Clojure at paper.li
I’ve been involved with clojure almost exclusively for a year as smallriver’s lead architect, working on the paper.li product and wanted to share my experience of clojure in the real world.
I had a previous experience with clojure where I put it to work where ruby on rails wasn’t a natural fit, and although smallrivers is a close neighbor of typesafe in Switzerland, my previous experience with the language made it prevail on scala.
Why clojure ?
While working on the backend architecture at a previous company I decided to evaluate three languages which met the needs I was faced with:
I decided to tackle the same simple task in all three languages and see how each would fare and how I felt about them. The company’s language at that time was Ruby and JS, and coming from a C background, I wanted a language which provided simplicity, good data structure support and concurrency features, while allowing us to still code quickly.
While naturally drawn to Erlang, I quickly had to set it apart because the stack that was starting to emerge at the time had JVM based parts and would benefit greatly from a language targetting the JVM. I was a bit bummed because some tools in the erlang world were very exciting and the lightweight actors were interesting for a part of our stack.
Scala made a very strong first impression on me, but in practice I was taken aback by some aspects of it: the lack of coherency of open source projects found on the net in terms of style, which made it hard to see which best practices and guidelines would have to be taught to the team, some of the code I found was almost reminiscent of perl a few year back, in the potential it had to become unmaintainable some time later. The standard build tool - SBT - also made a very weak impression. It seemed to be a clear step back from maven which given the fact that maven isn’t a first class citizen in the scala world seemed worrying.
Clojure took the cake, in part because it clicked with the lisper in me,
in part because the common idioms that emerged from the code I read bore
a lot of similarity with the way we approached ruby. The dynamic typing
promised succinct code and the notation for vectors, maps and sets
hugely improved the readability of lisp - look at how hashes work in
emacs lisp if you want to know what i mean. I was very excited about
dosync and a bit worried by the lack of leightweight erlang style
actors even though I could see how
agent’s could help in that regard.
As I’ll point out later on, we ended up not using these features at all
The task at hand
When I joined Smallrivers to work on paper.li, it became natural to choose clojure. The team was small and I felt comfortable with it. There was a huge amount of work which needed to be started quickly so a “full-stack” language was necessary to avoid spreading across too many languages and technologies, and another investigation in how the other languages had evolved in the meantime was not possible. The main challenges to tackle were:
- Being able to aggregate more content
- Improve the quality of the processing done on content
- Scaling the storage cluster accordingly
- Automate the infrastructure
The “hiring” problem
One thing that always pops up in discussions about somewhat marginal languages is the hiring aspect, and the fear that you won’t be able to find people if you “lock” yourself in a language decision that strays from the usual suspects. My experience is that when you tackle big problems, that go beyond simple execution but require actual strong engineers, hiring will be a problem, there’s just no way around it. Choosing people that fit your development culture and see themselves fit to tackle big problems is a long process, integrating them is also time consuming. In that picture, the chosen language isn’t a huge deciding factor.
I see marginal languages as a problem in the following organisations:
- Companies tackling smaller problems, or problems already solved. These are right in choosing standard languages, if I built a team to build an e-commerce site I wouldn’t go to clojure.
- Larger companies which want their employees to jump from project to project, which makes sense from a managerial standpoint.
What we built
The bulk of what was done revolves around these functional items:
- A platform automation tool, built on top of pallet.
- Clojure facades for the tools relied upon (elastic search, cassandra, redis, kafka).
- An ORM-type layer on top of cassandra
- Our backend pipelines
- A REST API
I won’t go in too much detail on our in-house code, but rather reflect on how things went over.
Coding style and programming “culture”
One of the advantages of lisp, is that it doesn’t have much syntax to go around, so our rules stay simple:
- the standard 2 space indent
- we try to stick to 80 columns, because i’m that old
- we always use
requireexcept for: clojure.tools.logging and pallet.thread-expr which are
- we avoid macros whenever possible
- we use dynamically rebindable symbols
Of course we embraced non mutable state everywhere possible, which in our case is almost everywhere. Whenever we need to checkpoint state, it usually goes to our storage layer, not to in memory variables.
When compared to languages such as C, I was amazed at how little rules are needed to enforce a consistent code look across projects, with very little time needed to dive into a part written by someone else.
We didn’t settle on a unique tool-suite at the office, when picking up clojure I made the move from vim to emacs because the integration is better and I fell in love with paredit. Spread amongst the rest of team, textmate, eclipse and intellij were used.
For building projects, leiningen was an obvious choice. I think leiningen is a great poster child for the greatest in clojure. A small and intelligent facade on top of maven, hiding all the annoying part of maven while keeping the nice distribution part.
For continuous integration, we wrote a small bridge between leiningen and zi lein-zi which outputs pom.xml for maven, which are then used to build the clojure projects. We still hope to find some time to write a leiningen plugin for jenkins.
Since a good part of what paper.li does relies on aggregation, async programming is very important. In the pure clojure world, the only real choice for async programming is lamina and aleph. To be honest, aleph turned out to be quite the challenge, a combination of the amount of outbound connections that our work requires and the fact that aleph seems to initially target servers more than clients.
Fortunately Zach Tellman put a lot of work into the library throughout last year and recent releases are more reliable. One very nice side effect of using a lisp to work with evented code is how readable code becomes, by retaining a sync like look.
For some parts we still would directly go to a smaller netty facade if we were to start over, but that’s a direct consequence of how much we learned along the way.
Libraries not frameworks
A common mantra in the clojure development community is that to ease integration the focus should be on libraries, not frameworks. This shows in many widespread projects such as compojure, pallet, and a host of common clojure tools. This proved very useful to us as clients of these libraries, allowing easy composition. I think pallet stands out most in that regard. Where most configuration management solutions offer a complete framework, pallet is just a library offering machine provisioning, configuration and command and control, which allowed us to integrate it with our app and build our abstractions on top of it.
We tried to stick to that mantra in all of our work, building many small composable libraries, we made some errors at the beginning, by underutilizing some of clojure features, such as protocols but we now have good dynamics for writing these libraries, by writing the core of them with as little dependencies as possible, describing the behavior through protocols, and then writing add-ons which bring in additional dependencies and implement the protocol.
Macros and DSLs
Another common mantra is to avoid overusing macros. It can’t be overstated how easy they make things though, our entity description library (which we should really prep up for public release, we’ve been talking about it for too long now) allows statements such as these (simplified):
(defentity :contributors (column :identifier (primary-key)) (column :type (required)) (column :name) (column :screen_name (index)) (column :description) (column :detail (type :compound)) (column :user_url) (column :avatar_url) (column :statuses_count (type :number)) (has-many :articles) (has-many :editions (referenced false) (ttl 172800)) (has-many :posts (key (timestamp :published_at)) (referenced false)))
The power of DSLs in clojure cannot be understated, with a few macros you can easily build full languages, allowing easy extending of the functionality. Case in point, extracting text from articles, like most people we rely on a generic readability type library, but we also need to handle some sites that need special handling. By using a small DSL you can easily push rules that look like (simplified):
(defsiterule "some.obscure.site" [dom] (-> dom (pull "#stupid-article-id")))
The great part is that you limit the knowledge to be transfered over to people writing the rules, you avoid intrusive changes to the core of your app and these can safely be pulled from an external location.
At the end of the day, it seems to me as though the part of the clojure community that came from CL had awful memories of macros making code unreadable, but when sticking to macros with a common look and feel, i.e:
def<resource>type macros, there are huge succintness take aways without hindering readability or maintenance of the code.
Every respectable codebase is going to need at least a few test. I’m of the pragmatist church, and straight out do not believe in TDD, neither in crazy coverage ratios. Of course we still have a more that 95% unit test coverage and the decoupled approach preached by clojure’s original developer, rich hickey1 allows for very isolated testing. For cases that require mocking, midge provides a nice framework and using it has created very fruitful throughout our code.
Concurrency, Immutable State and Data Handling
Funnily, we ended up almost never using any concurrency feature, not a single
dosyncmade it in our codebase, few
atom’s and a single
agent(in https://github.com/pyr/clj-statsd to avoid recreating a Socket object for each datagram sent). We also banned
futureusage to more closely control our thread pools. Our usage of
atom’s is almost exclusively bound to things that are write once / read many, in some cases we’d be better off with rebindable dynamic symbols.
We rely on immutable state heavily though, and by heavily I actually mean exclusively. This never was a problem across the many lines of code we wrote, and helped us keep a sane and decoupled code base.
With facades allowing to represent database fields, queue entries, and almost anything as standard clojure data structures and with powerful functions to work on them, complex handling of a large amount of data is very easily expressed. For this we fell in love with several tools which made things even easier:
- the threading operators
- the pallet thread-expr library which brings branching in
when->, and so on
seq-utils/index-byand all these functions which allow easy transformation of data structs and retain a procedural look
I cannot stress how helpful this has been for us in doing the important part of our code right and in a simple manner. This is clearly the best aspect of clojure as far as I’m concerned.
Moreover, building on top of Java and with the current focus on “Big Data” everywhere, the interaction with large stores and tools to help building batch jobs are simply amazing, especially cascalog.
- the threading operators
The case of Clojurescript
While very exciting we did not have a use for clojurescript, given the size of the existing JS codebase, and the willingness of the frontend developers to stick to a known.
The simple existence of the project amazes me, especially with the promise of more runtimes, there are various implementations on top of lua, python and gambit (a scheme that compiles to C). With projects like cascalog, pallet, lein, compojure, noir and clojurescript, the ecosystem addresses all parts of almost any stack that you will be tempted to build and we didn’t encounter cases of feeling cornered by the use of clojure - admiteddly, most of the time, a Java library came to the rescue.
The community is very active, and has not reach critical mass yet, which makes its mailing-list and irc room still usable. There are many influent public figures, some who bring insight, some who bring beautiful code. Most are very open and available to discussion which shaped our approach of the language and our way of coding along the way.
It’s been an exciting year and we’re now a full fledged 80% clojure shop. I’m very happy with the result, more so with the journey. I’m sure we could have achieved with other languages as well. As transpires throughout the article, the whole team feels that should we start over, we would do it in clojure again.
It helped us go fast, adapt fast and didn’t hinder us in any way. The language seems to have a bright future ahead of it which is reassuring. I would encourage people coming from python and ruby to consider it as a transition language or as their JVM targetting language, since many habits are still valid in clojure and since it helps slightly change the way we look at problems which can then be reapplied in more “traditional” languages.
Rich hickey’s talk simple made easy and his coining of the term “complecting” illustrates that http://www.infoq.com/presentations/Simple-Made-Easy ↩︎