from the Drive team:
Hi everyone,
As we've just had the AMA today, I thought it would be a good moment to introduce myself. My name is Andrew and I am the Director of Engineering for Proton Drive. All of the engineering work for the Drive products - our apps on iOS, Android, Windows, MacOS, Linux (!) and Web, along with Docs and spreadsheets and the Proton Drive backend - rolls up to me. I work closely with our Product team to spec out the features but my specialty is the tech, and that's what I'm here to talk about with you today.
So let's get into it! I want to take "a minute" (or 20... it's a long read) to share with the community a little bit about our technical journey over the last 3 years. I'm not going to talk too much about features and product here, mostly it's going to be technical. If that's not your thing, you can skip the rest of this post and go all the way to the end for the punchline.
People have wanted a Linux client for a long time, so I'm going to explain a little bit of why we haven't had one, why we are now building one, what's changed, why we're doing more than that, and what it means.
A Challenging Technical Past
3 years ago the Proton Drive team was in a much worse place than we are today. At a glance we were dealing with quite a few major technical problems at the same time:
- We had completely separate code bases for Android, Windows, and Web. Additionally:
- iOS and MacOS were together but the code was riddled with "if Mac do this, if iOS do that" statements that did not help us.
- Web code had unfortunately allowed some of the data models to intertwine with UX/UI code, making things very brittle and slow.
- Each team was basically a silo universe with little or no interaction with any of their peers, with the exception of Mac and iOS, whose use cases were wildly divergent.
- Incompatibilities between clients was not uncommon and we had quite a few very painful incidents resulting from platform-specific bugs.
- Performance was in a very bad spot on most of the clients.
- We had no consistency in measuring critical metrics.
TL;DR our foundations were shaky. On top of these shaky foundations we were busily cranking out basic features, but those features were often plagued by major bugs and/or performance issues amplified by the shaky foundations. Quality overall was not where it needed to be, but the constant fire-fighting and technical debt made it very difficult to get the momentum to dig ourselves out. Not a good time.
🐧 Linux client status: A distant dream. Not enough resources, not enough time, and no way we were going to try and build yet another client with all this going on.
Maturation and Unification
We recognized that the situation needed to change so we made some big long-term bets, tightened our belts, and got to work paying down technical debt and changing a lot of our internal processes. There are a lot of great stories from this time period but here's a few key changes that we have made over the last 3 years:
- We reorganized the client teams into a Desktop group and a Mobile group, recognizing that it was the use cases and features that should unite people - not the programming language or a specific chunk of shared code. Windows and Mac may have very different programming paradigms but they need to solve the same problems in the same ways. The same is true for Android and iOS.
- We put a strong culture of blameless post-mortems in place. Post-mortems are the analyses we run after we have a problem, where we dive deep into the root cause and take pains to make sure it can never happen again.
- We added a lot of performance/reliability instrumentation and logging (although there is still work to be done here, it's vastly improved).
- We took a zero tolerance policy towards any issue that could result in data loss, forcing the creation of improved tools and processes that can effectively combat even the worst disasters.
- We invested into training up a stable core of engineering managers, with small teams and tight focus.
- And a bunch of various technical rules were set, such as avoiding long-lived branches and increasing the strictness of code reviews, feature flags, rollback plans, automated tests and deployments.
- Our engineers put in a crapload of hours. Thank you all, if you're reading. You're the heroes of this story.
- And along the way we also somehow, by some miracle, managed to also ship Photos, Sharing, Albums, Docs, and Sheets. I'm still to this day, not sure how we managed to do that at the same time.
These changes bought us some breathing space. By improving processes across the entire team and investing more in quality and planning (and spending a lot of overtime hours) we were finally able to arrive at a moment many of us had been waiting for, we were able to extract from one of the clients (Windows) a high-quality core that could serve as the basis for a platform-agnostic Software Development Kit. Here's why this matters:
Our end-to-end encrypted (e2ee) platform makes a lot of things much harder than it is for non-e2ee systems. We have to have a lot of safeguards to ensure that any file encrypted and uploaded by one client is guaranteed to be downloadable and decryptable by all other clients, forever. That is a tall order.
Meanwhile the server is effectively blind to the data it is receiving, since everything is encrypted there's no way to take "easy" paths like "just verify that the checksum of the file on the server matches the checksum of the file on the client" (since the file on the server is a sequence of encrypted blocks, that do not in any way resemble your original file, and since we stream everything, this is a huge pain in the backside).
Until this moment in time, basically every feature we implemented required clients to go through an excruciating process of carefully aligning with one another and performing a ton of automated and manual tests to ensure interoperability, and that took a ridiculous amount of effort. It was also extremely detail-oriented and error prone. In effect, our client engineers spent an absolutely unreasonable amount of time working on uploading, downloading, encrypting and decrypting instead of the features or the experience on their respective platforms.
With an SDK, all this would change.
🐧 Linux client status: Not yet time. We have finally got some breathing space but now we need to capitalize on it and use the momentum to build an SDK, not get distracted with a 6th client codebase that would slide us backwards into the situation we just were about to get out of. Note that this is not an issue of "just hire more people" - rather, it's an architectural necessity.
Ripples Become A Tsunami: SDK Tech Disruption
As we began to work on the SDK, many things started to drastically shift inside the team:
- Teams that had never needed to communicate much on technical implementation details (beyond a few meetings to align on APIs and standards) began to talk to each other and have heated debates about cross-client architecture, leading to increased information exchange and improved relationships between teams.
- Integration of SDK functionality forced teams to clean up architectural messes and design flaws that had lived in the clients from the start, in order to match the interface requirements for using the SDK (basically it forced a cleanup of several design patterns).
- The SDK itself absorbed effectively all of the complex, error-prone code that was responsible for chunking, verifying, signing, encrypting, decrypting, uploading and downloading everything. It also absorbed the complexity of permissions, sharing, cloud filesystem CRUD operations, and more. Clients could now start to significantly slim down their codebases, reducing the cognitive burden on every client developer.
- The SDK could implement a single, fast, properly parallelized implementation that instantly upgraded most of the clients that it was retrofitted into.
For sure along the way we had a lot of challenges. The Crypto library we use today on non-Web clients is Go-based, which means pulling in the Go virtual machine. The SDK for non-web clients is written in C# which meant generating bindings for Swift and Kotlin, plus loading the native library during load on all platforms. For those keeping score at home that means the Mac, iOS and Android clients got to absorb a dotnet library with an embedded Go runtime (scary). Our attempts to unify Web failed, mostly because WASM support isn't quite where we need it to be just yet, so we had to have two implementations of the SDK... one for Web and one for everyone else. Still, 2 implementations is materially superior to 5 (Linux would have made it 6).
The SDK implementation slowly caused a coding renaissance in every client team. Cleanup of technical debt spread along the surface of the code that interfaced with the SDK, leading to a lot of improvements and the fixing of a lot of bugs and performance problems that have lain dormant for years. As we brought the knowledge of Web, Mobile and Desktop clients together we were able to make a much better system than any individual client had previously had, so each client ended up with a powered-up core that also put some pressure on us to clean up and improve the rest of the app. There's still a lot to do here but things are drastically improved in many critical places.
And we didn't make the SDK itself the goal - instead we used it as an implementation choice. Every new feature we needed to build, we built into the SDK. When we had to significantly overhaul code, we replaced it where possible with SDK code. Every single team contributed to the SDK - Android, iOS, Windows, Mac, Web, even the backend team - everyone got involved. Through a sustained, concerted effort we kept the focus there month after month and slowly but steadily built up that core.
🐧 Linux client status: Finally in sight. We set a new goal: to prove that the SDK was up to the right bar, and to force us to "drink our own champagne", we decided that we would build a Linux client alongside the SDK. We set some new and ambitious requirements:
Every HTTP call the Linux client makes to a Proton Drive endpoint has to be made through the SDK, with no exceptions. Anything that the SDK doesn't already do, and is needed by the Linux client, and is platform-agnostic, must be built in the SDK (no shortcuts).
Crypto 1.5 Shows What We Can Do In The SDK Era
This will be a short section but a very important one. The first major change we've used the SDK deliver is our Crypto 1.5 update. It is hard to overstate how incredible this change is: we have upgraded the crypto algorithm on every single Proton Drive client at once. And it's monstrously faster. The new algorithm is hardware-accelerated, which slashes CPU and battery drain on Mobile by an absolutely ridiculous amount. Internal data so far shows that the median improvement of upload speed on web exceeds 500%, though we are conservative in what we communicate since summary statistics can be misleading.
3 years ago this would have been ridiculously hard. We would have needed to coordinate binary releases across 5 different clients with 5 different feature flags controlling 5 different rollouts, and the testing alone would have taken us through the rest of this year. Instead, with the SDK deployed to all clients, we were able to control the entire rollout as effectively a single process, upgrading the entire world together and delivering this improvement to everyone in a fraction of the time it would have otherwise taken us.
And this is just the beginning. Crypto 1.5 is called Crypto 1.5 for a reason, and that reason is because there will be a Crypto v2 that makes even more radical changes that will further ramp up the speed while also reducing complexity. It's going to be another huge improvement.
The SDK allows us finally to make sweeping, impactful changes across the ecosystem. Of course, with great power comes great responsibility (fun fact: I once mistakenly attributed that quote to Winston Churchill instead of Spiderman's uncle, but to be fair it sounds like something he might have said...). We're taking great care to rigorously test every major change to the SDK to ensure that rollouts are smooth. We know that with a unified SDK the blast radius of any error is drastically increased, it's the one downside to unification. Fortunately the clients all have very different rollout strategies and timelines, so we have a lot of gradual processes to act as safety nets in the event that we make any big mistakes (and we have the tools to handle it if we do, thanks to our zero-tolerance policy on data loss).
And I haven't even talked about the increase in upload and download success rates, but that's also been extremely positive. Not to mention a host of other improvements related to performance and reliability.
🐧 Linux client status: In progress! At last! But let's talk a little bit more so read on to the next section...
What Linux Really Needs - A CLI, an SDK, and a Client
In considering the needs of the Linux community, we have heard (very loudly and very frequently) the calls for a Linux client. But anyone who has been around in the Linux community knows that is not necessarily what is actually needed. The Linux community understands the value of software freedom and choice when it comes to their tools and technologies; being forced to use a UI when what you really want or need is a command line, is not great; and no UI or command-line tool will ever be capable of doing everything that the Linux community wants or needs. So we decided that it made much more sense to deliver the Linux community a complete package rather than a narrow, locked-in UI. And it wasn't going to be much of a detour to deliver that much better package. So we will offer:
- A full-powered standalone Linux desktop client that has a sync engine and the same feature set as Mac and Windows (still a ways out but now well underway). And the engineers are in charge of the UI here since the Linux community is more technical in nature than our other environments :) For the things you can't do in the UI, or for the things you want to do in a more automated way we have...
- A comprehensive CLI that can do most(*) of the things a normal Desktop or Mobile client can do except for running a sync engine. This allows scripting and automation via cron, integration into arbitrary workflows, etc. And if that's not enough and you still need more power we have...
- The SDK which completely abstracts away the network interactions to the Proton Drive backend. Using this, you can implement almost limitless functionality on top of the Proton Drive core (though please remember that limits still apply and your app may be throttled if it causes availability problems or behaves in very broken ways)
() When we say "most" here it excludes a bunch of things that are currently only possible on-device, for example the "scan document" capability that the Android application has that relies on features shipped in Android and therefore cannot be made available via the command-line. Features like this will probably never come to the CLI.*
When Will It Happen?
This is of course, what everyone wants to know. I wish I could give you a date for when we will ship the Linux client, but I cannot. We are still early in the development of that client and we have set out some pretty ambitious goals for it as described above. But we've opted to deliver to the community as early as possible, our first versions of the SDK and CLI. We're using the SDK right now in every single Proton Drive client application so, although we aren't yet officially supporting it for third-party development, it's a production-grade core that is currently powering every official client.
What I can say is that we are now most definitely on the right path. The SDK is working very well. Our team has an internal proof of concept demonstrating an initial version of the Windows sync engine running on Linux and powering a UI where you can indeed upload and download files. It's happening! But remember, a proof of concept is just that - there's still a ton to do. We have to implement a whole bunch of features that Windows and Mac provide, then there's all the instrumentation, logs, edge case handling, unit testing, load testing, integration testing, and so on.
We will release this when it is ready.
Conclusion
We will strive to get a Linux beta into the hands of the community as soon as is practical. We know it has been a long wait, and we know you're tired of hearing "soon". But it's still going to be a little while, as we make sure that the client can do what it needs to do in a safe and performant manner. It will of course be open-sourced as well. I will be very surprised if a beta does not happen by the end of the year, but I'm not able to make a promise that this will be the case, there's still a lot of unknowns that need to be explored before we can really settle on the timeline.
I hope that this (very) lengthy post gives the community a little more insight into why it has taken so long, why we have taken the approach we have taken, why you really can now have faith that it will be delivered in the near future. The Linux desktop client is just one part of a three-part technical strategy for Linux that involves the client and the CLI and the SDK, and we're pretty sure it will have been well worth the wait.
If you're a developer and you'd like to discuss the SDK more, check out the Proton official Discord server and join our #drive-sdk channel.
