This is a short yet rambling tale of my experiences adopting GraphQL on a somewhat mature project: a Clojure server and Clojurescript front end, two years old with a current team of 5 developers. I discuss where we started, why we chose to adopt GraphQL, the libraries we used, the mistakes we made, how we feel about it now, and some advice on eating cake.
The application in question is relatively straightforward: it is a UI for displaying, in a hopefully coherent and intuitive manner, large amounts of financial data which is calculated by other applications.
The data can be logically described as a graph, albeit the special kind of graph known as a tree. It is fetched by the server from many other upstream applications and glued together into the tree that made most sense to our customer, users and developers - what will henceforth be referred to as the canonical form.
The domain is quite complicated, so let's use an analogy and say that our tree represents a sport, comprising of divisions, teams and players. Here's how it looks:
The vast majority of data in the graph resides in the lower layers
The graph is quite bottom heavy in terms of volume of data - there are only a handful of divisions and teams, but many players per team who play large numbers of games and generate a lot of statistical data per game.
The industry-standard way to represent a large amount of financial data is to throw it all into a large table with all the numbers visible at once and make the poor user scan it constantly for the one or two bits of information they actually need to know.
We wanted our application to be more valuable and helpful to our users. With an engaged product owner we designed screens which could meaningfully aggregate the fine-grained data and compare it to business-driven rules, resulting in a simple green-for-good or red-for-bad representation that directed the user's attention to the relevant places. Imagine a division summary showing the top 3 teams and 5 most valuable players, or a team screen showing each player and their best and worst statistics for the most recent game. There's no need to drown the user in data! The relevant facts are a tiny fraction of the total:
Highlighting the bits of data that are actually relevant to the user: a tiny fraction of the total
We decided early on to keep things very simple and simply transfer the entire graph from the server to the client at the start of a new client session, and keep it up-to-date via delta payloads sent over an SSE channel. We then derive streams of data on the client for each view which select the appropriate slices (we use the excellent re-frame for this). This gave us a number of benefits:
We are lucky that browsers have come far enough that tens of megabytes of data can be consumed at a canter, making this approach possible. It allowed us to think of the application using this very simple model:
When you have a very tall cake it's easiest to eat it in thin slices
Our graph of data was getting bigger and bigger as we covered more divisions, more teams per division, and more games were played generating more stats: the initial payload transferred to the client was reaching 120mb (gzipped transit) and SSE updates were increasingly frequent, putting load on the server and all the clients. Desktop users started complaining of slowness getting in to the app, and it was even slower for mobile users.
It was obvious that most of the data being pulled down wasn't needed all at the same time. If you're looking at the home page you need to know the names of the divisions and the names of the top teams in each division. You don't need every player's statistics for every game they've ever played! For any given view, you need a particular slice of all your data - for high level overviews you need a wide but shallow slice, and for detailed views you need narrow but deep slices. There is no page on the app where we need all the data in the graph. The argument for displaying only the relevant data to the user became the argument for only fetching the data relevant to the display.
We wanted a way to reduce the data requirements of the client without losing the benefits outlined above of having the canonical graph everywhere. The views should be able to declare the minimal set of data they need and we should be able to transfer just that while retaining the canonical structure.
Fragmenting our API into lots of little endpoints that each served a specific page was an unappealing option; each would have a lot of overlap but at the same time would not be reuseable or interchangeable and would also mean churn across the server and client. Plus, it's 2018 - there must be something better!
Adding a server-side store like Datomic with pull requests sounded very heavyweight - our "database" was just an in-memory atom containing immutable data, and exposing the internal structure and relations of our store sounded like it would make refactoring difficult.
GraphQL sounded good because as a specification rather than an implementation it keeps our internals hidden, has plenty of third party tooling and gives us a way to let the client decide what data it needs from the server so we can tailor it to the views. It has a (draft) subscriptions feature which we could use to push updates in the graph to our client and rather fortunately there existed a Clojure implementation called lacinia by Walmart Labs which was already battle-tested in their production environment.
Initially tools like GraphiQL were flashy and exciting. I played around with the example GraphiQL and had the feeling of being a pioneering explorer, slowly uncovering new parts of the graph and enthusing about the possibilities of concise and perfectly tailored data. It's simple enough that non-technical users could use it to find their way around which could also be a great bonus for business analysts and others close to the development team.
The next step was to spike our own little implementation into our existing webserver - enter lacinia and lacinia-pedestal. They are two excellent libraries that do all the heavy lifting for you, requiring you to describe your graph as pure Clojure data and allowing you to specify your resolvers as pure functions. It is implemented with
clojure.spec meaning if you make a mistake you usually get a nice error message to help you fix it. We integrated it into our existing pedestal server with very little fuss - we just included their interceptors on the routes we wanted to serve the graph from, and it sits alongside our Swagger API allowing us to migrate gradually.
lacinia does a great job of interpreting the query and dispatching to your resolver and streamer functions to deal with returning the data. It doesn't get in the way and is a great example of writing an extensible library. One really nice feature is that you may declare a resolver function for any node in the graph, giving you complete control over how to partition your fetching: if you have a node which is expensive to fetch, you only need to perform the fetch if the client explicitly requests that field.
It was very exciting to make our first few queries, but almost my next thought was
do I really have to have camel or snake case for my keys?
I spent a while researching this and contemplating submitting changes to lacinia but it was pointed out to me that tooling like GraphiQL, one of the main selling points, wouldn't work. In an effort to introduce GraphQL without having to rewrite all our UI code I came up with locksmith which takes a lacinia GraphQL schema and provides functions for efficiently transforming between idiomatic Clojure data and GraphQL data.
cljsjs due to the way it is packaged. To this end I wrote my own client called re-graph which handles queries, subscriptions and mutations over HTTP and WebSocket. This, combined with locksmith meant that the only changes really required in our client code were to add the queries each component needed.
We initially used GraphiQL to write our queries to take advantage of the autocomplete and pretty printing before pasting the big query strings back into our client code. Formatting them with indentation quickly became an issue and we looked into venia, but at the time it was missing a feature we needed to alias nested keys. That feature is now present, but our queries have not really changed since first writing them and the urge to de-stringify them has faded.
We set up our subscription to fire whenever our data store changed, as this was the simplest thing to do. This re-runs the query and pushes the data down the websocket to the client, keeping them up-to-date with the latest state on the server. At busy times the subscription was firing many times a second but executing the query was taking one or more seconds resulting in a backlog building up and extremely high CPU. We managed to end up with 40,000 threads all trying to execute queries, unsurprisingly causing the application to grind to a halt. To protect ourselves the subscription is now debounced to fire no more than once every 2.5 seconds and I proposed this change to lacinia to stop unlimited threads being created and to drop superceded interim values. The hard lesson is lacinia doesn't protect you from your own performance issues - that part is up to you.
It also turns out that lacinia does not deduplicate subscriptions with the same id, resulting, if you are not careful, in multiple threads doing identical work to return identical data on the client side. We have added deduplication inside our subscription streamer but I may also propose this is incorporated into lacinia-pedestal. Another hard lesson: lacinia doesn't protect you from bad clients.
We took the opportunity to simplify our graph which felt good but increased the amount of work to do - it became a small rewrite rather than a drop in solution. We thought if we don't do it now, it will be harder later and we might never do it. In retrospect I wish we had dropped in GraphQL as-is to get the benefits and do the refactoring later, but we'll never know now which way was better.
In addition to the simplification we were also tempted to further reduce the data transferred by shortcutting the graph and adding derived fields like "bestPlayer" at the team level. This saves on data - to calculate the best player you'd have to transfer data about all the players and sort them on the client side - but it breaks our adherence to the one canonical form.
Hidden logic or verbosity? The GraphQLer's dilemma
This is a philosophical trade-off which we haven't balanced satisfactorily yet. On the one hand it reduces the data requirements even further and simplifies the client but on the other it complects data and logic which seems to break the GraphQL paradigm of pure data and explicit behaviour. I'm not happy with these derived fields but they are a huge temptation when you own both the client and the server. I think a better way would be to request the list of players as before, but sorted by performance and limited to 1. The data structure remains unchanged, is minimised, and the implicit server logic in the name "bestPlayer" becomes explicit logic driven by the client.
Our client side code has also become slightly harder to re-use across UI components because some functions rely on fields which we may have forgotten to include in all the queries that use it. It took some time to weed these out, but is an area where
clojure.spec could help if we had functions that failed if they were missing input fields that we had missed from our queries.
lacinia and lacinia-pedestal made it very easy to get going with GraphQL and the way our client code was arranged allowed us to migrate just a couple of views, including the home page, as a spike. From spending 60+s to transfer 120mb we now take a small fraction of a second to load a few kilobytes. Our users are happy once again and we intend to migrate the rest of the application and retire the SSE channel.
Performance is your responsibility and you need to think carefully about your implementation, including the size of your graph, typical payloads, frequency of subscription updates and number of concurrent users.
Finally real-life queries can seem very verbose - you have to explicitly list every field you want back from the server, which is tiresome when you want them all, but the lack of a wildcard is by design: you won't get the benefits of GraphQL if you don't think about what fields you want and why.
Our server gathers data from many other applications and in every case each response returns more data than we want. This increases the load on both us and the services we are calling; if they all had a GraphQL endpoint we could drastically reduce the amount of data flying around. Pushing the GraphQL revolution upstream is already something we're looking into to help the next generation of applications and encourage small and lightweight UIs that can easily be put together.
Performance is something we need to monitor carefully. The savings on the client are offset to some extent by greater demands on the server. I've put together a library called lacinia-gen which takes a lacinia graph description and creates generators for the data with the intention of using it to find hotspots and improve the overall performance.
Refactoring of our graph specification is also something I'd like to look into. It's hard to get it right and strike the right balance of being useful and remain as just pure data, but I think pure data is the way; time will tell.
The libraries I wrote for our GraphQL journey - re-graph, locksmith and lacinia-gen - are all in active use on our project and have served us well. All are on GitHub, and as always feedback and contributions are very welcome.
Sign up to the JUXT newsletter