Kotlin/Native (Stranger) Threads

by Kevin Galligan | May 11, 2018 | Future of mobile, Thoughts

Chapter 1: Getting Started

This is going to be a multi-part series on threading in Kotlin Native. To build a reasonably functional mutltiplatform implementation, understanding threads is critical. The Kotlin/Native design and the way threads are intended to be used is quite different than what we’re used to in Java land. I’ll be learning this as we go, but as many have pointed out, if you really want to understand something, try explaining it to others.

Why Stranger Things theme? KN is opinionated about threading and state. To keep you out of trouble, it creates a world where different threads mostly can’t see each other’s data, and have very little interaction. However, underneath that world is a similar looking, but far more powerful (and potentially dangerous) world where you have access to everything. It’s the Upside Down, basically¹. We won’t be talking about that much right now, but we’ll definitely need to poke around and make some new structures to enable productive app development.

Tools

You’ll need some kind of test environment to run things in. For now I’d suggest downloading CLion. Read this on installing Kotlin/Native support. As mentioned in the beginning, I’m just getting started with this. The same is true of CLion. I hope we’ll be able to do KN in Intellij directly, but CLion appears to be the tool for the bulk of KN and C++ development. That means learning CLANG, at a minimum. Not right away, though. We can poke around for a while without understanding CLANG.

To verify everything is working, create a new KN project and use the HelloWorld sample.

CLion New Project Dialog

Hit the play button in the action panel and you should see “Hello, Native World!” printed to console.

Source

We’ll be trying out some stuff. To run the source, clone this:

Docs

This is probably a good time to step back and say if you don’t have a pretty decent understanding of how threads work in general, you can run through the samples, but you may be lost often. The intro stuff will probably be useful, but its safe to say we’ll also wind up in the weeds periodically. Will update with good intro resources if I come across them.

Anyway, back to KN. Read the CONCURRENCY doc in the KN repo. It explains a lot. Today we’ll go over the basics of Worker, how to pass “hot” data, and what data freezing is.

Worker

Workers are the core unit of concurrency provided for out of the box with KN. I’ll start by saying this:

Each thread, including the main thread and each worker, has it’s own copy of state.

I’ll repeat this a few times in this post, so if there’s an important takeaway, it’s that.

You create workers and pass them jobs to complete. I would imagine there are implementations where this doesn’t correspond to a thread, but right now, for iOS at least, that will boil down to a thread. A worker is a thread with a job queue. Also, remember, it’s own copy of state.

Simple Example

Let’s make a worker and have it do something.

Call ‘startWoker’ to start a worker. This will create the thread and make it available for jobs.

You create jobs with “schedule”. That method takes a TransferMode param, a producer lambda, and the actual job, also as a lambda.

Inside the job lambda is the work you actually do.

Scheduling returns a Future. You can use this to get the job result inside the calling thread. We call ‘consume’ simply to wait until the job is done.

If you’re manually creating workers and leaving that context, shut them down properly with ‘requestTermination’.

Let’s dig into each piece in more detail.

Worker Lifecycle

The worker is a managed resource that you create, and if you’re going to be forgetting about it, you probably want to shut down. That corresponds to two methods: startWorker() and requestTermination().

There’s not much to say about ‘startWorker’, other than it creates a worker queue. ‘requestTermination’ is a little more interesting. By default it’ll finish all jobs in its queue, but you can pass in a boolean ‘false’ to tell it to shut down asap.

If you’re creating workers, you should shut them down. If you’re creating a set of workers that are intended to stick around with your app at all times, don’t worry about it. If your process gets killed, so with your threads.

Consume

The schedule call returns a Future. On that you can check status and get results of the job. ‘consume’ will force your thread to wait for the job to finish, and pass the result of the job into the lambda. In our simple case, we return nothing, so the lambda accepts nothing.

Scheduling Work

Now that the simple stuff is out of the way, it’s time to look at scheduling work. This is where we need to get very precise.

The schedule method takes 3 parameters.

1. TransferMode — Checked or Unchecked. The “safe” option is Checked.
2. Producer lambda — This should return the data that will be passed into your job.
3. Job lambda — This is the “work” that will be run on the different thread. It has a single input param, which is the output of the producer lambda. You can return a value, which can be read by your calling process.

The primary complication around all of these parameters is that your data needs to cross between threads. As mentioned, each thread has it’s own state, and isn’t allowed to see others’ state. How can that happen? To stick with the post theme, you’re sending something through the Upside Down and back again. To do that safely, there are some considerations.

The Lambdas

The goal of how workers interact is to protect state. KN will do some magic to check that the state you’re passing isn’t referenced anywhere else OR that your state is frozen. Frozen is a runtime immutable state that we’ll discuss later.

In other words, KN wants to make sure your state is safe. Either nobody else can see it, or nobody can change it (including you).

How does it check? That’s an Upside Down topic for a different chapter.

Producer Lambda

This lambda returns an instance of data that is passed into the job lambda. It’s run in the same thread that calls it.

Why a lambda you ask? The runtime wants to make super sure you’re not dragging data around. The KN runtime is serious about references. So much so that passing hot data into a worker isn’t necessarily an easy operation.

This will fail because there’s a local reference:

fun failedReference(){
    val worker = startWorker()

    val dat = SomeData("asdf")
    worker.schedule(TransferMode.CHECKED, {dat}){

Even if you don’t plan on doing anything to ‘dat’, the KN runtime will get upset. More curious, this works:

var glob = SomeData("asdf")
fun getAndClearGlob():SomeData{
    val temp = glob
    glob = SomeData("qwert")
    return temp
}

fun globalReference(){
    val worker = startWorker()

    worker.schedule(TransferMode.CHECKED, {getAndClearGlob()}){

After calling ‘getAndClearGlob’, nothing is pointing at the object originally at ‘glob’. That seems logical! How about this?

var glob = SomeData("asdf")
fun getAndClearGlob():SomeData{
    val temp = glob
    glob = SomeData("qwert")
    return temp
}

fun globalReference(){
    val worker = startWorker()

    glob = SomeData("inside")
    
    worker.schedule(TransferMode.CHECKED, {getAndClearGlob()}){

That will fail. You and I know you’re not doing anything with the reference created inside the method, but KN’s worried about it.

All Threads Have Their Own State

I’m reminding you of this again here so I can point out an issue with the sample above. Keeping data at the “global” scope just so you can pass it to a worker is probably not a great design, but because all threads get their own state, the worker will have it’s own copy of ‘glob’ that does nothing but take up space and init/management cycles.

If you really want to ruin your day, try this:

val worker = startWorker()

fun doNotDoThis(){
    worker.schedule(TransferMode.CHECKED, {}){
        for(i in 0..1000000)
        {
            //Let's kill time
            val b = (i+i).toDouble()/1000.toDouble()
        }
    }.consume {  }

    worker.requestTermination()
}

The top level ‘worker’ can be called from multiple methods to start jobs. Nice! Right? Well, ‘startWorker’ creates the worker, which gets its own copy of state, which creates a worker, which gets its own copy of state…

That’s super bad. Got bit by that right out of the gate.

We’ll probably have a much longer discussion of how to structure “hot” data to get it over the wall to other threads. Here are some other strategies.

Freeze

If you’re following general best practices, the data being passed through threads should be immutable anyway. KN has a special method called ‘freeze’ that makes your data very seriously immutable at a runtime level.

fun freezeLocal(){
    val worker = startWorker()

    val localData = SomeData("asdf")
    localData.freeze()

    worker.schedule(TransferMode.CHECKED, {localData}){
        println("In thread ${it.a}")
    }.consume {  }

    println("In main ${localData.a}")

    worker.requestTermination()
}

Freeze actually modifies metadata flags on runtime objects which tell the runtime that seriously nothing will touch the data. It’s more than just having immutable values. You can now pass that frozen object to multiple threads and reference it from your calling thread. No problem!

Well, some problems.

Common Kotlin has no concept of “freeze”. That’s a minor issue. We should be able to do something with extension functions to get a noop method into common and the JVM. More pressing is that KN and everybody else is going to be different at runtime. Could the concept of freeze be ported to the JVM? Would that be valuable? Questions for another time, and certainly somebody else 😛.

A more immediate issue is class design. The SQLDelight demo, for example. SQLiteDatabase and the Helper objects involved all have state that is protected by Java threading constructs. That design is going to need to be rethought.

Or you can just do this:

fun freezeLocal(){
    val worker = startWorker()

    val localData = SomeData("asdf")

    worker.schedule(TransferMode.UNCHECKED, {localData}){
        println("In thread ${it.a}")
    }.consume {  }

    println("In main ${localData.a}")

    worker.requestTermination()
}

No freeze. Just pass in as TransferMode.UNCHECKED. In this case, probably not a huge deal. The String in SomeData is a val. Nothing is changing anyway. In other, more complex cases, we’ll need to be more careful.

Workers?

I’m looking at this through the lens of shared mobile architecture. That means interacting with the host system. You can’t have a worker for the home thread, so out of the box, you’d need to implement some kind of regular Future checking mechanism. That doesn’t sound great.

iOS has thread event queues. That’s how you communicate back to the main thread. In future posts we’ll explore the idea of leveraging the same checks and concepts involved in Worker “transfer” to ensure the data safety and memory management of KN, but actually perform the threading with iOS’s dispatch queues, and be able to post messages back to the main thread. This will almost certainly involve a new library and possibly some PR’s to the KN repo. See how it goes!

For multiplatform, we’ll need some Worker-ike interface that can resolve back to Looper/Handler on Android. In JVM, we won’t have the data checking and freeze stuff, so testing is going to have to be robust enough to run likely scenarios and catch data integrity check issues on the iOS side.

Or just pass everything unchecked and watch out for this guy.

Demogorgon, obv

Future Chapters

Some possible future chapters:

1. Wrapping iOS dispatch queues instead of “Worker”.
2. Rethinking SQLite and (possibly) SQLDelight design. Listeners watch queries, which pass data base to the main thread (ultimately). In JVM, that data can exist across threads, but not in KN (out of the box).
3. Exploring global state options. You can, but that’s like hanging out near the portal.
4. Enter the Upside Down and poke around a bit.

Nice suits

Touchlab

Is looking for orgs that want to get started with Kotlin multiplatform, and hiring mobile developers interested in it. Reach out.

¹ The demogorgon is the memory manager? Probably best not to overthink the analogy. Also, Netflix legal team. Love the show! Please don’t get mad about the images. XOXO

SQLite/SQLDelight ❤ Kotlin Multiplatform

by Kevin Galligan | May 4, 2018 | Future of mobile, Thoughts

A few weeks back I posted a piece about Mobile Oriented Architecture and how I think coding products for screens will evolve. At the end, I said I was keeping an eye out for the Kotlin/Native multiplatform commit.

I was. Really. Read the commit emails every morning. Yet somehow I missed the fact that multiplatform was possible.

You can build multiplatform projects using Kotlin/Native. I won’t say it’s super productive yet, but you can do it. I gave a talk about it last night, and have some code to share.

Update, Video!

TL;DR

If you want to run SQLDelight on iOS, skip to “One more thing” below. Anyway…

Kotlin Mobile Today

Kotlin multiplatform for mobile consists of common Kotlin, Kotlin/JVM, and Kotlin/Native. The slide above represents roughly how building multiplatform would look today, from a “how much code are you sharing” perspective. Obviously, the JVM side has a ton of stuff available. On the iOS side, the Native team has wrapped a bunch of resident libraries with the interop, but that’s kind of it. The middle represents what you could theoretically share. Just logic, and to a limited degree (size of that area varies widely according to how much logic, obviously).

To be useful, we need libraries. The essential libraries. We don’t need everything currently available on the JVM side, just the critical stuff. Android is a mature ecosystem, which means you have a lot of special purpose stuff. If you open a production Android app and look at the library list, it’ll probably be huge, but if you really looked at it, the list of things critically necessary for shared architecture would be pretty short.

• Networking
• Threading
• Something reactive
• Properties (like shared preferences. Typed storage for when sql is too much)
• SQL (or at least a “boxy” database. Something with a schema)
• Testing, logging, utilities, yada yada

The lesser used H&H image

All of these need fully Kotlin-ized rewrites. It sounds like a lot of work, but a few things to keep in mind.

Copying is easier than creating

One of the reasons why the discussion around “cross platform” is broken is we think 2 apps means 2x work. It does not.

Product development involves iteration, which involves throwing away a lot of work. You shouldn’t need to do the same on your second platform. Similarly, “smartphone” mobile has a decade of iterating on libraries and best practice ideas. To build a new ecosystem, you can just take the current result of that work and build analogs. It’s non-trivial, but it’s not starting over either.

Ecosystem interest is intense

To say a lot of developers like Kotlin would be an understatement. There is intense interest in creating things for Kotlin. Some communication and coordination will be useful, so 10 groups don’t build the same thing (although a little competition doesn’t hurt either). However, I’m predicting as more examples of multiplatform emerge and the tools stabilize, we’ll see a flood of multiplatform libraries emerge. This will pick up roughly end of summer.

Wide open

If you’ve come into Android fairly recently and wanted to do some open source work, you’ll notice it’s pretty crowded. Getting attention can be difficult. This particular field is wide open but very likely to be very active very soon. It is a great time to get involved.

A bit of my own background. The last several years have been largely focused on the non-technical side of running a mobile development shop. As such, there has been a bit of open source, but nothing super committed. During early Android, however, there were few libraries and I had a lot more time on my hands.

I did not write ORMLite. I did, however, work with the author to make the db interface generic, then implement the Android-ORMLite bridge. It is *very* likely that ORMLite would never have existed on Android, or come much later if that didn’t happen. A lot of apps used it, because, for better or worse, there wasn’t much else available. As long as you don’t make a bad library, you can have a big impact in an early ecosystem, as going from nothing to something moves the needle a lot more than writing something that’s different, but competing with similar things.

What do I think of ORMLite? That’s a different post. Summary, it was generally more productive than straight SQL, and we learned a lot 🙂

Anyway, I’m a database nerd.

Doppl

Doppl is a set of libraries built on top of J2objc, to facilitate sharing logic and architecture between Android and iOS. It works, but pitching a Java/Objective-C thing in the time of Kotlin/Swift is difficult. I also think Kotlin multiplatform has a more meaningful future (webassembly FTW).

We are using it when the situation dictates, but are internally very much focused on Kotlin. On the plus side, “Doppl” was essentially an implementation of Android architecture on iOS. A lot of similar skills, concepts, and in some cases code, can be used.

SQLite on Kotlin

I think Kotlin Multiplatform needs a “driver level” SQLite implementation. By that I mean something that is at least similar to the Android SQLite implementation, assuming there’s no performance impact. There has been a debate about that implementation since the beginning of Android, and a desire to “fix” it, but I suspect that hasn’t happened yet because it turned out to not be a big enough problem. So I’m not going to try that now.

In theory, we have an implementation of Android’s SQLite stack in Doppl/J2objc that can simply be added and typealiased. Once that’s stable and has some solid testing around it, reimplement the J2objc layer in Kotlin.

SQLite on Android

SQLite on Android

All SQLite code ultimately talks to the shared library ‘sqlite’ that exists pretty much everywhere. On top of that, AOSP has C++ that provides methods to push/pull data in and out of the sqlite library. On top of that is the Java interface everybody is familiar with, that talks to C++ with JNI. It’s a fair amount of stuff, but conceptually very simple.

SQLite with Doppl

SQLite with Doppl

To replicate this functionality on iOS, we took the ArOSP C++ and Java code. There was some necessary trimming. For C++, it referenced a lot of libraries available to the broader system, which we truncated to reduce package size. For Java, structures that aren’t relevant for iOS were removed (anything for ContentProvider, basically).

Wishful Thinking Diagram

This is conceptually what Kotlin Mobile Multiplatform SQLite looks like.

J2objc?

I can tell you from experience that most developers are not excited about Java that gets morphed into Objective-C. I will say this: It’s super stable, and as far as numerical distribution, is probably on more phones than the rest of the “code sharing” solutions combined. However, I find it really hard to get past the weirdness in people’s minds.

Beyond that, there are actual mechanical issues. Just adding it means 5–10m extra download size. If you’re using it for lots of your app, that’s acceptable. For a library, not so much. Also, while it comes from Java and in its own ecosystem feels like Java, I would call it the uncanny valley of types. The interface between Kotlin and J2objc requires a lot of little adapters.

Wishful Thinking Diagram

Early on my thinking about how to quickly get Kotlin multiplatform working was to use J2objc as the stand-in for the JVM. It is, quite literally, a Java JRE implemented in Objective-C. However, don’t. Calling into specific APIs is totally doable, but as a generic solution, it really wouldn’t go well.

Anyway, the “Wishful Thinking” diagram actually looks like this.

It’s complicated

This diagram isn’t to scale. The majority of the heavy lifting is in J2objc-land, but compared to the simplicity of the Android side, there’s a lot of adapters and delegates going on.

Over the next few weeks, the plan is to reduce that down by writing Kotlin that talks directly to the C++. Something like this.

For today, though, the implementation we have works. I don’t think there’s a significant performance impact, if only because we’re ultimately pushing something to disk, so a couple extra function calls aren’t going to do much.

One more thing…

The big talk “reveal” is that I wasn’t going to actually demo SQLite. I’ve been chatting with Alec Strong about SQLDelight and it’s Kotlinization. As it turns out, he’s been rewriting the output to generate Kotlin, and not just JVM Kotlin. Common Kotlin. They don’t have a Sqlite interface to run it on, and that’s what I’m working on.

So, the “demo” is actually…

Kotlin Multiplatform SQLDelight

The Kotlin output on SQLDelight is pretty amazing, and the separation of a driver interface will allow SQLDelight to be run in non-Android contexts pretty easily. On the SQLDelight side, I only had to change one thing. SqlPreparedStatement bindLong and bindDouble use ‘long’ and ‘double’ as their parameter names, which blew up in Swift. That’s a Kotlin/Native framework generator fix (either escape or fail upstream), but I just did a workaround to get through the day.

The Kotlin/Native compiler, AFAIK, doesn’t have a package manager and needs source to build, so I just grabbed the runtime code for SQLDelight and threw it into the driver build. This works, but SQLDelight with Kotlin is in SNAPSHOT mode and under heavy development. I blew up the demo 2 hours before start by doing a clean build. So, there’s a good chance this won’t compile when you try it. I’ll add a version that doesn’t depend on remote code soon, but flying out to I/O in the morning, so unlikely till the following week.

The code

Steps in README. Sample was taken from Kotlin/Native calculator.

Future

The demo is open source now. We want to do some more work on the driver code before opening that. Kotlin/Native is *very* different with regards to threads, and the SQLite Android code does some thread management internally. That needs a think. It’s not that bad, but you can’t just take the same logic and move it over. Also, a more complete testing strategy needs to be implemented. It’s first priority after I/O.

Whether *this* Sqlite library is what ultimately gets used for SQLDelight on iOS, or if somebody releases something more purpose-built, is less clear, but see up top about a little competition.

Predictions

I’m making some predictions, which is always a bad idea, but here goes.

Late-summer: Several good examples and usable libraries. My goal is to have Droidcon NYC’s app be fully Kotlin shared architecture.

KotlinConf 2018: I’ll just assume JetBrains is going to push hard to have something really solid by then. If only a psychological milestone.

Mid-2019: Mainstream. I have no idea of actual adoption, but if you want to share a bunch of code on mobile, the tools, and at least the critical libraries will be available.

Say Hello!

We (Touchlab) are very much looking for orgs that want to get started asap with Multiplatform. We also build apps without the fancy shared code. If interested please reach out.

We are also hiring. Mostly right now that’s standard app work, but if you are interested in Kotlin and Multiplatform, this is a great place to be.