Eigen is where we worried about how we were building apps, other apps are limited in their scope, but Eigen is basically the mobile representation of Artsy. We're never not going to have something like Eigen.
We eventually came to the conclusion that we needed to re-think our entire UIKit stack for Eigen. Strictly speaking, Objective-C was not a problem for us, our issues came from abstractions around the way we built apps.
Re-writing from scratch was not an option. That takes a lot of time and effort, which will happily remove technical debt, but that's not our issue. We also don't need or have a big redesign. However, a lot of companies used the Objective-C -> Swift transition as a time to re-write from scratch. We asked for the experiences from developers who had opted to do this, they said it was a great marketing tool for hiring - but was a lot of pain to actually work with day to day. They tend to talk about technical debt, and clean slates - but not that Objective-C was painful and Swift solves major architectural problems. With the notable exception of functional programming purists.
In the end, for Eigen, we came to the conclusion that we wanted to work with a component-based architecture. This architectural choice comes from studying how other larger apps handle code-reuse.
We were considering:
- View Controllers being a mix of Components which could be extended using protocols in Swift.
- JSON defined Components ( which would have ended up like Spotify's (hub) or Hyperslo's (Spots) ).
- Building a Component structure heavily inspired by React ( like Bending Spoons's (Katana) ).
Had we continued with native apps via native code, we'd have put more resources behind Swift, which had quite a bit running for it:
It was consistent with our existing code. We wrote hundreds of thousands of lines of code in Objective-C and maybe around a hundred thousand of Swift. The majority of the team had 5+ years of Cocoa experience and no-one needs to essentially argue that continuing with that has value.
Swift code can interact with Objective-C and can work on its own. We can write Swift libraries that can build on-top of our existing infrastructure to work at a higher level of abstraction. Building a component-based infrastructure via Swift could allow easy-reuse of existing code, while providing a language difference for "new app code" vs "infra."
People are excited about Swift. It's an interesting, growing language, and one of the few ones non-technical people ask about. "Oh you're an iOS developer, do you use Swift?" is something I've been asked a lot. The developers outside of the mobile team have signed up multiple times for Swift workshops and want to know what Swift is, and what its trade-offs are.
It's evolving the language changes at a fast rate, with new ideas coming from, and influencing other languages. People inside the community influence and shape its growth. There are some great claims being made about Swift by people we respect.
Swift improves on a lot of Objective-C. Most of the patterns that we use in Objective-C are verbose, and they can become extremely terse inside Swift. Potentially making it easier to read and understand.
We would be using the official route. Apple obviously wants you to be using Swift, they are putting a lot of resources into the language. There are smart people working on the project, and it's become more stable and useful every year. There aren't any Swift-only APIs yet, but obviously they'll be coming.
It's a known-unknown territory. We have a lot of knowledge around building better tooling for iOS apps. From libraries like Moya, to foundational projects like CocoaPods. Coming up with, and executing dramatic tooling improvements is possible. Perhaps we had overlooked a smarter abstraction which would have worked around the downsides, and thus making it worth expanding our search.
If we end up building something which gains popularity, we get the advantage of working with a lot of fresh perspectives, and being able to gain from other people working on the same project. This is what happened with Moya. It's a pattern Basecamp discuss when they talk about rails by beginning with a real project and abstracting outwards.
The dominant two issues come from differences in opinions in how software should be built
Types. Types are useful. Overly strict typing systems make it too hard to quickly change codebases.
Strictly typed languages work really well for building systems, or completely atomic apps - the sort Apple have to build on a day to day basis. When I say an atomic app, I mean one where the majority of the inputs and outputs exist within the domain of the application. Think of apps with their own filetypes, that can control inputs and outputs really easily.
Even in Objective-C, a looser-typed language where you were not discouraged from using meta--programming, handling JSON required a tonne of boilerplate laden, inelegant code when working with an API. Considering how bread-and-butter working with an API is for most 3rd party developers it should come as no surprise that the most popular CocoaPods are about handling JSON parsing, and making network requests.
Problems which Apple, generally speaking, don't have. They use iCloud, or CloudKit, or whatever, and expect you will also. The official Apple opinion was neatly summed up on the official Swift blog on how to handle JSON parsing exhibits the problem well.
Swift’s built-in language features make it easy to safely extract and work with JSON data decoded with Foundation APIs — without the need for an external library or framework.
They do, but it's not great code to write nor maintain. I don't know anyone who does what they recommend in production.
The stricter type system in Swift made it harder to work on JSON-driven apps.
Slow. Native development when put next to web development is slow. Application development requires full compilation cycles, and full state restart of the application that you're working on. A trivial string change in Eigen takes 25 seconds to show up. When I tell some developers that time, they laugh and say I have it good.
The moment that this really stood out for me was when I re-discovered Injection for Xcode which ruined my appetite for building apps the traditional way. It reduced an iteration cycle to about a second. With Apple's resources, and the fact that Injection for Xcode has existed for years by a single developer, it's frustrating that iOS is a mobile platform with no support for code reloading. I filed bug reports (radars), they were marked as duped with no comment. I talked to Apple engineers at WWDC, the idea was dismissed as "didn't work" when it was tried before.
I've heard developers say they use Playgrounds to work around some of these problems, and the Kickstarter app has probably the closest I've seen to an actual implementation of this, so check that out if you're hitting these issues.
The Swift compiler is slow. Yes, it will improve. One of my favourite Swift features, inferred typing, can accidentally increase compile times non-obviously. Which can make it feel arbitrary about what code takes longer to compile or not. We eventually automated having our CI warn us whether the code we were adding was slow as it felt hard to predict.
You may want to read our announcement of switching to React Native in anticipation of this. However the big three reasons are:
- Better developer experience.
- Same conceptual levels as the rest of the team.
- Ownership of the whole stack.
However, the key part of this post is how does this compare to native development? Also, have these arguments stood up to the test of time a year later?
Sidenote: I found it hard to write this without being able to comprehensively reference what we are doing now, and so, I'll be referencing a sibling article: JS 2017.
References from JS 2017: Relay, Jest
As everyone inside the community has both the ability and the tools to contribute to the ecosystem you get better tools.
Apple make great tools. I do miss Xcode sometimes. It's cohesive, beautifully designed and doesn't show its age. It's a perfect Mac citizen.
Though it's important to note that they make tools for Apple first and then for us 3rd party devs. Outside influence obviously exists, but they're shipping whatever they want and you can only influence that via Radars and through going to a conference once a year and talking directly to the dev tools team. Only the Swift language is Open Source (and SwiftPM)
There are so few well built, large developer tools for the Apple ecosystem. Developers are wary of being copied by Apple - something so prevalent that there is a common word for it, being Sherlocked. The project I've worked on for 5 years, CocoaPods, had an announcement of being sherlocked in late-2015 - you just have to deal with it. The idea that only Apple should be shipping these kind of things kills community momentum.
If you're going to build something amazing, only to have all support pulled out from under you once it gets popular because Apple copied it and made it for free and with a full time team behind it - why bother?
This makes it tough for us, as the 3rd party community, to build useful tools on the kind of scale that is normal in other developer ecosystems.
Reference from JS 2017: Jest
Better Abstractions, Better Developer Experience
I've mentioned that the apps we build have problems specific to API-driven applications. This means that the majority of our work tends to be that we have the full data already, and need to iterate to get the right styling and logic mapping correct, in doing so we want to also make it easy to re-use code.
All of these frameworks have the same domain problems that our iOS apps have, external API stores, complex user device state and a mature user-interface API (either the DOM, or UIKit.)
With React, the core concept of a virtual DOM means that you can simplify a lot of complicated state-management for your application. It becomes trivial, removing the need for more complicated state-handling ideas like functional or reactive programming.
With Relay, we got an genuinely ground-breaking change in how interactions get handled with our API. I don't want to work against an API without a tool like Relay again.
References from JS 2017: React, Relay
Both of these tools provide a developer experience better than iOS native tooling. React's strict state management rules allow external tools to extend a React application easily, so the onus is not on the React team to make better tools. Other projects provide tools like: debuggers, external state viewers, runtime code injection, component storyboarding all of which can be running simultaneously as you are building your application. Imagine being given the flow of all state in your app in every bug report.
A single press of save would take your changes, inject it into your current running application, keep you in the exact same place, depending on the type of change it could re-layout your views, and so you can stay in your editor and make your changes. From 25 seconds, to less than one. For a lot of my work, I can put my tests, debuggers and the application on another screen, and just move my head to watch changes propagate on pressing save.
React Native feels like the best of both worlds: Elegant, fast to work with application code, which the whole dev team understands. Falling back to native tooling when we think it will be best for the project.
Reference from JS 2017: TypeScript
Same Tools, Different Dev
This isn't a great position to be in.
We wanted all developers to feel like they can contribute to any area of the company. For the past 5 years, the native mobile projects had close to zero contributions from anyone outside of the mobile team. Due to differences in tooling, and the idea that there was a cultural difference between us. Since the mobile team moved to React Native we have received features and bug fixes from the web team, and fit in better overall.
This expansion of a mobile team developer's scope has made it much easier for us to reason about finding better ways to share code with the web team. At the end of 2015, the Collector Web team introduced GraphQL to Artsy. I wrote about how this affected the mobile team. This acts as an API layer owned by the front-end side of Artsy. Meaning that it could contain a lot of API-derived client-specific logic. Previously, this work was done by the web team, and then consumed by mobile - now both teams build their APIs and consume them.
Reference from JS 2017: GraphQL
This is not something we have explored too deeply, however we expect to be able to port a lot of our React Native to Android. I got a rough prototype ported in 2 days work. By working at React-level, and allowing the React Native bindings to handle the interactions with the host OS, we've been writing cross-platform code.
We consider ourselves blocked on Android support, specifically by not having an engineer in our team with deep experience in Android. Moving to React Native does not obviate our native skills, you're going to be significantly better in that environment with those skills than without. As we mentioned in our announcement:
If you’re not already knowledgeable about iOS development, are not motivated to put in the time to learn about the
platform specific details, and think making rich iOS applications with React Native will be a breeze, you’ll
come home from a very cold fair indeed.
We need someone with a similar depth of knowledge in the Android ecosystem as our iOS, but we may need one or two for the entire team. The rest can continue to be a mix of Web and iOS engineers. You gain a subset of cross-platform skills using React Native. Had we continued down the path of using Swift, our skills would continue to be siloed.
With respect to Swift on Android, potentially, logic code could be shared between platforms but realistically for our setup that's just not worth it. We're moving that kind of logic into the GraphQL instance and sharing across all clients, not only native platforms. If you're sharing model code, you could generate that per-project instead from the server. Since GraphQL is strongly-typed, we're doing this for both TypeScript + GraphQL and TypeScript + Relay.
We don't know where this will end, but we've prototyped porting one of our view controllers from React Native to a website. It's almost source-compatible. This such a completely different mindset from where we were a year ago.
Owning the stack
Pick an abstraction level of our application above UIKit and we can fork it. All our tools can be also be forked. We can fix our own issues.
In native, there are no concepts like, "We'll use Steipete's fork of UIKit for UIPopover rotation fixes" or "My version of Xcode will run tests when you press save.". Well, hopefully the latter may be fixed in time, but the "you have no choice but to wait, and maybe it won't happen" aspect is part of the problem.
You have your tools given to you, in a year you get some new ones and lose some old ones. In contrast, we've built many extensions for VS Code for our own use, and helped out on major ones. When VS Code didn't do what I wanted, I started using use my own fork.
Reference from JS 2017: VS Code
It's worth highlighting that all of this is done on GitHub, in the open. We can write issues, get responses, and have direct line to the people who are working on something we depend on. This is a stark contrast to the Radar system used internally at Apple, and which external developers have write-only access to. For external contributors radar is opaque, and often feels like a waste of time. On the other hand, a GitHub issue doesn't have to wait for the repo maintainers, others can get value from it and it's publicly indexed. If we had put all our effort into Radars instead of issues like this, the whole community would be worse off.
This isn't all doom and gloom. With Swift the language, and SwiftPM the package manager, Apple are more open with the feedback cycle using tools like Slack, Mailing Lists, JIRA and Twitter.
In contrast, and this may be the last major time it happens, but people refer to the time it took to migrate in the scale of weeks during the Swift 2 -> 3 migration. Having the language evolve is great, sometimes in ways that you agree with and sometimes in ways you don't. Being able to use your own version of your tools frees you to make it work for you and your business. We have been talking about extending TypeScript specifically for our applications.
React Native, one year later
In our announcement we talked about the lack of nuanced post-mortems on React Native. We're now a year in, we can at least try to help out in that space. We're sticking with React Native for the foreseeable future. It would take some drastic changes in the Apple ecosystem for us to re-consider this decision. So here's the summary after 1 year.
- We can share concepts with web
- Tools are built for apps like ours
- To do it right requires engineers willing to dive deep in JS
- You need native experience to have a polished app
- Dependency stack is still obscenely big
- Opens native engineers to more projects, makes yours more welcoming to others
- Problems do, and will occur, but everything is fixable by forking
- Extensive communication with native code gets tricky to test and maintain
- We ended up re-using quite a lot of existing native code
- It makes working in native code feel more like a chore, as you lose the JS developer experience
- Spending so much time in another environment will erode native knowledge
- Makes a lot of sense in an additive approach to existing apps
- We're not making plans to re-write other Apps into React Native, they are fine as-is
- New apps going forward we will default to React Native apps, unless there is a good reason to not
So, should you use React Native? Maybe. If you have an API driven app, probably.
It's definitely worth a week of prototyping for any engineering team, then if that goes well you should look into GraphQL and Relay. They really make React Native shine.
Want to get started?
If you'd like to look into GraphQL + Relay, but don't want to start building a server yourself, consider these GraphQL as a Services: