Guest post by Petr Marek (source)

Everybody driving a car needs the navigation to get to the destination fast and avoid traffic jam. One of the biggest problems is how to enter fast the destination and how to find where are the congestions, what is the traffic situation. YodaQA Traffic is a project attempting to answer the traffic related questions quickly and efficiently. Drivers may ask questions in natural language like: “What is the traffic situation in the Evropská street?” or “What is the fastest route from Opletalova street to Kafkova street?” You can try out the prototype (demo available only for limited time) – try to ask for example “traffic situation in the Wilsonova street” .

YodaQA Traffic still has some limitations. Currently we only have a browser version not suitable for smart phones. It is answering traffic questions for Prague’s streets only.

But as usual, this whole technology demo is open source – you can find it in the branch f/traffic-flow of our Hub project.

How does it work and where we get the data from?

All YodaQA are first analyzed to recognize and select traffic questions. We do it in two steps. The first step is to recognize the question topic. We use six topics like traffic situation, traffic incident or fastest route. The topic is determined by comparing semantic similarity of the user’s question with a set of reference questions. We estimate the similarity with our Dataset-STS Scoring API. Each reference question is labeled by a “topic”. The Sentence Pair Similarity algorithm selects the reference question “topic” with the highest similarity to the question.

Next we need to recognize the location, i.e. to recognize the street name. This is handled by another tool called the Label-lookup which we normally use for entity linking in YodaQA. It compares questions words with a list of all street names in the Prague. We exported the list of streets names in Prague from OpenStreetMap. We do not do exact match, we try to select the closest street name from the list.

The last step is to decide whether the question is really the traffic question, because the Dataset-STS API and Label-lookup can find topic and street name even in a pure movie question like “When was the Nightmare on Elm Street released?”. The Dataset-STS and Label-lookup return not only topic or street name but also the score, fortunately. We created dataset of over 70 traffic questions and over 300 movies questions and founded the minimal score thresholds, with which the recognition makes the lowest classification error on this dataset.

Once we know the type of question and the location we start a small script accessing the traffic situation data from HERE Maps. The only complication is that the the API doesn’t return traffic situation for particular street, but bounding box only. To overcome this problem we have to find a bounding box for a desired location, using an algorithm we developed for this purpose. Then we call the traffic flow API to acquire the information for all streets in the bounding box. Finally, we filter out the traffic situation for the desired street.

It was great fun to work on this application, it is not perfect but it shows how to create intelligent assistants helping people solving various everyday situations. We are also excited to see, how the users will use the new functionality of YodaQA and how it will help them.

Guest post by Petr Marek (source)

Everybody driving a car needs the navigation to get to the destination fast and avoid traffic jam. One of the biggest problems is how to enter fast the destination and how to find where are the congestions, what is the traffic situation. YodaQA Traffic is a project attempting to answer the traffic related questions quickly and efficiently. Drivers may ask questions in natural language like: “What is the traffic situation in the Evropská street?” or “What is the fastest route from Opletalova street to Kafkova street?” You can try out the prototype (demo available only for limited time) – try to ask for example “traffic situation in the Wilsonova street” .

YodaQA Traffic still has some limitations. Currently we only have a browser version not suitable for smart phones. It is answering traffic questions for Pragueâ€s streets only.

But as usual, this whole technology demo is open source – you can find it in the branch f/traffic-flow of our Hub project.

How does it work and where we get the data from?

All YodaQA are first analyzed to recognize and select traffic questions. We do it in two steps. The first step is to recognize the question topic. We use six topics like traffic situation, traffic incident or fastest route. The topic is determined by comparing semantic similarity of the userâ€s question with a set of reference questions. We estimate the similarity with our Dataset-STS Scoring API. Each reference question is labeled by a “topic”. The Sentence Pair Similarity algorithm selects the reference question “topic” with the highest similarity to the question.

Next we need to recognize the location, i.e. to recognize the street name. This is handled by another tool called the Label-lookup which we normally use for entity linking in YodaQA. It compares questions words with a list of all street names in the Prague. We exported the list of streets names in Prague from OpenStreetMap. We do not do exact match, we try to select the closest street name from the list.

The last step is to decide whether the question is really the traffic question, because the Dataset-STS API and Label-lookup can find topic and street name even in a pure movie question like “When was the Nightmare on Elm Street released?”. The Dataset-STS and Label-lookup return not only topic or street name but also the score, fortunately. We created dataset of over 70 traffic questions and over 300 movies questions and founded the minimal score thresholds, with which the recognition makes the lowest classification error on this dataset.

Once we know the type of question and the location we start a small script accessing the traffic situation data from HERE Maps. The only complication is that the the API doesnâ€t return traffic situation for particular street, but bounding box only. To overcome this problem we have to find a bounding box for a desired location, using an algorithm we developed for this purpose. Then we call the traffic flow API to acquire the information for all streets in the bounding box. Finally, we filter out the traffic situation for the desired street.

It was great fun to work on this application, it is not perfect but it shows how to create intelligent assistants helping people solving various everyday situations. We are also excited to see, how the users will use the new functionality of YodaQA and how it will help them.

I have some 2” wooden blinds in my house that I’ve been wanting to motorize. Why? I’m lazy and I thought it would be cool to have.

The best commercial solution for retrofitting existing blinds seems to be Somfy. They have wireless battery-powered systems and fancy-looking remotes. For new motorized blinds, Bali seems to be popular, and they use Somfy for the motorization. There are also some kickstarter things (MOVE, MySmartBlinds), but the last time I looked those didn’t really do what I want. Somfy likely has a good product, but it’s very expensive. It looks like it would cost about $150 per blind, which is just way too much for me. They want $30 just for the plastic wand that holds the batteries (8 x AA). We’re talking about a motor and a wireless controller to tell it what to do. It’s not rocket surgery, so why should it cost $150?

My requirements are:

• Ability to tilt the blinds to one of three positions (up, middle, down) remotely via some wireless interface. I don’t care about raising or lowering the entire blind.
• There must be some API for the wireless interface such that I can automate them myself (close at night, open in morning)
• Tilt multiple blinds at the same time so they look coordinated.
• Be power efficient – one set of batteries should last more than a year.

Somfy satisfies this if I also buy their “Universal RTS Interface” for $233, but that only makes their solution even more expensive. For the 6 blinds I wanted to motorize, it would cost about $1200. No way.

I’ve been meaning to get into microcontrollers for a while now, and I thought this would be the perfect project for me to start. About a year ago I bought a RedBear BLE Nano to play with some Bluetooth stuff, so I started with that. I got a hobby servo and a bunch of other junk (resistors, capacitors, etc) from Sparkfun and began flailing around while I had some time off around Christmas. The Arduino environment on the BLE Nano is a little weird, but I got things cobbled together relatively quickly. The servo was very noisy, and it’s difficult to control the speed, but it worked. Because I wanted to control multiple devices at once, BLE was not a really great option (since AFAIK there is no way to ‘broadcast’ stuff in a way that is power-efficient for the listeners), and I started looking at other options. Eventually I ran across the Moteino.

The Moteino is an Arduino clone paired with a RFM69W wireless radio, operating at either 915Mhz or 433Mhz. It also has a very efficient voltage regulator, making it suitable for battery powered applications. The creator of the board (Felix Rusu) has put in a lot of work to create libraries for the Moteino to make it useful in exactly my type of application, so I gave it a try. The RFM69 library is lovely to work with, and I was sending messages between my two Moteinos in no time. The idea is to have one Moteino connected via USB to a Linux box (I already have a BeagleBone Black) as a base station which will relay commands to the remote devices. I got my servo working again with the Moteino quickly, as most of the code Just Worked.

I started out with a hobby servo because I knew it would be easy to control, but the noise and lack of speed control really bothered me. I needed to try something else. I considered higher quality servos, a gear motor with encoder or limit switches, stepper motors, worm gear motors, etc. I was going to end up building 6 of these things to start with, so cost was definitely a big factor. I ended up settling on the 28BYJ-48 stepper motor because it is extremely cheap (about $2), relatively quiet, and let me control the speed of rotation very precisely. There is a great Arduino library for stepper motors, AccelStepper, which lets you configure acceleration/deceleration, maximum speed, etc. It also has an easy-to-use API for positioning the motor. I found a 5mm x 8mm aluminum motor coupling to connect the motor to the blinds shaft. I then used a zip tie and a piece of rubber to secure the motor to the blinds rail. This doesn’t look very professional, but it’s not something you really see (my blinds have a valance that covers the rail). A better solution involving some kind of bracket would be great, but would increase the cost and require a lot more time. Using the stepper, I was able to smoothly, quietly, and cost-effectively control the blinds.

I then started to look into power consumption. If you don’t do put anything to sleep, the power usage is pretty high. The LowPower library from Felix makes it easy to put the CPU to sleep, which helps a lot. When sleeping, the CPU uses very little power (about 3µA I think), and the radio will wake you up via interrupt if a message arrives. The radio uses roughly 17mA in receive mode, however, so that means we’d only get about a week of battery life if we used a set of high-quality AAs (3000mAh / 17mA = 176h). We need to do a lot better.

The RFM69 has a useful feature called Listen Mode that some folks on the LowPowerLabs forums have figured out how to use. In this mode, you can configure the radio to cycle between sleeping and receiving in order to reduce power consumption. There are a lot of options here, but it was discovered that you only need to be in the RX phase for 256µS in order for a message to be detected. When the radio is asleep it uses about 4µA. So if you sleep for 1s and receive for 256µS, that means your average power consumption for the radio is about 12µA. This is a dramatic improvement, and it means that the device can still respond in roughly one second, which is certainly adequate for my application. Of course, you can always trade even more responsiveness for power efficiency, and people using this method on coin cell batteries certainly do that. There is one user on the forums who has an application with an expected battery life of over 100 years on a single coin cell! I have an extension of the RFM69 library, RFM69_WL, which collected some of the other listen mode code that was floating around and extends it so you can set your own sleep/RX durations.

I’ve measured/calculated my average power consumption to be about 46µA if I run the motor for 12s per day. That comes out to over 7 years of life on a set of 4 AA batteries (Energizer Ultimate Lithium), which is an almost unbelievable number. There are several factors I am not really considering, however, such as RF noise (which wakes the radio causing increased power consumption), so the real life performance might not be very close to this. Still, if I can get 2 years on a set of 4 AAs I’ll be pretty happy.

Usually when you buy a 28BYJ-48 it will include a driver board that has a ULN2003A, some connectors, and a set of LEDs for showing which phase of the stepper is active. This is fine for testing and development, but it was going to be pretty clunky to use this in a final solution. It was time to design my first PCB!

I found out early on that it was pretty difficult to talk to other people about problems with your project without having a schematic, so I made one of those. I started with Fritzing, but moved to EAGLE when it was time to do the PCB. It seemed to be the standard thing to use, and it was free. EAGLE has a pretty steep learning curve, but some tutorials from Sparkfun helped a lot. I also got some help from the folks on the LowPowerLabs forums (TomWS, perky), who I suspect do this kind of thing for a living. You can get the EAGLE schematic and board design here.

View post on imgur.com

View post on imgur.com

I ordered my first batch of boards from Seeed Studio, as well as a bunch of supporting components from Digikey. The PCB orders typically take a little over two weeks, which is quite a bit more waiting than I’m accustomed to. I was pretty excited when they arrived, and started checking things out. I soon realized I had made a mistake. The component I had in my design for the motor connector (which is a JST-XH 5-pin) was the wrong pitch and size, so the socket I had didn’t fit. Whoops. The hardware world does not play well with my “just try some stuff” mentality from working with software. I found an EAGLE library for the JST-XH connectors, used the correct part, and ordered another batch of PCBs. This time I actually printed out my board on paper to make sure everything matched up. I had run across PCBShopper while waiting for my first batch of boards, so I decided to use a different fabricator this time. I chose Maker Studio for the second order, since I could pay about the same amount and get red boards instead of green. Another two weeks went by, and finally last week I received the boards. I assembled one last weekend using my fancy (and cheap!) new soldering station. It didn’t work! Shit! The Moteino was working fine, but the motor wasn’t moving. Something with the motor driver or connection was hosed. After probing around for a pretty long time, I finally figured out that the socket was installed backwards. It seems the pins in the EAGLE part I found were reversed. Ugh. With a lot of hassle, I was able to unsolder the connector from the board and reverse it. The silkscreen outline doesn’t match up, but whatever. It works now, which was a big relief.

View post on imgur.com

I thought about putting the board in some kind of plastic enclosure, but it was hard to find anything small enough to fit inside the rail while also being tall enough to accomodate the Moteino on headers. I’m planning to just use some extra-wide heat shrink to protect the whole thing instead, but haven’t done that yet.

Below are some photos and videos, as well as the entire list of parts I’ve used and their prices. Each device costs about $40, which is a pretty big improvement over the commercial options (except maybe the kickstarter stuff). Also important is that there are no wires or electronics visible, which was critical for the Wife Acceptance Factor (and my own, honestly).

I’m sure a real EE will look at this and think “pfft, amateur!”. And that’s fine. My goals were to learn and have fun, and they were definitely accomplished. If I also produced something usable, that’s a bonus.

View post on imgur.com

View post on imgur.com

View post on imgur.com

Bill of Materials

In order to talk to the devices from a host computer, you’ll also need a Moteino USB ($27). To program the non-USB Moteinos you’ll need a FTDI adapter. LowPowerLabs sells one of those for $15, but you may be able to find a better deal elsewhere.

The blog has been a little bit silent – a typical sign of us working too hard to worry about that! But we’ll satisfy some of your curiosity in the coming weeks as we have about six posts in the pipeline.

The thing I would like to mention first is some fundamental research we work on now. I stepped back from my daily Question Answering churn and took a little look around and decided the right thing to focus for a while are the fundamentals of the NLP field so that our machine learning works better and makes more sense. Warning: We’ll use some scientific jargon in this one post.

So, in the first months of 2016 I focused huge chunk of my research on deep learning of natural language. That means neural networks used on unstructured text, in various forms, shapes and goals. I have set some audacious goals for myself, fell short in some aspects but still made some good progress hopefully. Here’s the deal – a lot of the current research is about processing a single sentence, maybe to classify its sentiment or translate it or generate other sentences. But I have noticed that recently, I have seen many problems that are about scoring a pair of two sentences. So I decided to look into that and try to build something that (A) works better, (B) actually has an API and we can use it anywhere for anything.

My original goal was to build awesome new neural network architectures that will turn the field on its head. But I noticed that the field is a bit of a mess – there is a lot of tasks that are about the same thing, but very little cross-talk between them. So you get a paper that improves the task of Answer Sentence Selection, but could the models do better on the Ubuntu Dialogue task then, or on Paraphrasing datasets? Who knows! Meanwhile, each dataset has its own format and a lot of time is spent only in writing the adapter code for it. Training protocols (from objectives to segmentation to embedding preinitializations) are inconsistent, and some datasets need a lot of improvement. Well, my goal turned to sorting out the field, cross-check the same models on many tasks and provide a better entry point for others than I had.

Software: Getting a few students of the 3C group together, we have created the dataset-sts platform for all tasks and models that are about comparing two sentences using deep learning. We have a pretty good coverage (of both tasks and models), and more brewing in some side branches. It’s in Python and uses the awesome Keras deep learning library.

Paper: To kick things off research-wise, we have posted a paper Sentence Pair Scoring: Towards Unified Framework for Text Comprehension where we summed up what we have learned early in the process. A few highlights:

• We have a lofty goal of building an universal text comprehension model, a sort of black box that eats your sentences and produces embeddings that correspond to their meaning, which you can use for whatever task you need to do. Long way to go, but we have found that a simple neural model trained on very large data is doing pretty good in this exact setting, and even if applied to tasks and data that look very different from the original. Maybe we are on to something.
• Our framework is state-of-art on the Ubuntu Dialogue dataset of 1M techsupport IRC dialogs, beating Facebook’s memory network models.
• It’s hard to compare neural models because if you train a model 16 times with the same data, the result will always be somewhat different. Not a big deal with large test datasets, but a very big deal with small test datasets which are still popular in the research community. Almost all papers ignore this! If you look at evolution of performance of models in some areas like Answer Sentence Selection, we have found that most differences over the last year are deep below per-train variance we see.

Please take a look, and tell us what you think! We’ll shortly cover a follow-up paper here that we also already posted, and we plan to continue the work by improving our task and model coverage further, fixing a few issues with our training process and experimenting with some novel neural network ideas.

More to come, both about our research and some more product-related news, in a few days. We will also talk about how the abstract-sounding research connects with some very practical technology we are introducing.

I’ve been running my personal blog on rootco.de for a few months now. The server is a minimal install of openSUSE Leap 42.1 running on a nice physical machine hosted at the awesome Hetzner, who offer openSUSE Leap as an OS on all of their Physical and Virtual server hosting. I use the standard Apache available in Leap, with Jekyll to generate this blog. You can actually see the source to this Jekyll blog on GitHub. And to manage it all I use the awesome SaltStack and keep all of my Salt configuration in GitHub also so you can see exactly how my system is setup.

Why am I sharing all of this? Well this weekend there was something I needed to fix.
http://rootco.de was running without HTTPS.

So What?

This site is a blog about Free Software & Open Source stuff, why on earth does it need to be running HTTPS?.

Because every single web service that can be HTTPS, should be HTTPS. There are lots of good articles going back years as to why, but the simplest reasons is that it helps ensure the content you visit when you go to my blog is the content I intended for my blog. It’s very hard for someone to tamper with the content delivered from a HTTPS website. While I’m not (yet) currently hosting any interactive services on my server, if I do I want to ensure they’re secured by HTTPS so the data I’m sending to my server is done so as securely as possible.

And in this day and age, there is rarely an excuse to not use HTTPS for everything. Certificates used to be expensive and complicated to setup, but thanks to the wonderful project LetsEncrypt anyone can now get certificates for their domains for FREE.

Getting Stated with LetsEncrypt

I started as anyone should, by reading the Getting Started Guide.
As there is not (yet) a certbot package for openSUSE Leap, I had to use the certbot-auto wrapper script.
The documentation recommends you install it using the following commands:

$ git clone https://github.com/certbot/certbot
$ cd certbot
$ ./certbot-auto --help

As I’m actually using SaltStack to manage my system, all I did instead was add the following to my Salt State for the rootco.de Web Server.

certbot:
  git.latest:
    - name: https://github.com/certbot/certbot
    - target: /opt/certbot
    - user: root

You can see the git commit HERE.

I then ran the following on my salt master to tell SaltStack to pull down the changes and apply them to the rootco.de Web Server.

$ git -C /srv/salt pull 
$ salt 'luke.rootco.de' state.highstate --state-output=changes

NOTE: I could have just waited, I actually have the above running as a cronjob every 30 minutes to make sure my server configuration stays as I have defined it in SaltStack

I then sanity checked the contents of /opt/certbot before proceeding. I really hate randomly downloading code from GitHub, so I spent a bit of time making sure what I downloaded made sense and matched what I expected, while wishing someone would take the time to package this up on the openSUSE Build Service so I could trust them and stop worrying. Once I was happy, I ran the following command to request a certificate for rootco.de and www.rootco.de:

$ /opt/certbot/certbot-auto certonly --webroot -w /srv/www/htdocs \
-d rootco.de -d www.rootco.de

The wizard automatically detected I was running openSUSE, installed a few packages it needed, then asked me for an email address, and that was it! I had my certificate created and on my server at /etc/letsencrypt/live/rootco.de/fullchain.pem. I followed the advice to backup /etc/letsencrypt as it contains lots of important configuration and the certificates/keys for my system. Now I had to get Apache to actually use the certificate.

Configuring Apache on Leap for LetsEncrypt

Because I was a little rusty, I reminded myself of the openSUSE Leap Apache Documentation. Good thing too, because I can completely forgotten that to tell Apache to use SSL you needed to run the following command:

$ a2enflag SSL

With that set, I went about setting up an Apache vhost configuration for SSL on rootco.de:

<VirtualHost _default_:443>
	DocumentRoot "/srv/www/htdocs"
	ErrorLog /var/log/apache2/error_log
	TransferLog /var/log/apache2/access_log
	SSLEngine on
	# Path to the LetsEncrypt created certificate fullchain.pem
	SSLCertificateFile /etc/letsencrypt/live/rootco.de/fullchain.pem 
	# Path to the LetsEncrypt created private key privkey.pem
	SSLCertificateKeyFile /etc/letsencrypt/live/rootco.de/privkey.pem
	CustomLog /var/log/apache2/ssl_request_log   ssl_combined
</VirtualHost>

Now, because I absolutely hate making any change to the rootco.de Server directly and want everything managed by SaltStack, I actually put this in my Salt States folder, and modified the rootco.de Web Server state to automatically deploy the file to the appropriate place on the server. You can see the git commit for that HERE. As I am impaitent and didn’t want to wait for my automatic deployment, I again manually refreshed the Salt States on my master and used salt to deploy this new configuration:

$ git -C /srv/salt pull 
$ salt 'luke.rootco.de.' state.highstate --state-output=changes

A quick systemctl restart apache2 on the server later and I was in business - https://rootco.de was live!

That’s great but…

LetsEncrypt certificates have a duration of 90 days. I want https://rootco.de to be running for a lot longer than that, so I needed to find a solution.

The LetsEncrypt Documentation talked about a renew function so I gave it a quick try:

$ /opt/certbot/certbot-auto renew --dry-run

This seemed to work fine so I added a simple cron job to run the following every 60 days:

$ /opt/certbot/certbot-auto renew >/dev/null 2>&1

I picked 60 days as LetsEncrypt certificates only let you renew 30 days before expiration, and I want to hit their servers as little as possible while still ensuring the certificate always gets updated before it expires. >/dev/null 2>&1 is there because I really don’t care about the logs - if it works, I will never need to look at it, and if it’s broken I’m going to have to be running the command manually anyway to figure out what went wrong.

Doing a cronjob in SaltStack is so easy, I used it rather than editing the crontab for root myself.

So now my server is running with HTTPS, with a nice shiny LetsEncrypt certificate for both https://rootco.de and https://www.rootco.de, and the whole thing will auto renew ever 60 days. And if something goes horribly wrong and my server gets messed up, all of this is easily redeployable using SaltStack, which is a nice extra bonus for me.

One last thing

LetsEncrypt is awesome. This service is really revolutionary and is the sort of thing which shouldn’t be taken for granted, so please help them out by Donating to LetsEncrypt.

I recently wrote a bigger project in the D programming language, the appstream-generator (asgen). Since I rarely leave the C/C++/Python realm, and came to like many aspects of D, I thought blogging about my experience could be useful for people considering to use D.

Disclaimer: I am not an expert on programming language design, and this is not universally valid criticism of D – just my personal opinion from building one project with it.

Why choose D in the first place?

The previous AppStream generator was written in Python, which wasn’t ideal for the task for multiple reasons, most notably multiprocessing and LMDB not working well together (and in general, multiprocessing being terrible to work with) and the need to reimplement some already existing C code in Python again.

So, I wanted a compiled language which would work well together with the existing C code in libappstream. Using C was an option, but my least favourite one (writing this in C would have been much more cumbersome). I looked at Go and Rust and wrote some small programs performing basic operations that I needed for asgen, to get a feeling for the language. Interfacing C code with Go was relatively hard – since libappstream is a GObject-based C library, I expected to be able to auto-generate Go bindings from the GIR, but there were only few outdated projects available which did that. Rust on the other hand required the most time in learning it, and since I only briefly looked into it, I still can’t write Rust code without having the coding reference open. I started to implement the same examples in D just for fun, as I didn’t plan to use D (I was aiming at Go back then), but the language looked interesting. The D language had the huge advantage of being very familiar to me as a C/C++ programmer, while also having a rich standard library, which included great stuff like std.concurrency.Generator, std.parallelism, etc. Translating Python code into D was incredibly easy, additionally a gir-d-generator which is actively maintained exists (I created a small fork anyway, to be able to directly link against the libappstream library, instead of dynamically loading it).

What is great about D?

This list is just a huge braindump of things I had on my mind at the time of writing

Interfacing with C

There are multiple things which make D awesome, for example interfacing with C code – and to a limited degree with C++ code – is really easy. Also, working with functions from C in D feels natural. Take these C functions imported into D:

extern(C):
nothrow:

struct _mystruct {}
alias mystruct_p = _mystruct*;

mystruct_p = mystruct_create ();
mystruct_load_file (mystruct_p my, const(char) *filename);
mystruct_free (mystruct_p my);

You can call them from D code in two ways:

auto test = mystruct_create ();
// treating "test" as function parameter
mystruct_load_file (test, "/tmp/example");
// treating the function as member of "test"
test.mystruct_load_file ("/tmp/example");
test.mystruct_free ();

This allows writing logically sane code, in case the C functions can really be considered member functions of the struct they are acting on. This property of the language is a general concept, so a function which takes a `string` as first parameter, can also be called like a member function of `string`.

Writing D bindings to existing C code is also really simple, and can even be automatized using tools like dstep. Since D can also easily export C functions, calling D code from C is also possible.

Getting rid of C++ “cruft”

There are many things which are bad in C++, some of which are inherited from C. D kills pretty much all of the stuff I found annoying. Some cool stuff from D is now in C++ as well, which makes this point a bit less strong, but it’s still valid. E.g. getting rid of the `#include` preprocessor dance by using symbolic import statements makes sense, and there have IMHO been huge improvements over C++ when it comes to metaprogramming.

Incredibly powerful metaprogramming

Getting into detail about that would take way too long, but the metaprogramming abilities of D must be mentioned. You can do pretty much anything at compiletime, for example compiling regular expressions to make them run faster at runtime, or mixing in additional code from string constants. The template system is also very well thought out, and never caused me headaches as much as C++ sometimes manages to do.

Built-in unit-test support

Unittesting with D is really easy: You just add one or more `unittest { }` blocks to your code, in which you write your tests. When running the tests, the D compiler will collect the unittest blocks and build a test application out of them.

The `unittest` scope is useful, because you can keep the actual code and the tests close together, and it encourages writing tests and keep them up-to-date. Additionally, D has built-in support for contract programming, which helps to further reduce bugs by validating input/output.

Safe D

While D gives you the whole power of a low-level system programming language, it also allows you to write safer code and have the compiler check for that, while still being able to use unsafe functions when needed.

Unfortunately, `@safe` is not the default for functions though.

Separate operators for addition and concatenation

D exclusively uses the `+` operator for addition, while the `~` operator is used for concatenation. This is likely a personal quirk, but I love it very much that this distinction exists. It’s nice for things like addition of two vectors vs. concatenation of vectors, and makes the whole language much more precise in its meaning.

Optional garbage collector

D has an optional garbage collector. Developing in D without GC is currently a bit cumbersome, but these issues are being addressed. If you can live with a GC though, having it active makes programming much easier.

Built-in documentation generator

This is almost granted for most new languages, but still something I want to mention: Ddoc is a standard tool to generate code documentation for D code, with a defined syntax for describing function parameters, classes, etc. It will even take the contents of a `unittest { }` scope to generate automatic examples for the usage of a function, which is pretty cool.

Scope blocks

The `scope` statement allows one to execute a bit of code before the function exists, when it failed or was successful. This is incredibly useful when working with C code, where a free statement needs to be issued when the function is exited, or some arbitrary cleanup needs to be performed on error. Yes, we do have smart pointers in C++ and – with some GCC/Clang extensions – a similar feature in C too. But the scopes concept in D is much more powerful. See Scope Guard Statement for details.

Built-in syntax for parallel programming

Working with threads is so much more fun in D compared to C! I recommend taking a look at the parallelism chapter of the “Programming in D” book.

“Pure” functions

D allows to mark functions as purely-functional, which allows the compiler to do optimizations on them, e.g. cache their return value. See pure-functions.

D is fast!

D matches the speed of C++ in almost all occasions, so you won’t lose performance when writing D code – that is, unless you have the GC run often in a threaded environment.

Very active and friendly community

The D community is very active and friendly – so far I only had good experience, and I basically came into the community asking some tough questions regarding distro-integration and ABI stability of D. The D community is very enthusiastic about pushing D and especially the metaprogramming features of D to its limits, and consists of very knowledgeable people. Most discussion happens at the forums/newsgroups at forum.dlang.org.

What is bad about D?

Half-proprietary reference compiler

This is probably the biggest issue. Not because the proprietary compiler is bad per se, but because of the implications this has for the D ecosystem.

For the reference D compiler, Digital Mars’ D (DMD), only the frontend is distributed under a free license (Boost), while the backend is proprietary. The FLOSS frontend is what the free compilers, LLVM D Compiler (LDC) and GNU D Compiler (GDC) are based on. But since DMD is the reference compiler, most features land there first, and the Phobos standard library and druntime is tuned to work with DMD first.

Since major Linux distributions can’t ship with DMD, and the free compilers GDC and LDC lack behind DMD in terms of language, runtime and standard-library compatibility, this creates a split world of code that compiles with LDC, GDC or DMD, but never with all D compilers due to it relying on features not yet in e.g. GDCs Phobos.

Especially for Linux distributions, there is no way to say “use this compiler to get the best and latest D compatibility”. Additionally, if people can’t simply `apt install latest-d`, they are less likely to try the language. This is probably mainly an issue on Linux, but since Linux is the place where web applications are usually written and people are likely to try out new languages, it’s really bad that the proprietary reference compiler is hurting D adoption in that way.

That being said, I want to make clear DMD is a great compiler, which is very fast and build efficient code. I only criticise the fact that it is the language reference compiler.

UPDATE: To clarify the half-proprietary nature of the compiler, let me quote the D FAQ:

The front end for the dmd D compiler is open source. The back end for dmd is licensed from Symantec, and is not compatible with open-source licenses such as the GPL. Nonetheless, the complete source comes with the compiler, and all development takes place publically on github. Compilers using the DMD front end and the GCC and LLVM open source backends are also available. The runtime library is completely open source using the Boost License 1.0. The gdc and ldc D compilers are completely open sourced.

Phobos (standard library) is deprecating features too quickly

This basically goes hand in hand with the compiler issue mentioned above. Each D compiler ships its own version of Phobos, which it was tested against. For GDC, which I used to compile my code due to LDC having bugs at that time, this means that it is shipping with a very outdated copy of Phobos. Due to the rapid evolution of Phobos, this meant that the documentation of Phobos and the actual code I was working with were not always in sync, leading to many frustrating experiences.

Furthermore, Phobos is sometimes removing deprecated bits about a year after they have been deprecated. Together with the older-Phobos situation, you might find yourself in a place where a feature was dropped, but the cool replacement is not yet available. Or you are unable to import some 3rd-party code because it uses some deprecated-and-removed feature internally. Or you are unable to use other code, because it was developed with a D compiler shipping with a newer Phobos.

This is really annoying, and probably the biggest source of unhappiness I had while working with D – especially the documentation not matching the actual code is a bad experience for someone new to the language.

Incomplete free compilers with varying degrees of maturity

LDC and GDC have bugs, and for someone new to the language it’s not clear which one to choose. Both LDC and GDC have their own issues at time, but they are rapidly getting better, and I only encountered some actual compiler bugs in LDC (GDC worked fine, but with an incredibly out-of-date Phobos). All issues are fixed meanwhile, but this was a frustrating experience. Some clear advice or explanation which of the free compilers is to prefer when you are new to D would be neat.

For GDC in particular, being developed outside of the main GCC project is likely a problem, because distributors need to manually add it to their GCC packaging, instead of having it readily available. I assume this is due to the DRuntime/Phobos not being subjected to the FSF CLA, but I can’t actually say anything substantial about this issue. Debian adds GDC to its GCC packaging, but e.g. Fedora does not do that.

No ABI compatibility

D has a defined ABI – too bad that in reality, the compilers are not interoperable. A binary compiled with GDC can’t call a library compiled with LDC or DMD. GDC actually doesn’t even support building shared libraries yet. For distributions, this is quite terrible, because it means that there must be one default D compiler, without any exception, and that users also need to use that specific compiler to link against distribution-provided D libraries. The different runtimes per compiler complicate that problem further.

The D package manager, dub, does not yet play well with distro packaging

This is an issue that is important to me, since I want my software to be easily packageable by Linux distributions. The issues causing packaging to be hard are reported as dub issue #838 and issue #839, with quite positive feedback so far, so this might soon be solved.

The GC is sometimes an issue

The garbage collector in D is quite dated (according to their own docs) and is currently being reworked. While working with asgen, which is a program creating a large amount of interconnected data structures in a threaded environment, I realized that the GC is significantly slowing down the application when threads are used (it also seems to use UNIX signals `SIGUSR1` and `SIGUSR2` to stop/resume threads, which I still find odd). Also, the GC performed poorly on memory pressure, which did get asgen killed by the OOM killer on some more memory-constrained machines. Triggering a manual collection run after a large amount of these interconnected data structures wasn’t needed anymore solved this problem for most systems, but it would of course have been better to not needing to give the GC any hints. The stop-the-world behavior isn’t a problem for asgen, but it might be for other applications.

These issues are at time being worked on, with a GSoC project laying the foundation for further GC improvements.

“version” is a reserved word

Okay, that is admittedly a very tiny nitpick, but when developing an app which works with packages and versions, it’s slightly annoying. The `version` keyword is used for conditional compilation, and needing to abbreviate it to `ver` in all parts of the code sucks a little (e.g. the “Package” interface can’t have a property “version”, but now has “ver” instead).

The ecosystem is not (yet) mature

In general it can be said that the D ecosystem, while existing for almost 9 years, is not yet that mature. There are various quirks you have to deal with when working with D code on Linux. It’s always nothing major, usually you can easily solve these issues and go on, but it’s annoying to have these papercuts.

This is not something which can be resolved by D itself, this point will solve itself as more people start to use D and D support in Linux distributions gets more polished.

Conclusion

I like to work with D, and I consider it to be a great language – the quirks it has in its toolchain are not that bad to prevent writing great things with it.

At time, if I am not writing a shared library or something which uses much existing C++ code, I would prefer D for that task. If a garbage collector is a problem (e.g. for some real-time applications, or when the target architecture can’t run a GC), I would not recommend to use D. Rust seems to be the much better choice then.

In any case, D’s flat learning curve (for C/C++ people) paired with the smart choices taken in language design, the powerful metaprogramming, the rich standard library and helpful community makes it great to try out and to develop software for scenarios where you would otherwise choose C++ or Java. Quite honestly, I think D could be a great language for tasks where you would usually choose Python, Java or C++, and I am seriously considering to replace quite some Python code with D code. For very low-level stuff, C is IMHO still the better choice.

As always, choosing the right programming language is only 50% technical aspects, and 50% personal taste

UPDATE: To get some idea of D, check out the D tour on the new website tour.dlang.org.

With the Open Source Event Manager (OSEM), one of the Ruby on Rails apps I hack on, we're heading down the road to version 1.0. A feature we absolutely wanted to have, before making this step, was an easy deployment to at least one of the many Platform as a Service (PaaS) providers. We deemed this important for two reasons:

1. Before people commit to use your precious little app they want to try it. And this has to be hassle free.
2. Getting your server operating system ready to deploy Ruby on Rails applications can be tedious, too tedious for some people.

So I have been working on making our app ready for Heroku which is currently the most popular PaaS provider for rails (is it?). This was an interesting road, this is my travelogue.

Storage in the public cloud

Storing files is incredibly easy in Rails, there are many ready made solutions for it like paperclip, dragonfly or carrierwave. The challenge with Heroku is, that on their free plan, your apps storage will be discarded the moment it is stopped or restarted. This happens for instance every time during a deployment or when the app goes to sleep because nobody is using it.

And even though it's easy to store files in Rails, we in the OSEM team have long discouraged this in our app. We rather try to make it as easy as possible to reference things you have shared somewhere else. Want to show a picture of your Events location? Use the ones you share on flickr or instagram anyway. Embed a video of a Talk? Just paste the Youtube or Vimeo link. Share the slides with your audience? Slideshare or Speakerdeck to the rescue!

OSEM commercials by Henne Vogelsang licensed CC BY 4.0

Still there are some places left in our app where we upload pictures. Pictures we think conference organizers are not necessarily free to share on other platforms, like sponsor logos or pictures of sponsored lodgings. So to be able to use OSEM on Heroku our file upload needed support for offloading the files to someone else's computer a.k.a. *the cloud*. In the end I have settled for the carrierwave plug in of cloudinary (pull request #970).

This means it's now as easy as configuring the cloudinary gem and making use of their free plan to shove off all the storage OSEM needs to them.

Storage was the first gap in OSEM that I closed, another piece of the puzzle was configuration.

Configuration in the environment

According to some clever people it's too easy to mistakenly check in your apps configuration file into your version control system. That's the reason your apps environment should provide all the settings. I'm not a big fan of putting stickers onto microwaves that say you can't use it to dry your cat. If people want to be stupid, let them.

But hey 12-Factor is a thing, let's roll with it! So in pull request #900 I removed all traces of OSEMs YAML configuration file and now all the settings happen in environment variables.

This was the second gap I had to close to be able to run our app on Heroku. Now all of the sudden things where falling into place and some very interesting things emerged for OSEM.

Continuous Deployment

One of the reasons I have gone down this road was to make it easy for people to try out OSEM. Now once you can run your app on Heroku you can also integrate your github repository and run a continuous deployment (every commit get's deployed right away). That made it possible for us to set up an OSEM demo instance for people to try which always runs the latest code. All we had to do is to make use of data from our test suite to populate the demo (pull request #982) and voila...

OSEM demo by Henne Vogelsang licensed CC BY 4.0

Continuous Review

So a continuous deployment of the latest code with some sample data. Does that sound useful for you as free software developer? I think being a free software developer first and foremost means collaborating with varying people from all over. People I work with on a daily basis and people who I never had contact with before. Collaboration mostly means reviewing each others changes to our code. It's pretty easy as we have rules and tools for that in place. What is not so easy is to do the same for changes to the functionality, the user experience design, of our app.

In the OSEM team we some times attach a series of screenshots or animated gifs to a github pull requests to convey changes to the user interaction and work flows, but this is usually no replacement for trying yourself. Then in the middle of me doing all of this Heroku Review Apps happened. Review apps are instant, disposable deployments of your app that spin up automatically with each pull request.

OSEM heroku pipeline by Henne Vogelsang licensed CC BY 4.0

Now, once someone sends a pull request where I want to review the user experience design I just press the 'Create Review App' button on Heroku and a minute later I get the a temporary instance populated with test data. Magic.

More things to come?

Another thing we might want to replace in the future is how we spin up developer instances. So far we use Vagrant which starts your OSEM checkout in a virtual machine. But nowadays you have to have docker containers in the mix right? Let's see what the future brings.

All in all, I must say, it was a nice trip into the Platform as a Service world. Surprisingly easy to do and even more surprisingly rewarding for the development work flow. What do you think?