Any contribution to the project is considered to be made under the terms of the Apache License. Contributor License Agreements (CLAs) can be used to enforce something similar with BSD or MIT licenses, but they aren’t guaranteed. The Apache License bakes the terms of the contribution in by default. 1. Apache grants a full rights to any current or future patents that might be derived from the contribution.

This last part is the reason to use Apache. When we started the Armstrong Project, I called up Jacob Kaplan-Moss to ask his opinions on licenses. He sold me on Apache with this line:

If I had [the licensing of Django] to do over again, it would be Apache today.

JKM’s endorsement on the grounds of patent protection was the reason that I advocated to use the Apache License on the Armstrong Project when we started instead of BSD, which is more common in the Python community (remember, community signifiers and all). I’m not worried about any current contributor, I’m worried about who might own the work a contributor makes in 1, 2, or 5 years.

Most newspapers are in a state of flux right now. Let’s say The Timbuk2 Independent contributes a few components to Armstrong. In a few years, they get bought by MegaNewsProfitExtraction, Inc. who then starts evaluating all of the intellectual property they’ve acquired. They realize the contribution from The Independent is patentable and apply for an receive a patent for their small contribution. Under a license like BSD or MIT MNPE, Inc. can now go around attempting to collect all of the patent licensing fees they’re due based on your use of Armstrong.

I don’t think that scenario is that far out there. Remember, you never write the rules for the guy you like, you write them for the one you don’t. Assuming this scenario, the best thing we can all do to protect ourselves is use a license that protects us from the future patent trolls that are lurking under the bridges of acquisitions.

Got other ideas? I’m interested in hearing them.

Generic Dangers

Here at the Texas Tribune, we started using a project called django-chunks some time last year. Consider this post a cautionary tale and think long and hard before you start using. We didn’t. We’re paying the price.

The Promise

django-chunks gives you the ability to inject arbitrary chunks of HTML into any template inside Django. You load up a template tag library, call a templatetag, and you’re off to the races. No more waiting on clients (or other departments) to get you copy. “Here’s the chunk, go to town,” you tell them.

That seems pretty good, right? We’re all lazy, that’s why we program. The idea of making a computer do something for us tends to be at the core of why most programmers got into programming. Second only to making computers do the work is making other people do the work, so django-chunks let’s us off load that work to someone else.

The Problem

Once you’ve started making things chunks, everything becomes a chunk. It’s a golden hammer of sorts.

This field needs to be copy edited once in its entire existence? Bring in the chunks!

This type of thinking is shortsighted at best, and harmful at worst. This morning I took a look at our join page only to discover that we were giving users a security error due to a mishandled chunk. An image URL had been entered as http instead of https, mixing insecure content in to a secure page. Yes, there was a chunk created who’s sole purpose in life was to be edited to change out an <img> tag!

The Solution

Don’t use chunks. The case above is where a simple model should have been used if we really need to let non-technical people change out the image. We have a hard requirement on https on that page, and as programmers we can enforce that.

Django is meant for building custom stuff. Go do that. Use things that help you build that custom stuff faster, but don’t go looking for turn-key. Someone else’s answer to the problem isn’t going to be as good as what you could have built.

Python for Beginners

Yesterday I attended the Pycon Web Summit and there was a lot of talk about getting new programmers started in Python. I’ve been thinking about this a lot the last year since helping found the Austin Web Python User’s Group and I think I have a solution.

Success early, success often

One of the key things we need to be able to do is get developers on every platform up and running quickly. An iPython shell is a wonderful place for a newbie. Do you remember the first time you typed code into a REPL and it did what you told it to. 2 + 2 returned 4, then I’ll bet you tried 2 + 3 just to see that it wasn’t some trick. That sense of wonder, excitement, and, most importantly, accomplishment needs to be priority number one as we move forward whether we’re starting someone on raw Python for data manipulation or Django for full web application development.

The number of steps between “I want to learn to do X” and actually making Python do something for you needs to be minimal. That means the first words can’t be “pip install django” to get Django installed. We need to teach newbies about pip and virtualenv and how to install Python and all of the steps that go in betwee, but not yet.

The tool

I think an environment built on top of Vagrant is the right solution. We can bootstrap a virtual machine that’s ready to start accomplishing things the second it’s launched. We can’t start teaching Python by telling people they have to go find, download, and install Ruby, Rubygems, VirtualBox, and Vagrant.

The solution to this problem is a one-click installer that gets Vagrant and all of its dependencies installed and presents you with a GUI to select the type of environment you want to create. Need Django, Pyramid, NumPy, SciPy, or hell, even a setup with [csvkit]? Select that and a few minutes later (assuming a broadband connection), you’re up and running with a prompt that lets you start working.

This is doable. I haven’t done it. I’m not sure that I could (I haven’t programmed for Windows in well over a decade). I want this out there though. I want more people thinking about it and hopefully someone can kick the process off. I’ll help in any way I can and I’ll definitely use it if someone starts the project.

Got a better idea? Let’s hear it.

Deploying TileStream to Heroku

This past week I attend the 2012 IRE conference. Remember all of those #nicar12 tweets you saw from me and few other programmery/journalisty type people? That’s the conference we were all hanging out at.

Custom maps were one of the big themes. There were a few TileMill talks and they were all packed. TileMill, for those who aren’t familiar, is a tool that let’s you create custom map tiles–the images that make up maps like Google Maps–so you can have a map that’s entirely unique. An example of this is the Idaho Unemployment Map by the folks over at State Impact.

We’ve been talking about using TileMill at the Texas Tribune for months now, but we’ve yet to actually deploy one. A few of us have TileMill locally and have played with it, but the tile serving component is something we haven’t touched.

I came back from the conference and got sick. Yesterday, while trying to kill some time without thinking of anything particularly important I decided to see what was involved with deploying TileStream.

TileStream is a tile server written in Node by MapBox, the creators of TileMill, to generate and server the tiles for a map you create. Since tiles are simply PNGs, it seems like you should be able to just generate a whole host of files, upload them to a server, and call it a day. The problem that a tile server solves is having to generate all of those tiles at once. Generating them, then uploading them once is a pain, but what happens if you need to make a change to them?

Lately, I’ve been on a “no new servers” kick. I’m tired of seeing the amount of time spent tweaking servers instead of working on code. DevOps is fun, don’t get me wrong, but sysadmins we are not. With that in mind, I decided to take a look at what’s involved in deploying TileStream to Heroku, a “cloud application platform” that supports a whole host of languages—including Node.

Preparing for Deploy

The very first thing you have to do is create a map an export it. That’s a topic unto itself, so I’m not going to cover it here. I created a simple copy of the state of Texas with all of its counties outlined and colored in. I forget where I procured the shape file, but some Googling should turn it up if you want to follow along.

Make sure to export the file as the mbtiles format when you export it. Where you export it to isn’t important right now, just remember where it’s at.

Next, you have to make sure Heroku is installed. If you already have a working Ruby and gems environment with Git and so installed, you can run gem install heroku to get the command line client. If you don’t, check out the Heroku Toolbelt for a quick start to get setup. Once you have the command line tools setup, log in to your Heroku account with heroku login and follow the directions.

The next step is to create a new Git repository. Heroku uses Git as its means of tracking files to deploy. You’re going to have to learn at least a little bit of Git if you’re going to use Heroku (side note: I’ve written two books on Git and highly recommend Pragmatic Version Control using Git if you’re new to version control). Once you have a Git repository, run the command heroku create -s cedar inside your working tree. You should see something similar to this:

prompt> heroku create -s cedar
Creating hollow-fire-2448... done, stack is cedar |
Git remote heroku added

hollow-fire-2448 is the name of my Heroku application. Yours will be different. Now you have to tell Heroku what to install. To do that for Node applications, Heroku uses a package.json file. That’s the file that Node applications use to set up the dependencies to make sure that everything is installed. For this server, you just need to declare a simple dependency on tilestream. My package.json file looks like this:

  "name": "texas-counties",
  "version": "0.0.1",
  "dependencies": {
    "tilestream": "1.0.0"

Add that to the repository using git add followed by git commit. The next step is telling Heroku how to run TileStream. Heroku uses a Procfile to handle starting and stopping applications. The Procfile is run using Foreman and can define all of the processes required to run an application. The format is <name>: <command> and for this application you only need to add one line:

web: tilestream --host --uiPort=$PORT --tilePort=$PORT --tiles=./tiles

There’s a couple of things going on there. First, notice that I’m explicitly adding a --host name and using the name of the app that Heroku told me when I called heroku create earlier. TileStream currently only responds to requests on hosts that it recognizes. You’re going to need to change that line to be whatever your Heroku app’s name is.

Next, notice that both --uiPort and --tilePort are set to the value of $PORT. Heroku exposes $PORT as an environment variable to let your application know what port to listen to for incoming connections.

Finally, you set the directory for tiles to ./tiles. Commit this, then push to Heroku to verify that everything went according to plan.

prompt> git push heroku master 
Counting objects: 6, done.
Delta compression using up to 2 threads.
Compressing objects: 100% (5/5), done.
Writing objects: 100% (6/6), 659 bytes, done.
Total 6 (delta 0), reused 0 (delta 0)

-----> Heroku receiving push
-----> Node.js app detected
… and a whole bunch more output …

Go ahead and stand up and stretch. Go grab some coffee or tea or water, whatever you vice. This step takes a few minutes while Heroku installs all of the dependencies and such for TileStream for the first time. It’s kind of awesome, though. Without a single bit of server administration, you’re just a few minutes away from having a fully operational TileStream server.

… waiting on Heroku to finish up …

Ok, done? Now run heroku open. This launches your browser and opens the URL of the Heroku application. If everything went well, you should see the empty TileStream server like this.

Empty TileStream

If you don’t get a page like the above, check the logs by running heroku logs to see if it gives you any clues. Another thing to double check is the process list. Run heroku ps to make sure that web.1 has a state of up.

That big error is the non-user-friendly way of saying there’s nothing in the tiles directory to read and display. Remember the mbtiles file you created earlier? Now it’s time to move it into place. Inside your Git repository, create a directory called tiles and copy the mbtiles file into it. Once the file is in place, add it to Git, then push the new commit to Heroku.

This push is going to take a little bit, depending on how fast your connection is. It has to send the entire mbtiles file over the wire to Heroku. Having done this a few times now, it seems like Heroku might throttle large uploads. I start out at a few hundred KB/sec, then it drops down to around 100KB/sec for about 30 seconds before settling in at 80KB/sec. Their business isn’t receiving huge files, so it would make sense if Heroku did throttle to make sure one large upload didn’t take over their entire pipe.

Once the push has finished, reload your browser window and you should see your new map, much like this:

TileStream with one map

And now, you have a tile server. Deploying to Heroku for this is a great fit for the standard news application. You need the ability to handle tons of traffic as you launch, then scale back until it hits maintenance mode where you only need a skeleton server running.

Heroku gives you one dyno–think of that as one process on a server–for free, with each additional dyno costing $0.05/hour (see the Heroku pricing page). That means you can spin up several dynos to handle the initial flood of traffic, then scale back to a smaller set and only have to pay for the initial spike. All, without any additional work on your end setting up or configuring servers.

Now, the one caveat to all of this is that I haven’t actually tried running TileStream like this with a production load. I’m not sure what kind of performance we could get out of it or what limitations there might be. The only way to answer that is to try. Hopefully we’ll be able to pluck one of the projects out of our pipeline and do some custom maps for it using TileMill and TileStream.

Where to from here?

The next thing you need to do is write some JavaScript to interact with the tile server. Leaflet has gained a lot of popularity and seems to be the default choice. I’ve yet to play around with, but that’s a topic for another blog post.

If you’re interested in seeing what all of the pieces look like together, my Heroku app is still online at I’ll try to leave it spinning, but if I take it down, I’ve posted the repository on GitHub so you can see all of the files in their original state.

Importance of Context

Today I discovered the 99% Invisible podcast on architecture and design. Their latest podcast, Pruitt–Igoe Myth, tackles the problems associated with the Pruitt–Igoe housing project which was built in the 1950s in St. Louis to provide affordable housing in the St. Louis urban core. Due to a variety of reasons, which the podcast explores, it was torn down in the 1970s. From Wikipedia:

[Pruitt-Igoe’s] 33 buildings were torn down in the mid-1970s, and the project has become an icon of urban renewal and public-policy planning failure.

After listening to the podcast, you come away with the impression that this isn’t a fair assessment. It was built at the beginning of the White Flight, in a part of the city that saw a decrease in population, not the projected 100,000 every decade increase that was forecasted. These and other issues contributed to it turning into the very thing it was trying to prevent: a slum.

The building is considered the example of the failure of Modernist architecture as it applied public house, but if you view it in the context above you can see that there are many external factors that contributed. It’s easy to pick one particular piece of the puzzle and lay the blame on that for the failure. It’s much harder to try and understand the complex relationship around what caused the issue.

Applied to Programming

This type of logical error is present in many (not all, but many) of the conversations about what framework or language to use, what methodology should be adopted, or even where to found your startup. It’s easy to point to one success or failure and declare “X is why Z happened, so if I want to duplicate Z, then I must/must not do X.” This type of cargo-cult behavior is dangerous and should be guarded against.

Yesterday I tweeted this:

Whoa! JustinTV is moving from #rails to #django. I’m telling ya, Python & the web with a little Django mixed in is about to blow up.

It gives the impression of just that type of “Y leads to X” kind of thought process that I’m against. To clarify, I whole-hearted expected what kvogt wrote when explaining why they’re moving to Django. To paraphrase: “it just makes sense right now to be on one platform.” isn’t going to suddenly take the world by storm after moving to Django any quicker than they would have if they had moved their Python backend to Ruby.

That said, I stand behind the final point of that tweet. There are tons of shops using Python and Django that aren’t vocal about their use. Python is powering business logic that runs on servers sending me music, tracking my location, displaying my news, and a whole host of other things. Python can do everything low-level system tasks to scientific and financial analytical calculations to high-level business logic for websites and everything in between.

I can’t help but thing there’s going to be more announcements this year: shops standardizing on one language and that one language being Python.

Using Basketweaver With GitHub

Last month I blogged about using Travis CI with Armstrong. Things have been going along fine until the last few weeks. Tests were failing due to network timeouts while talking to PyPI. Never one to take failing tests lightly, I set out to fix it.

From local testing, it appeared that there was some sort of selective filtering happening at the server level on PyPI that was causing our tests to fail. All of our tests in the CI environment follow these tests:

  • Install all of the development requirements with pip install -r requirements/dev.txt
  • Install the local package
  • Execute the tests using fab test

I could follow these steps to the letter locally in a fresh virtualenv, but the second they hit the Travis-CI server they would time out while trying to install everything. We’ve seen similar behavior at the Tribune when we roll out new servers. PyPI appears to be up, but installs fail due to timeouts.

Once I confirmed this, I started looking at alternatives to as our main index for testing. My initial thought was to have a dynamic server that would act as a proxy to PyPI and cache everything locally. This requires the least amount of work long-term—assuming the server stays up. The problem was that nothing worked quite the way I wanted. The closest I found was collective.eggproxy. It felt a little odd and wasn’t very configurable without going the Paster route, so I decided to fall back on basketweaver.

Basketweaver builds a static index suitable for using with pip via the --index-url option. It takes a directory of files, then generates the HTML that pip can scrape to determine if the package exists. This HTML can be hosted anywhere that can serve a static HTML page, such as GitHub Pages.

Working with GitHub

There’s a few hoops to jump through when deploying to GitHub Pages. First, make sure you include an empty .nojekyll file. GitHub assumes everything you want to publish is in Jekyll, but this file tells GitHub to not parse your files.

Next, and I can’t count the number of times I’ve done this, GitHub Pages doesn’t give you directory indexes. Basketweaver generates its index in the /index/ directory so you can’t hit the plain GitHub Pages URL and expect to see anything more than an error message. Make sure to add the /index/ after your GitHub Pages URL to view the it once you’ve published your changes.

The next thing I do is rework where basketweaver looks for files to build the indexes. I really don’t want to look at a full directory of files at my root directory, instead I want all of the files stored in the creatively named ./files/ directory. Basketweaver installs a file called makeindex which I can never remember, so I created a file that remembers it for me.

The last thing to do is to use the newly created index when installing packages. For Armstrong, we do this with:

pip install -i \
    -r requirements.txt

I haven’t gone to the trouble of setting up a CNAME for yet, so we’re using the main address.

There’s one final gotcha: PyPI uses routing that treats and as the same URL. That’s why pip install Django and pip install django both work even though the former is the correct package name. The URL spec is ambigous as to whether this is correct, but most web servers are case sensitive, including GitHub Pages.

This will get you if you have dependencies on packages that don’t use all lowercase names, such as South, Fabric, or Django. All three of these are dependencies of Armstrong. The fix is to make sure that your install_requires and requirements files have the correct case. The easiest way to determine this is to look at the output of pip freeze and make sure you’re using the same package name as it generates.


At the end of the day, this keeps our tests from being held hostage whenever PyPI goes on the fritz or starts randomly filtering requests as it seemed to do this past week. All that said, we’re still borrowing other people’s infrastructure. GitHub had a little blip while I was writing this post, underlining that you get what you pay for.

While you can use Basketweaver and GitHub to create a mirror of sorts for your packages, make sure you control the infrastructure if its mission critial that everything always stay up. That, or pay for it so there’s someone to call when it goes down.

Editing Mode

In case you didn’t know, I use computers. A lot. Between working as a programmer, writing books, and the occasional leisure time spent playing on computers, the vast majority of my life is spent with a screen of some sort in front of it. That time means I come across and try a lot of different tools, and some of them actually make my life better.

One such tool I’ve started using extensively while writing my latest book is Notability. It’s a note taking application that let’s you import PDFs that you can write directly on top of. This is important because I can’t see the typos from within my text editor.

Switching environments when switching tasks is an important concept I picked a while back. For me, that switching comes when I build the book and switch over to a PDF version on my iPad to read it. With Notability, I can take the PDF version of my book and change rooms or sometimes just turn the chair around away from the desk, and switch into “editing” mode.

I’m not alone in using an iPad for editing. I hadn’t found an app that worked well for note taking though, so I often switched back to my text editor to write notes. Having to mentally switch context back and forth and back and forth as I physically switched devices hurt my productivity. Being able to do it all in one app has made iPad editing much more feasible.

Once I’ve finished an edit pass and have a whole slew of changes to make, I switch back to my computer with my iPad close by. Notability lets you change the color of the pen you use, so I swap it out for green, and slightly larger for impact, then start slashing through all of the red as I mark edits off. The satisfaction from marking something off with a physical slash can’t be overstated.

I’ve been using Notability for about a month now and don’t know how I managed to edit without it. I highly recommend it if you have an iPad and are doing any type of writing/drafting work.

Question for the Reader

I’ve been considering a series of short posts like this about tools that I use and how they fit into my work flow. I love to watch people work and see how they interact with their systems, though. Is this something that interests you?

Travis and Python

Today I took my name back and got Armstrong tests running on Travis CI. Travis CI is the distributed, community run continuous integration server that the Ruby community has put together. It lets you do all manner of fun things, like testing in dozens of different Ruby version configurations.

You’re probably wondering what Armstrong is doing there with all of this talk of Ruby. No, I didn’t rewrite Armstrong in Rails last night. No, I didn’t convert all of our fabfiles over to Rakefiles either. Instead, I subverted it from within.

Travis CI uses a .travis.yml file for all of its configuration. There are two key fields that it gives you that let you do fun things with it: before_scripts and scripts.

before_scripts runs before anything starts. It’s like setup in the xUnit world, but for your whole environment. Each of the Armstrong components ships a requirements/dev.txt file, so I tell Travis to do a pip install -r of that during setup. That’s right, Travis CI has pip installed!

Next, I’ve set the script to use our test runner, fab test and we’re set. I had to add a few environment variables to turn off our coverage reports—they don’t provide much value when there’s no one there to view them—and we don’t need to do a re-install like we do on a local environment.

You can see this in action by checking out the current build status for the armstrong.core.arm_wells component here. Here’s the .travis.yml file’s contents:

   - 1.9.3
  - sudo pip install -r requirements/dev.txt
  - sudo pip install .
  - fab test
  email: false
    - ""

There’s work happening to bring native Python support. Native support means being able to test against multiple versions and such. Be sure to check out the #travis channel on Freenode if you’re interested in helping out.