Using OpenNLP for Named-Entity-Recognition in Scala

A common challenge in Natural Language Processing (NLP) is Named Entity Recognition (NER) - this is the process of extracting specific pieces of data from a body of text, commonly people, places and organisations (for example trying to extract the name of all people mentioned in a wikipedia article). NER is a problem that has been tackled many times over the evolution of NLP, from dictionary based, to rule based, to statistical models and more recently using Neural Nets to solve the problem.

Whilst there have been recent attempts to crack the problem without it, the crux of the issue is really that for approach to learn it needs a large corpus of marked up training data (there are some marked up corpora available, but the problem is still quite domain specific, so training on the WSJ data might not perform particularly well against your domain specific data) and finding a set of 100,000 marked up sentences is no easy feat.  There are some approaches that can be used to tackle this by generating training data - but it can be hard to generate truly representative data and so this approach always risks over-fitting to the generated data.

Having previously looked at Stanford's NLP library for some sentiment analysis, this time I am looking at using the OpenNLP library. Stanford's library is often referred to as the benchmark for several NLP problems, however, these benchmarks are always against the data it is trained for - so out of the box, we likely won't get amazing results against a custom dataset. Further to this, the Stanford library is licensed under GPL which makes it harder to use in any kind of commercial/startup setting. The OpenNLP library has been around for several years, but one of its strengths is its API - its pretty well documented to get up and running, and is all very extendable.

Training a custom NER

Once again, for this exercise we are going back to the BBC recipe archive for the source data - we are going to try and train an OpenNLP model that can identify ingredients.

To train the model we need some example sentences - they recommend at least 15,000 marked up sentences to train a model - so for this I annotated a bunch of the recipe steps and ended up with somewhere in the region of about 45,000 sentences.

As you can see in the above example, the marked up sentences are quite straight forward. We just wrap the ingredient in the tags as above (although note that if the word itself isn't padded by a space either side inside the tags, it will fail!).

Once we have our training data, we can just easily setup some code to feed it in and train our model:

This is a very simple example of how you can do it, and not always paying attention to engineering best practices, but you get the idea for whats going on. We are getting an input stream of our training data set, then we instantiate the Maximum Entropy name finder class and ask it to train a model, which we can then write to disk for future use.

When we want to use the model, we can simply load it back into the OpenNLP Name Finder class and use that to parse the input text we want to check:

So, once I had created some training data in the required format, and trained a model I wanted to see how well it had actually worked - obviously, I don't want to run it against one of the original recipes as they were used to train the model, so I selected this recipe for rosemary-caramel millionaire shortbread, to see how it performed, here are the ingredients it found:

  • butter
  • sugar
  • rosemary
  • caramel
  • shortbread

All in all, pretty good - it missed some ingredients, but given the training data was created in about 20 minutes just manipulating the original recipe set with some groovy, that's to be expected really, but it did well in not returning false positives.

In conclusion, if you have a decent training set, or have the means to generate some data with a decent range, you can get some pretty good results using the library. As usual, the code for the project is on GitHub (although it is little more than the code shown in this post).

Turn your GitHub Page into a Personalised Resume

A little while ago, I decided I wanted to update my CV, and figured given I was in tech it made sense for my CV to be online.  I was aware of GitHub Pages - which give you a nice looking URL which seemed like a perfect location for my tech CV.

Once I had it looking pretty decent, and updated to modern Bootstrap styling so it was fully responsive, I thought I would stick it on GitHub, as a GitHub page.  GitHub provides support for everyone to have a free hosted page with normal HTML/JS resources etc (which is pretty nice of them!) and gives you a nice, share-able URL like http://{username}

Whilst I was reading about GitHub pages, I noticed that they have native support for Jekyll - which is a static HTML generator tool for building websites - which is when I had my second realisation of the day - I could make my CV open-source-able by making it a configurable Jekyll projects that lets users just update some config in their GitHub account and hey-presto, have a nicely styled, personalised tech CV!

So I started porting it over to Jekyll: which just involved moving the configurable, user specific items into a config file (_config.yml) and then breaking the HTML sections into fragments to make it more manageable to understand what is going on.  The idea of Jekyll is pretty straight forward - its just a simple tokenised/template approach to putting together static HTML output, but it does work well and I really didn't find myself wanting for anything in the process.  The GitHub native support was also really nice, all I needed to do was just upload the source of the project to my GitHub account and GitHub handled the build and serving of the site out of the box!

And that's all the configuration it takes! The YAML format is pretty readable - it largely just works with indenting, and hopefully taking a look over the nested sections of data, its fairly easy to understand how you can modify parts to make it customise-able.

You can see my GitHub CV page here - Out of the box you can configure lots of aspects: custom text blocks, key skills, blogs, apps, github projects, stackoverflow, etc.

How can you have a custom GitHub CV?

It really is super simple to get your own GitHub CV:
  1. Create a Github account (if you don't already have one)

  2. Go to the project repository and fork the repository

  3. Change the name of the repository (in the settings menu) to {{yourusername}}

  4. Edit the /_config.yml file in your repository - it should be pretty straight forward as to what the links/details are that you need to add.

  5. Visit your new profile page: {{yourusername}} and start sharing it!

Unsupervised Learning in Scala using Word2Vec

A pretty cool thing that has come out of recent Machine Learning advancements is the idea of "Word Embedding", specifically the advancements in the field made by Tomas Mikolov and his team at Google with the Word2Vec approach. Word Embedding is a language modelling approach that involves mapping words to vectors of numbers - If you imagine we are modelling every word in a given body of text to an N-dimension vector (it might be easier to visualise this as 2-dimensions - so each word is a pair of co-ordinates that can be plot on a graph), then that could be useful in plotting words and starting to understand relationships between words given their proximity. What's more, if we could map words to sets of numbers, then we could start thinking about interesting arithmetic that we could perform on the words.

Sounds cool, right? Now of course, the tricky bit is how can you convert a word to a vector of numbers in such a way that it encapsulates the details behind this relationship? And how can we do it without painstaking manual work and trying to somehow indicate semantic relationships and meaning in the words?

Unsupervised Learning

Word2Vec relies on neural networks and trains on a large, un-labelled piece of text in a technique known as "unsupervised" learning.

Contrary to the last neural network I discussed which was a "supervised" exercise (e.g. for every input record we had the expected output/answer), Word2Vec uses a completely "unsupervised" approach - in other words, the neural network simply takes a massive block of text with no markup or labels (broken into sentences or lines usually) and then uses that to train itself.

This kind of unsupervised learning can seem a little unbelievable at first, getting your head around the idea that a network could train itself without even knowing the "answers" seemed a little strange to me first time I heard the concept, especially as a fundamental requirement for a NN to converge on optimum solution requires a "cost-function" (e.g. some thing we can use after each feed-forward step to tell us how right we are, and if our NN is heading in the right direction).

But really, if we think back to the literal biological comparison with the brain, as people we learn through this unsupervised approach all the time - its basically trial-and-error.

It's child's play

Imagine a toddler attempting to learn to use a smart phone or tablet: they likely don't get shown explicitly to press an icon, or to swipe to unlock, but they might try combinations of power buttons, volume controls and swiping and seeing what happens (and if it does what they are ultimately trying to do), and they get feedback from the device - not direct feedback about what the correct gesture is, or how wrong they were, just the feedback that it doesn't do what they want - and if you have ever lived with a toddler who has got to grips with touchscreens, you may have noticed that when they then experience a TV or laptop, they instinctively attempt to touch or swipe the things on the screen that they want (in NN terms this would be known as "over fitting" - they have trained on too specific a set of data, so are poor at generalising - luckily, the introduction of a non-touch screen such as a TV expands their training set and they continue to improve their NN, getting better at generalising!)

So, this is basically how Word2Vec works. Which is pretty amazing if you think about it (well, I think its neat).

Word2Vec approaches

So how does this apply to Word2Vec? Well just like a smartphone gives implicit, in-direct feedback to a toddler, so the input data can provide feedback to itself. There are broadly two techniques when training the network:

Continuous Bag of Words (CBOW)

So, our NN has a large body of text broken up into sentences/lines - and just like in our last NN example, we take the first row from the training set, but we don't just take the whole sentence to push into the NN (after all, the sentence will be variable length, which would confuse our input neurons), instead we take a set number of words - referred to as the "window size", let's say 5, and feed those into the network. In this approach, the goal is for the NN to try and correctly guess the middle word in that window - that is, given a phrase of 5 words, the NN attempts to guess the word at position 3.

[It was ___ of those] days, not much to do

So its unsupervised learning, as we haven't had to go through any data and label things, or do any additional pre-processing - we can simply feed in any large body of text and it can just try to guess the words given their context.


The Skip-gram approach is similar, but the inverse - that is, given the word at position n, it attempts to guess the words at position n-2, n-1, n+1, n+2.

[__ ___ one __ _____] days, not much to do

The network is trying to work out which word(s) are missing, and just looks to the data itself to see if it can guess it correctly.

Word2Vec with DeepLearning4J

So one popular deep-learning & word2vec implementation on the JVM is DeepLearning4J. It is pretty simple to use to get used to what is going on, and is pretty well documented (along with some good high-level overviews of some core topics). You can get up and running playing with the library and some example datasets pretty quickly following their guide. Their NN setup is also equally simple and worth playing with, their MNIST hello-world tutorial lets you get up and running with that dataset pretty quickly.


A little while ago, I wrote a web crawler for the BBC food recipe archive, so I happened to have several thousand recipes sitting around and thought it might be fun to feed those recipes into Word2Vec to see if it could give any interesting results or if it was any good at recommending food pairings based on the semantic features the Word2Vec NN extracts from the data.

The first thing I tried was just using the ingredient list as a sentence - hoping that it would be better for extracting the relationship between ingredients, with each complete list of ingredients being input as a sentence.  My hope was that if I queried the trained model for X is to Beef, as Rosemary is to Lamb, I would start to get some interesting results - or at least be able to enter an ingredient and get similar ingredients to help identify possible substitutions.

As you can see, it has managed to extract some meaning from the data - for both pork and lamb, the nearest words do seem to be related to the target word, but not so much that could really be useful. Although this in itself is pretty exciting - it has taken an un-labelled body of text and has been able to learn some pretty accurate relationships between words.

Actually, on reflection, a list of ingredients isn't actually that great an input, as it isn't a natural structure and there is no natural ordering of the words - a lot of meaning is captured in the phrases rather than just lists of words.

So next up, I used the instructions for the recipes - each step in the recipe became a sentence for input, and minimal cleanup was needed, however, with some basic tweaking (it's fairly possible that if I played more with the Word2Vec configuration I could have got some improved results) the results weren't really that much better, and for the same lamb & pork search this was the output:

Again, its still impressive to see that some meaning has been found from these words, is it better than raw ingredient list? I think not - the pork one seems wrong, as it seems to have very much aligned pork as a poultry (although maybe that is some meaningful insight that conventional wisdom just hasn't taught us yet!?)


Whilst this is pretty cool, there is further fun that can be had - in the form of simple arithmetic. A simple, often quoted example, is the case of countries and their capital cities - well trained Word2Vec models have countries and their capital cities equal distances apart:

(graph taken from DeepLearning4J Word2Vec intro)

So could we extract similar relationships between food stuffs?  The short answer, with the models trained so far, was kind of..

Word2Vec supports the idea of positive and negative matches when looking for nearest words - that allows you to find these kind of relationships. So what we are looking for is something like "X is to Lamb, as thigh is to chicken" (e.g. hopefully this should find a part of the lamb), and hopefully use this to extract further information about ingredient relationships that could be useful in thinking about food.

So, I ran that arithmetic against my two models.
The instructions based model returned the following output:

Which is a pretty good effort - I think if I had to name a lamb equivalent of chicken thigh, a lamb shank is probably what I would have gone for (top of the leg, both pieces of slow twitch muscle and both the more game-y, flavourful pieces of the animal - I will stop as we are getting into food-nerd territory).

I also ran the same query on the ingredients based set (which remember, ran better on the basic nearest words test):

Which interestingly, doesn't seem as good. It has the shin, which isn't too bad in as far as its the leg of the animals, but not quite as good a match as the previous.

Let us play

Once you have the input data, Word2Vec is super easy to get up and running. As always, the code is on GitHub if you want to see the build stuff (I did have to fudge some dependencies and exclude some stuff to get it running on Ubuntu - you may get errors relating to javacpp or jnind4j not available - but the build file has the required work arounds in to get that running), but the interesting bit is as follows:
If we run through what we are setting up here:

  1. Stop words - these are words we know we want to ignore -  I originally ruled these out as I didn't want measurements of ingredients to take too much meaning. 
  2. Line iterator and tokenizer - these are just core DL4J classes that will take care of processing the text line by line, word by word. This makes things much easier for us, so we don't have to worry about that stuff
  3. Min word frequency - this is the threshold for words to be interesting to us - if a word appears less than this number of times in the text then we don't include the mapping (as we aren't confident we have a strong enough signal for it)
  4. Iterations - how many training cycles are we going to loop for
  5. Layer size - this is the size of the vector that we will produce for each word - in this case we are saying we want to map each word to a 300 dimension vector, you can consider each vector a "feature" of the word that is being learnt, this is a part of the network that will really need to be tuned to each specific problem
  6. Seed - this is just used to "seed" the random numbers used in the network setup, setting this helps us get more repeatable results
  7. Window size - this is the number of words to use as input to our NN each time - relates to the CBOW/Skip-gram approaches described above.

And that's all you need to really get your first Word2Vec model up and running! So find some interesting data, load it in and start seeing what interesting stuff you can find.

So go have fun - try and find some interesting data sets of text stuff you can feed in and what you can work out about the relationships - and feel free to comment here with anything interesting you find.

Machine (re)learning: Neural Networks from scratch

Having recently changed roles, I am now in the enviable position of starting to do some work with machine learning (ML) tools. My undergraduate degree was actually in Artificial Intelligence, but that was over a decade ago, which is a long time in computer science in general, let alone the field of Machine Learning and AI which has progressed massively in the last few years.

So needless to say, coming back to it there has been a lot to learn. In the future I will write a bit more about some of the available tools and libraries that exist these days (both expanding on the traditional AI Stanford libraries I have mentioned previously with my tweet sentiment analysis, plus newer frameworks that cover "deep learning").

Anyway, inspired by this post, I thought it would be a fun Sunday night refresher to write my own neural network. The first, and last, time that I wrote a neural network was for my final year dissertation (and that code is long gone), so was writing from first principals. The rest of this post will be a very straight forward introduction to the ideas and the code for a basic single layer neural network with a simple sigmoid activation function. I am training and testing on a very simple labeled data set and with the current configuration it scores 90% on the un-seen classification tests.  I am by no means any kind of expert in this field, and there are plenty of resources and papers written by people who are inventing this stuff as we speak, but this will be more the musings of someone re-learning this stuff.

As always, all the code is on GitHub (and, as per my change in roles, this time it is all written in Scala).

A brief overview

The basic idea is that a Neural Network(NN) attempts to mimic the parallel architecture of the human brain. The human brain is a massive network of billions of simple neural cells that are interconnected, and given a stimulus they each either "fire" or don't, and its these firing neural cells and synapses that enable us to learn (excuse my crude explanation, I'm clearly not a biologist..).

So that is what we are trying to build: A connected network of neurons that given some stimuli either "fire" or don't.

A Neuron

Ok, so this seems like a sensible place to start, right? We all know what a network is, so what are these nodes in our network that we are connecting? These are the decision points, and in themselves are incredibly simple. They are basically a function that takes n input values  and multiplies them by a pre-defined weight (per input), adds a bias and then runs it through an activation function (think of this as our fire/don't-fire function):

In terms of our code, this is pretty simple (don't worry about the sigmoid derivative values we are setting, we are just doing this to save time later):

As you can see, the neuron holds the state about the weights of the different inputs (a NN is normally fixed in terms of number of neurons, so once initialised at the start we know that we will get the same number of inputs).

You may have also noticed the first line of the class, where we require an ActivationFunction to be applied - as I mentioned, this is our final processing of the output. In this example I am just using the Sigmoid function:

As you can see, it's a pretty simple function. Much like the brain, the neurons are very simple processors, and its only the combination of them that makes the brain so powerful (or deep learning, for that matter)

The network

Ok, so a sinple "shallow" NN has an input layer, a single hidden layer and an output layer (deep NNs will have multiple hidden layers).  The network is fully connected between layers - that is to say, each node is connected to every node in the next layer up, and thats it - no neurons are connected to other layers and no connections are missed.

The exact setup of the network is largely problem dependent, but there are some observations:

  1. - The input layer has to correlate to your inputs: If you have a data set that has two input values, let's say you have a dataset that contains house price data, and you have the input values number of rooms and square foot then you would have to have two input neurons.

  2. - Similarly, if you were using the NN for a classification problem and you know you have a fixed number of classifications, the output neurons would correlate to that. For example, consider you were using the NN to recognise hand written digits (0-9) then you would likely have 10 output neurons to group those possible outputs.

 The number of hidden neurons, or the number of layers is a lot more problem dependent and is something that needs to be tuned per problem.

If you have ever worked with graph type structures in code before, then setting up a simple network of these neurons is also relatively straight forward given the uniform structure of the network.

The weights

Oh right, yes. So all these neurons are connected, but its a weighted graph, in CS terms. That is, the connection between each node is assigned a weight - as a simple example, if we take the house price dataset as an example, after looking at the data we might determine that the number of rooms is a more significant factor in the end result, so that connection should have a greater weighting than the other input.  Now that is not exactly what happens, but in simple terms it's a good way of thinking about it.

Given that the end goal is for the NN to learn from the data, it really doesn't matter too much what we initialise the weights for all the connections to, so at start-up we can just assign random values. Just think of our NN as a newborn baby - it has no idea how important different inputs are, or how important the different neural cells that fire in response to the stimuli are - everything is just firing off all over the place as they slowly start to learn!

So what next?

Ok, so we have our super-simple neurons, that just mimic a single brain cell and we have them connected in a structured but randomly weighted fashion, so how does the network learn? Well, this is the interesting bit..

When it comes to training the network we need a pretty large dataset to allow it to be able to learn enough to start generalising - but at the same time, we don't want to train it on all the data, as we want to hold some back to test it at the end, just to see just how smart it really is.

In AI terms, we are going to be performing "supervised" learning - this just means that we will train it with a dataset where we know the correct answer, so we can make adjustments based on how well (or badly) the network is doing - this is different to "unsupervised" learning, where we have lots of data, but we don't have the right answer for each data point.

Training: Feed forward

The first step is the "feed-forward" step - this is where we grab the first record from our training data set and feed the inputs into our randomly initialised NN, that input is fed through the all of the neurons in the NN until we get to the output layer and we have the networks attempt at the answer. As the weighting is all random, this is going to be way off (imagine a toddlers guess the first time they ever see a smart phone).

As you can see, the code is really simple, just iterate through the layers calculating the output for each neuron. Now, as this is supervised learning we also have the expected output for the dataset, and this means we can work out how far off the network is and attempt to adjust the weights in the network so that next time it performs a bit better.

Training: Back propogation

This is where the magic, and the maths, comes in.  The TL;DR overview for this is basically, we know how wrong the network was in its guess, so we attempt to update each of the connection weights based on that error, and to do so we use differentiation to work out the direction to go to minimise the error.

This took a while of me staring at equations and racking my brain trying to remember my A-level maths stuff to get this, and even so, I wouldn't like to have to go back and attempt to explain this to my old maths teacher, but I'm happy I have enough of a grasp of what is going on, having now coded it up.

In a very broad, rough overview, here is what we are going to do now:

  1. - Cacluate the squared error of the outputs. That is a fairly simple equation of
    1/2 * (target - output)^2

  2. - From there, we will work out the derivative of this function. A lot of the errors we will start to work with now uses differentiation and the derivatives of the result - and this takes a little bit of calculus, but the basic reason is relatively straight forward if you think about what differentiation is for.

  3. - We work back through the network, calculating the errors for every neuron combining the error, derivatives and the weighting (to determine how much a particular connection played in an error, it also needs to be considered when correcting the weighting.

Once we have adjusted the weight for each connection based on the weight, we start again and run the next training data record.  Given the variation in the dataset, the next training record might be quite different to the previous record trained on, so the adjustment might be quite different, so after lots (maybe millions) of iterations over the data, incrementally tweaking and adjusting the weights for the different cases, hopefully it starts to converge on something like a reasonable performance.

Maths fun: Why the derivative of the errors?

So, why is calculating the derivative relevant in adjusting the errors?

If you imagine the network as a function of all the different weights, its pretty complicated, but if we were to reduce this, for the sake of easier visualisation, to just be a 3-d space of possible points (e.g. we have just 3 weights to adjust, and those weights are plotted on a 3-d graph) - now imagine our function plots a graph something like this:

(taken from the wikipedia page on partial derivatives)

The derivative of the function allows us to work out the direction of the slope in the graph from a given point, so with our three current weights (co-ordinates) we can use the derivative to work out the direction in which we need to adjust these weights.


The whole NN in scala is currently on Github, and there isn't much more code than I have included here.  I have found it pretty helpful to code it up from scratch and actually have to think about it - and has been fun coding it up and seeing it train up to getting 90% accuracy on the unseen data set I was using (just a simple two-in two-out dataset, so not that impressive).

Next up I will see how the network performs against the MNIST dataset (like the Hello-World benchmark for machine learning, and is the classification of handwritten digits).

Oh, and if you are interested as to why the images look like a dodgy photocopy, they are photos of original diagrams that I included in my final year dissertation in university, just for old time sake!

RESTful API Design: An opinionated guide

This is very much an opinionated rant about APIs, so it's fine if you have a different opinion. These are just my opinions. Most of the examples I talk through are from the Stack Exchange or GitHub API - this is mostly just because I consider them to be well designed APIs that are well documented, have non-authenticated public endpoints and should be familiar domains to a lot of developers.

URL formats


Ok, lets get straight to one of the key aspects. Your API is a collection of URLs that represent resources in your system that you want to expose. The API should expose these as simply as possible - to the point that if someone was just reading the top level URLs they would get a good idea of the primary resources that exist in your data model (e.g. any object that you consider a first-class entity in itself). The Stack Exchange API is a great example of this. If you read through the top level URLs exposed you will probably find they match the kind of domain model you would have guessed:


And whilst there is no expectation that there will be anyone attempting to guess your URLs, I would say these are pretty obvious. What’s more, if I was a client using the API I could probably have a fair shot and understanding these URLs without any further documentation of any kind.

Identifying resources

To select a specific resource based on a unique identifier (an ID, a username etc) then the identifier should be part of the URL. Here we are not attempting to search or query for something, rather we are attempting to access a specific resource that we believe should exist. For example, if I were to attempt to access the GitHub API for my username: I am expecting the concrete resource to exist.

The pattern is as follows (elements in square braces are optional):

Where including an identifier will return just the identified resource, assuming one exists, else returning a 404 Not Found (so this differs from filtering or searching where we might return a 200 OK and an empty list) - although this can be flexible, if you prefer to return an empty list also for identified resources that don’t exist, this is also a reasonable approach, once again, as long as it is consistent across the API (the reason I go for a 404 if the ID is not found is that normally, if our system is making a request with an ID, it believes that the ID is valid and if it isn't then its an unexpected exception, compared to if our system was querying filtering user by sign-up dates then its perfectly reasonable to expect the scenario where no user is found).


A lot of the time our data model will have natural hierarchies - for example StackOverflow Questions might have several child Answers etc. These nested hierarchies should be reflected in the URL hierarchy, for example, if we look at the Stack Exchange API for the previous example:

Again, the URL is (hopefully) clear without further documentation what the resource is: it is all answers that belong to the identified questions.

This approach naturally allows many levels of nesting as necessary using the same approach, but as many resources are top level entities as well, then this prevents you needing to go much further than the second level. To illustrate, let’s consider we wanted to extend the query for all answers to a given question, to instead query all comments for an identified answer - we could naturally extend the previous URL pattern as follows

But as you have probably recognised, we have /answers as a top level URL, so the additional prefixing of /questions/{ids} is surplus to our identification of the resource (and actually, supporting the unnecessary nesting would also mean additional code and validation to ensure that the identified answers are actually children of the identified questions)

There is one scenario where you may need this additional nesting, and that is when a child resource’s identifier is only unique in the context of its parent. A good example of this is Github’s user & repository pairing. My Github username is a global, unique identifier, but the name of my repositories are only unique to me (someone else could have a repository the same name as one of mine - as is frequently the case when a repository is forked by someone). There are two good options for representing these resources:

  1. The nested approach described above, so for the Github example the URL would look like:

    I like this as it consistent with the recursive pattern defined previously and it is clear what each of the variable identifiers is relating to.

  2. Another viable option, the approach that Github actually uses is as follows:

    This changes the repeating pattern of {RESOURCE}/{IDENTIFIER} (unless you just consider the two URL sections as the combined identifier), however the advantage is that the top level entity is what you are actually fetching - in other words, the URL is serving a repository, so that is the top level entity.

Both are reasonable options and really come down to preference, as long as it's consistent across your API then either is ok.

Filtering & additional parameters

Hopefully the above is fairly clear and provides a high level pattern for defining resource URLs. Sometimes we want to go beyond this and filter our resources - for example we might want to filter StackOverflow questions by a given tag. As hinted at earlier, we are not sure of any resources existence here, we are simply filtering - so unlike with an incorrect identifier we don’t want to 404 Not Found the response, rather return an empty list.
Filtering controls should be entered as part of the URL query parameters (e.g. after the first ? in the URL). Parameter names should be specific and understandable and lower case. For example:

All the parameters are clear and make it easy for the client to understand what is going on (also worth noting that for example returns an empty list, not a 404 Not Found). You should also keep your parameter names consistent across the API - for example if you support common functions such as sorting or paging on multiple endpoints, make sure the parameter names are the same.


As should be obvious in the previous sections, we don’t want verbs in our URLs, so you shouldn’t have URLs like /getUsers or /users/list etc. The reason for this is the URL defines a resource not an action. Instead, we use the HTTP methods to describe the action: GET, POST, PUT, HEAD, DELETE etc.


Like many of the RESTful topics, this is hotly debated and pretty divisive. Very broadly speaking, two approaches to define API versioning is:
  • Part of the URL
  • Not part of the URL
Including the version in the URL will largely make it easier for developers to map their endpoints to versions etc, but for clients consuming the API it can make it harder (often they will have to go and find-and-replace API URLs to upgrade to a new version). It can also make HTTP caching harder - if a client POSTs to /v2/users then the underlying data will change, so the cache for GET-ting users from /v2/users is now invalid, however, the API versioning doesn’t affect the underlying data so that same POST has also invalidated the cache for /v1/users etc. The Stack Exchange API uses this approach (as of writing their API us based at

If you choose to not include the version in your API then two possible approaches are HTTP request headers or using content-negotiation. This can be trickier for the API developers (depending on framework support etc), and can also have the side affect of clients being upgraded without knowing it (e.g. if they don’t realise they can specify the version in the header, they will default to the latest).  The GitHub API uses this approach

I think this sums it up quite nicely:

Response format

JSON is the RESTful standard response format. If required you can also provide other formats (XML/YAML etc), which would normally be managed using content negotiation.

I always aim to return a consistent response message structure across an API. This is for ease of consumption and understanding across calling clients.

Normally when I build an API, my standard response structure looks something like this::

[ code: "200", response: [ /** some response data **/ ] ]

This does mean that any client always needs to navigate down one layer to access the payload, but I prefer the consistency this provides, and also leaves room for other metadata to be provided at the top level (for example, if you have rate limiting and want to provide information regarding remaining requests etc, this is not part of the payload but can consistently sit at the top level without polluting the resource data).

This consistent approach also applies to error messages - the code (mapping to HTTP Status codes) reflects the error, and the response in this case is the error message returned.

Error handling

Make use of the HTTP status codes appropriately for errors. 2XX status codes for successful requests, 3XX status codes for redirecting, 4xx codes for client errors and 5xx codes for server errors (you should avoid ever intentionally returning a 500 error code - these should be used for when unexpected things go wrong within your application).

I combine the status code with the consistent JSON format described above.

Groovy Retrospective: An Addendum - Memory usage & PermGen

I can't really have a Groovy retrospective without mentioning memory.

Over the last four years I have spent more time than any sane person should have to investigating memory leaks in production Groovy code. The dynamic nature of Groovy, and it's dynamic meta-programming presents different considerations for memory management compared to Java, simply because perm gen is no longer a fixed size. Java has a fixed number of classes that would normally be loaded in to memory (hot-reloading in long living containers aside), where as Groovy can easily change or create new classes on the fly. As a result, permgen GC is not as sophisticated (pre moving permgen to the normal heap anyway) and largely in Java if you experienced an Out-Of-Memory permgen exception then you would just increase your permgen size to support the required number of classes being loaded by the application.

To be fair, the majority of the problems encountered were due to trying to hot-reload our code coupled the setup of the application container (in this case tomcat) and having Groovy on the server classpath rather than bundled with the application (much like you wouldn't bundle Java itself with an application, however, bundling Groovy with your application is recommended).

A couple of points of interest, if you are considering Groovy (especially relevant in a long running process where you attempt to reload your code):

The MetaClassRegistry

When you meta-program a Groovy class, e.g. dynamically add a method to a class,  its MetaClass is added to the MetaClassRegistry, which is a static object on the GroovySystem class. This means that any dynamically programmed class creates a tie back to the core Groovy classes.

The main consideration to keep in mind when meta programming in a Groovy environment is that if you want to reload your classes you now have a link between your custom code and the core Groovy code so you must either 1) explicitly clear out the MetaClassRegistry; 2) reload the core Groovy classes as well (throw everything out on reload)

I think coming from a Java environment, where you would likely use the JAVA_HOME on the server for long running applications, it can often seem logical to have a similar server classpath entry for Groovy also - but actually, the easiest approach is to bundle the groovy classes with your application so is a normal candidate for reloading.

If you decide not to reload Groovy, you can add explicit code to clear out the registry - this is pretty simple code, but a note of warning, without just throwing everything away there are still plenty of risks that you can leave links that stop your classes being collected (which was the case for me for a long time until just making all the third party classes (groovy included) candidates to be thrown away and reloaded.  Even with throwing away all third party libraries you can still be caught out if you use shutdown hooks (e.g. jvm shutdown hook to clean up connections will tie your classes back to your underlying JRE, meaning that no classes can be collected until you restart your application!)

Anyway, above is code to clear the registry, it assumes you have access to the GroovyClassLoader, but you can also follow the same approach by just grabbing the MetaClassRegistry from any arbitrary Groovy class and iterate through that. Give it a try, if you play around a little you will probably find its quite easy to create a leak if you want to!

Anonymous Classes

Another thing to keep in mind in Groovy applications is the generation of classes by your application. As you would expect, as you load in Groovy classes to your application they will be compiled to class files (or if you are pre-compiling into a JAR or something) which will be added to PermGen. However, in addition to this, your code being executed may also result in additional (possibly anonymous) classes also being created and added to PermGen - so without care, these can start to fill up that space and cause OOM exceptions (although generated classes will often be very little, so might take a while before it actually errors).

An example of what might do this is loading Groovy config files - if you are loading sensibly and just doing it once then it won't be an issue, but if you find yourself re-loading the config every request/execution then it can keep adding those to PermGen.  Another example of where this happens (surprisingly) is if you are using Groovy templating. Consider the following code:

(Taken from the SimpleTemplateEngine JavaDocs)

The example is a simple example of binding a Groovy template with a map of values - maybe something that you would do to send an email or create a customized document for someone - but behind the scenes Groovy will create a class that is added to permgen for each execution. Now this isn't a  lot, however, if you are dealing with high throughput it can certainly add up pretty quickly.

Lazy Garbage Collection

Another interesting behaviour that I observed over the last few years is that, in the JVM implementation I was using, the PermGen garbage collector collected lazily.  As I mentioned, because nothing interesting traditionally happened in the permanent generation in the JVM, the garbage collectors didn't do anything interesting. Further more, because it was always assumed that the contents of perm gen were fairly static (as the name suggests) the collection happens fairly in-frequently, and often only kicks in for a full collection (which is more costly).  What this means is that even if everytime you reload you free up lots of classes for collection (say, your entire application), it might not GC permgen for several reloads as a full collection isn't required, and the JVM will just lazily perform the collection when permgen is almost full.

If you look at the diagram above, it displays a common pattern I observed - each little step in the up swing of the chart represents a full application reload, but you will see that there are multiple reloads before the permgen usage approaches the limit, and it is only when the usage is close to the limit that it actually performs the collection.

The challenges this can present is that if you push the permgen close to the limit but still not triggering a full collection, then the following reload it can once again spike the permgen (because the entire application and its associated third party code is being reloaded into memory), this can push it over the limit and cause OOM exceptions.  This was not something I ever saw on production environments, but was fairly common in desktop/development environments where less resources were available.

(another interesting observation in this particular pattern is that the GC seems to be intermittently collecting fewer classes. I never got to the bottom of that question mark: there was no difference in activity between application reload each time, so there is no change in application behaviour to trigger a leak and the middle period of reloading maintains a constant level of memory usage patterns - which also shows no leak behaviour)

Hopefully someone smarter than me about all this stuff is reading this and can shed some better insights into it, otherwise, hopefully its helpful if you are about to start doing crazy things with perm gen..