Reverse State Monad in Scala. Is it possible?

Hello all!

In this post we’re going to have some fun with a mind-breaking thing called Reverse State, and explore the limits of laziness in Scala along the way.

When I see something interesting implemented in a “foreign” programming language, I often have a desire to port it to Scala – just out of pure wondering how it would look. And sometimes using familiar language also allows to deeper understand the concepts presented. Some time ago I did it with a great book called “Neural Networks and Deep Learning”: here are most of the exercises from the book written in Scala.

This time a totally different thing caught my eye. It was a really nice article about Reverse State monad and it’s implementation in Haskell. I have never heard about it before, so implementing it in Scala seemed like an exciting exercise.

And it didn’t disappoint, although outcome was not the one I expected 🙂 So I decided to make it a story instead of just plain code. I’ll use the technique that Eugene Yokota applied in his “Learning Scalaz”: we’ll follow the source (in Eugene’s case it was Learn You A Haskell For Great Good) piece by piece, discussing and writing the code along the way. Let’s go!


To really follow along, reader should be familiar with Monoid and Traverse typeclasses as well as with State monad in Scala. Also a quick read of the original article won’t hurt.


There’s a big introductory part that’s dedicated to different ways of scanning a data structure. Let’s not skip this block and use it as a warmup.

Given a list of Ints, can you produce a cumulative sum of those integers? For example, if we had the list [2, 3, 5, 7, 11, 13], we want to have [2, 5, 10, 17, 28, 41].

There are actually many different ways to write this function. Depending on your taste on imperative programming, you can choose anywhere between highly imperative ST-based destructive updates and an idiomatic functional style.

This should be simple, there’s scanLeft in Scala std lib:


For example, what if you want to produce a cumulative sum that is accumulated from the right? So if we had the list [2, 3, 5, 7, 11, 13], we want to have [41, 39, 36, 31, 24, 13].

Pretty much the same thing here:


As a Haskell programmer, we have the instinct to generalize things.

Say no more. Scala developers love it too. We’ll use type classes from cats, but scalaz would work the same way:

But how might we implement such a function? Let’s consider cumulative first. What we really need is to keep track of the running sum as we traverse, and then returning the running sum as the new value. The State monad then becomes helpful.

State is a well-known concept in Scala, so we can easily follow-up with our implementation of cumulative:

Comments are required here I guess. “Traversing with State” is a powerful technique to go over some data structure, while accumulating information along the way. Processing of each next element allows you to modify the accumulated State “effect”. In this case we’re just accumulating the running sum (according to the provided Monoid) and using the same sum as the result value of the state calculation.

runA is analogous to evalState in Haskell – it evaluates the thing and returns just the result value (ignoring the accumulated state). And, since for stack safety reasons State calculations in cats are wrapped into Eval, we have to execute it to get our value out.

Ok, let’s now, using the original article’s help, try to implement cumulativeR with Reverse State.

Enter the reverse state monad

As mind-boggling as it is, let’s try to digest this definition:

The reverse state monad, on the other hand, has the same API, except that you can set the state, so that the last time you ask for it, you will get back the value you set in the future.

Image result for smart guy meme

Oh man… Well, let’s try to at least port the provided implementation to Scala. But I’ll change two things with comparison to the original:

  1. I’ll swap the results in the signature of the runF function to be consistent with normal State in cats, where the state is returned on the left, and the value is on the right.
  2. To implement cumulativeR we only need an Applicative, so I’ll not try provide an instance of Monad for ReverseState at this moment.

So this is our ReverseState applicative. We’re drowning in Evals, but that’s the cost of stack safety: everything here is really similar to the original State in cats, except for the ap function.

And, actually, no “clever use of laziness” is happening in the ap. Seems like it will show up in flatMap, but so far we’re fine without it  –  cumulativeR implementation works already:

We can check that it’s output is equivalent to the scanRight-based implementation:

Of course, due to laziness, similar example in Haskell will not calculate anything until we explicitly ask for an element of the list or trigger the evaluation somehow else.

Let’s go ahead and implement the Scan generalization as presented in the original.

Can we do better? Can we generalize this to more kinds of “cumulative” operations? What if, instead of a simple running sum, what if we want a running average? Or a running standard deviation? Or some entirely new thing such as the running maximum multiplied by the minimum? The only difference between all of those tasks is that the specific state transforming function (the function that was passed to ReverseState) is different.

Since we don’t have proper universal quantification in Scala, I’ll just lift the x into the type parameters list and name it S (state). I could make it closer to the original using shapeless polys, but that’s not the topic of the post.

Here, we simply unwrap a given state monad action, wrap it again in our ReverseState, do the traversal, then unwrap it again.

I find it beautiful! And it works, although I decided not to present standard deviation and max*min scans here. The former would require a lot of math and the latter needs proper composition abstractions for Scan, which fall out of the scope of this post.

So that’s it! We implemented everything introduced in original post, and we did it in Scala! Except for…

FlatMap! Where is my FlatMap ???

The true power of state lies in the ability to sequence stateful computations using bind (or flatMap as we know it in Scala). But does it work for ReverseState?

In Haskell it definitely does. Laziness of Haskell runtime allows bind to be a finite computation. Let’s take a closer look to the definition from the original article:

It’s clear that there’s a circular computational dependency between future and a: each of them is calculated in terms of the other. But that is fine – as long as we operate on finite data, at some point next “future” state won’t be needed and Haskell runtime will evaluate only as much as required for the result to be produced.

So what about Scala?

I would be happy to be proven wrong, but after hours of thought and experiments, after trying to wrap pretty much every tiny thing in Eval, I came to conclusion that there’s no possible way to implement flatMap for ReverseState in Scala.

Although there’s a way to encode a circular dependency in Scala, there has to be an explicit exit from the “loop”. In other words, computation of such a circular dependency in Scala will only complete when under some runtime condition the dependency is gone. The reason is simple –  JVM runtime is strict, thus it can’t suspend computations, that are not needed right now.

This restriction still allows some pretty interesting laziness tricks, like loeb function, for example. But let’s take a look at how an implementation of flatMap for ReverseState might look like in Scala:

The circular dependency in the result is unconditional – the next leg of calculation is created regardless of any previous results.

Eval won’t help here, because to work inside Eval we need to sequence it with flatMap. So we won’t even be able to construct our Eval computation, since it would require circularly dependant flatMap calls on Eval. The flatMap calls themselves are eager and there’s no way to avoid that.

So, depending on whether we wrap the result into Eval.defer, we either get an infinite loop or a stack overflow for programs that involve flatMap-ing ReverseState.

Seems like we reached the limits of laziness in Scala here.


There’s one case though where flatMap for ReverseState will work properly in Scala. It’s when your state type S is a lazy data structure (a standard Stream, for example).

It may seem like some random exceptional fact, but actually it’s the same case of providing the runtime with a condition to stop evaluation and break the circular computational dependency. This time it’s just less explicit and takes the form of Stream‘s laziness.

Thanks to Oleg Nizhnik (@odomontois) for pointing me in this direction.


In this post we found out, that ReverseState is not a Monad in Scala. Again, I would really love to be proven wrong here, so if you happen to find a working instance – please, ping me!

It’s not a Monad, but it’s an Applicative, which means we still can use it in some meaningful computations 🙂

As an example of such, we looked at right-to-left stateful traversals. Big thanks to Zhouyu Qian from Capital Match for his post about ReverseState in Haskell, that served as a foundation for the post you just read.

Thanks for reading!


Interactive playground with all of the presented code is available here.

Writing a simple Telegram bot with tagless final, http4s and fs2

Hello all!

Recently I’ve started diving in fs2 and http4s. They looked so awesome, that to properly introduce myself I decided to implement something interesting. A bot for Telegram messenger seemed like a nice idea. So here’s a small tutorial on how to implement a bot in a purely functional way using tagless final encoding, fs2 for streaming and http4s client to talk to telegram API.

Most of the code snippets here have comments, that explain in more detail what’s going on. I encourage the reader to not skip them — many small details are not covered in the text. Otherwise it would be twice as large.

Disclaimer 1: There are feature rich telegram bot libraries out there. If you need something to be done quickly and you don’t mind bringing akka to your dependencies — you’ll be better off with those solutions.

Disclaimer 2: I’m in no way an expert in both fs2 or http4s, so if you find a more optimal way to do something presented here — please, leave me a comment!  🙂

“Gimme da code!”. Here’s the github repo.

So what are we going to build?

A todo-list bot. To not overload the tutorial with specifics of Telegram Bot API, we’ll make it very simple:

  • It will keep a separate todo-list for each chat (either personal or group one)
  • Bot will be “long polling” the messages that were sent to it. I selected long polling, because it’s simpler to implement and, more importantly, it works in development mode.
  • The interface will look like this:
    • /show command will make the bot answer with the current todo-list.
    • /clear command will erase all the tasks for this chat
    • any other message will be interpreted as an instruction to add a new todo-list item (with message content used as list item content)

So in the end interaction with the bot should look like this:

User interface of a simple todo-list bot


Let’s proceed to designing the algebras.

Designing Algebras

Despite the simplicity of the bot, there’s quite a bunch of algebras we’re going to operate with:

  • A simplified Telegram Bot API algebra, that will only contain the requests we need
  • Some kind of storage algebra, that will handle the storage of our todo-lists
  • Logger algebra for purely functional logging
  • Higher-level todo-list bot algebra, that will be composed out of other algebras above

Logger algebra

We’ll start with the simplest one. Logging is quite a common task, and there are projects already that provide logging algebras out of the box. We’ll use log4cats.

For those who are wondering, a simplified example of Logger algebra would look like this:

Log4cats provides several implementations, we’ll pick slf4j.

Storage algebra

Storage algebra should also be quite simple to digest:

Basically, for each particular chat, it allows us to add items, query the whole list and erase it.

In a real world app, this algebra would be interpreted into some database storage service. But for our purposes an in-memory implementation will be enough.

For purely functional and asynchronous concurrent access to shared state, we’ll use fs2.async.Ref. We could use Ref from cats-effect 1.0.0-RC2, but at the moment of writing it was not possible due to http4s depending on incompatible version of cats-effect.

Ref will store a Map[ChatId, List[Item]], which is essentially what we want to store. Here’s the implementation:

Telegram Bot API

Telegram Bot API algebra will look this way:

Yep, for this bot we’ll only need to poll incoming messages and post responses.

Notice, that I introduced a separate effect S[_] for streamed result. The idea came from this recent article. While in this particular example we’ll only use fs2.Stream[F, ?] as the streaming effect, it makes sense to abstract it out anyway. As far as I understand, abstracting over streaming effects in Scala is an area of research, so in future we may find out a better way to work with streams in tagless final encoding.

Here goes one of the most exciting parts of this post. Let’s implement bot API algebra!

First let’s see what it is constructed from. There’s a comment for each constructor parameter.

Now let’s implement the logic using this toolkit. sendMessage is a simple one, so we’ll start with it. It’s going to call the endpoint with the same name:

Here we’re using query parameters flavour of the API, which is not very convenient in general. For production bots there are other ways, including JSON.

Also, we don’t care about result body here, so we pass Unit  type to the client.expect  call.

Now to the juice — streaming updates.

Telegram Bot API  — stream updates

In plain words, what we’re going to do is to repeatedly call the getUpdates endpoint. Since getUpdates  returns a list of updates, we’ll also have to flatten the result of each call. Finally, our stream has to be aware of the last requested offset — we don’t want to receive any duplicates.

So let’s begin step by step.

repeatedly call

Here we create a stream of single Unit  value, repeat  it and lift into our effect F  using covary.

Here we don’t add any delay between stream elements, because we’ll instead rely on long polling timeout for throttling, which is recommended in the API docs. Just in case you wonder — throttling an fs2 stream is very simple.

Next step.

call the getUpdates endpoint


be aware of the last requested offset

This feels like a stateful stream stage. Moreover, it also requires to run side-effects to obtain the new state — new offset can only be obtained from the pack of fresh updates.

Let’s dive a little deeper here. For pure stateless mapping fs2 defines a map  function on streams:

Then, if we need to run an effectful mapping, we use evalMap (notice the mapping function is now a Kleisli-like effectful function)

If we want to go from a simple map to a stateful map, there’s mapAccumulate. Now you need to specify initial state and a way to obtain the new state after mapping each element:

Notice also, that the resulting stream for each incoming element emits both the stream state after the mapping and the mapped element itself.

Probably you see where it’s going — we need both! Each of our polls needs the latest offset. But to update the offset for the next request we have to execute the poll! So we want something like mapAccumulate, but the calculation step has to be effectful.

And there’s such combinator, it’s (quite expectedly) called evalMapAccumulate:

Now we’re ready to proceed with our stream of updates:

requestUpdates stage has to do 2 things: get the new messages from the Telegram Bot API and calculate the new offset:

Good! One last step remains.

flatten the result

Each response contains a list of updates. For a better API user experience we’d like to have each update as a standalone stream element. This is quite simple to achieve with flatMap:

Basically, we transform each polled bunch of updates into a separate stream of updates (using Stream.emits ), and concatenate them together into a single flattened stream.

And that’s it. Types line-up, our stream of updates is ready to be processed by some domain specific algebras.

Todo-list bot logic

Now that we have all the lower level machinery in place, let’s develop the business logic.

First of all, let’s define all the possible inputs to our bot. We’ll also need a way to convert raw incoming messages into domain specific commands:

And now we have everything we need to define todo-list bot algebra:

This is a higher level overview. The only interaction this algebra provides is to launch the bot process. The process itself is a simple stream that polls the updates (using the lower level bot API) and handles each update with an effect — hence the evalMap we’re already familiar with.

Let’s take a look at both stages.

pollCommands is quite simple: we start long polling from zero offset and map all non-empty messages into domain commands. Conversion involves some hoops since message.text is an option.

Also, in a real world implementation after each processed command we’d persist the offset somewhere, so that when bot is restarted, that offset would be used instead of zero. But, actually, it’s not that bad — once getUpdates was called with some non-zero offset, long polling API marks all the commands before that offset as read, and they are no longer served, even if you then call the API with zero offset.

handleCommand  is the place where we’re going to invoke our storage algebra. This one should be pretty clear, despite the verboseness:

All commands are handled in a similar fashion: update the storage and trace the change to the log.

There’s one small caveat with addItem. I want to make the bot a little less boring than a stone, so why won’t it have several different answers to choose from when an item is added?
Such cases are when we should carefully track and suspend side-effects. Random number generation is a side-effect, so we use the Sync[F]  to suspend it until the “end of the universe”.
Also, List.head  is unsafe, so we use F.catchNonFatal  to lift the error into F . It’s possible, because Sync[F]  extends MonadError[Throwable, F]:

Wiring up and “launching” the bot process

Now we have everything to construct our bot.

Quite a lot going on here:

  • We derive the JSON decoder for API response using auto derivation from circe generic
  • Http client is created safely using This produces a stream of a single element, and guarantees that all allocated resources (connection pool) are released when the stream either finishes or crashes.
  • We flatMap the http client stream into our todo-list bot process stream. To create it, we cook all the ingredients. Some of these “cooking” steps are also side-effectful: for example the Logger interpreter or the  Ref instance for our storage algebra.
  • Stream.force  in the end just allows to go from F[Stream[F, A]]  to a Stream[F, A] that we’d like to return

End of the universe

We can be really proud of ourselves — we managed to develop a nice useful computation without even specifying our effect type, let alone executing any side-effects! That’s cool.

Now we’re ready to pull the trigger: specify our effect type and actually run the computation. We’ll use fs2.StreamApp[IO]  helper so that we don’t even have to write “unsafe” with our own hands.

To test the bot out, you’ll have to create a one for yourself. It’s a no-brainer, just follow the instructions here. Once you get a token, just plug it in and you’re ready to go!

To not expose my test bot token, app grabs it from TODOLIST_BOT_TOKEN env variable. If that doesn’t fit you, just put your token directly into the  TodoListBotProcess  constructor instead.


So it turns out that you don’t need akka or other complex frameworks to build a Telegram Bot. Moreover, you can do it in a purely functional way, and it actually looks beautiful! And as a bonus — the process of writing a tagless final program is a real joy, I definitely recommend to try it out 🙂

Some ideas to implement as an exercise for interested readers:

  • Allow to edit items (through editing original messages in the chat)
  • Wrap the bot with an administration http api, to send announcements for example

That’s all I have this time, thanks for reading!

Multiple faces of Scala Iterator trap

This post is about an issue with Scala collections I encountered recently. Although it is purely an issue of Scala standard library, I haven’t faced it for almost 3 years of my Scala development experience.

The consequences of not being familiar with the matter can lead to very painful results: unexpected stack overflow exceptions in your application.

The core

This issue can arise in variety of different forms, so let’s first look at the core: the simplest reproducer. It involves folding over a standard Iterator:


As any Scala developer knows from day one, foldLeft is stack safe. But don’t let the feeling of safety trick you here. This code blows up the stack:


And, of course, it is not foldLeft. The biggest “WTF?” moment here is that stack overflow is caused by this trivial-looking call:


How is that?

It appears, that for Iterator most transformation methods, like map, filter, etc., do not immediately evaluate the resulting collection. Instead, they build another Iterator on top of the original one.

Calling these transformations several times in a row builds a chain of Iterators. Each one delegates in some way to the respective underlying iterator. Let’s look for the evidence in the source:


As you can see, hasNext of the new iterator just calls the hasNext of underlying one. At this moment the vulnerability should be pretty clear: it’s possible to build a chain of arbitrary depth, that can eventually implode.

What’s important to note, that it’s not the transformation itself, that blows the stack. To make the iterator crash you have to call something that triggers the chain. It can be something as simple as .toString or .isEmpty. As we will see later, it can make debugging harder.

The faces

In hindsight, this issue looks straightforward and even obvious. However, it can take so many forms and shapes, that it can be a total riddle from the first look. Debugging experience can be very painful, especially when you are not familiar with the issue before it comes to the surface. I had that painful experience, so let’s dive into details.

I guess, this kind of behaviour is fine for Iterator by it’s very nature. Also, one can fairly say, that Iterator is not a frequent citizen of our code bases. I would agree — I can’t recall using it directly. But things are not so simple.

There are places, where underlying Iterator can leak out of well-known standard collections. I, personally, burnt my fingers with mapValues method on standard Map.

Let’s now take it to the next level. The foldLeft reproducer we’ve seen is very focused. Everything is in one place and, thus, quite simple to debug. But with advent of Reactive Programming, various distribution abstractions like actors and streams are used more and more often.

What it means, is that you may have the very same Iterator trap, scattered over multiple stages of your stream. Or some actor can silently build up an iterator chain and then message it to some other actor, which triggers evaluation and blows up the stack.

I my case, an akka-persistence-query stream crashed after several thousands of replayed events. It’s quite mind boggling, when you see a stack overflow error inside the stream stage, where there’re no signs of infinite recursion or complex calculations.

To illustrate, here’s a simplified version of my case. We build an event-sourcing application with CQRS. When streaming events to the query-side, we want to ensure, that each event is processed exactly once by each consumer.

For that purpose, there’s a small Deduplicator stage, which accumulates a map of last processed event numbers. This allows the stage to mark each event as an original or a duplicate separately for each consumer:


Although not trivial, the code is reasonably simple. And it is a time bomb.

Notice, that deduplicated value is constructed using mapValues [line 20]. Then, after another mapValues it becomes the stage state for the next stream element [line 26].

So this is effectively the same fold, just an asynchronous one. After some number of events the Iterator chain becomes larger, than our stack can fit in. The bomb is ready.

Last ingredient for the explosion is the trigger. And our innocent Deduplicator guy just passes the trigger down the stream [line 28]. The actual crash happens somewhere in the consumer code, which happens to access the Map.

The fix

One can fairly argue, that Deduplicator is not optimally implemented and can be more efficient. But for the sake of our study, let’s fix it without rewriting.

The solution to an overly long Iterator chain is to break it. First idea that comes to mind is to trigger evaluation of mapValues result immediately.

There are many ways to do that. For example, there are several good suggestions in this SO thread. Developers from Lightbend argued, that .view.force call is the best way to “evaluate” the collection.

Indeed, in our example, calling deduplicated.view.force produces a fresh map, without any ticking time bombs inside. Stack overflow threat is gone.


So this is it. Now any time you’ll see a stack overflow with Iterator involved, you know what to look for.

While debugging this issue, I was lucky to have access to paid Lightbend subscription. The support was incredible, I saved tons of time using their experience and guidance, while looking for the root cause and eliminating it.

That’s all I have today, thanks for reading!

Scala.js In A Big Web Application Talk

My first talk ever at a big IT conference was about my production experience with Scala.js. It nicely concluded a full year of extensive Scala.js frontend development, that my team was doing at Evolution Gaming starting from May 2016.

I tried to lay out various pros and cons of using Scala.js in a big browser SPA, based on that real experience. I also tried to go into details and provide examples for most of the points.

Talk took place on 16.05.2017 at Riga Dev Days conference in Riga, Latvia.

Slides (english, Speakerdeck)
Video (48m, english):

Cross-platform polymorphic date/time values in Scala with type classes

At Evolution Gaming me and Artsiom work on internal scheduling application, that has a huge ScalaJS frontend. We have to deal with lots of complex date/time logic, both on backend and browser sides.

I quickly realised, that sharing same business logic code across platforms would be a massive advantage. But there was a problem: there were (and still is) no truly cross-platform date/time library for Scala/ScalaJS.

After a small research I settled with a type class-based solution that provides cross-platform java.time.* -like values with full TimeZone/DST support. In this post we will:

  • take a quick look at the current state of date/time support in Scala/ScalaJS;
  • see how to get cross-platform date/time values today with the help of type classes.

Described approach works quite well in our application, so I extracted the core idea into a library, called DTC. If you’re a “Gimme the code!” kind of person, I welcome you to check out the repo.


I assume, that reader is familiar with ScalaJS and how to set up a cross-platform project. Familiarity with type classes is also required.

The Goal

There’s no solution without a goal. Precise goal will also provide correct context for reasonings in this article. So let me state it.

My primary goal is to be able to write cross-platform code that operates on date/time values with full time zone support.

This also means that I will need implementation(s) that behave consistently across JVM and browser. We’re Scala programmers, so let’s choose JVM behaviour semantics as our second goal.

Current state of date/time in Scala and ScalaJS

So we have our goal, let’s see how we can achieve it.

In this section we’ll go over major date/time libraries and split them into 3 groups: JVM-only, JS-only and cross-platform.


We won’t spend too much time here, everything is quite good on JVM side: we have Joda and Java 8 time package. Both are established tools with rich API and time zone support.

But they can’t be used in ScalaJS code.


We’re looking at JS libraries, because we can use them in ScalaJS through facades. When it comes to date/time calculations, there are effectively two options for JavaScript: plain JS Date and MomentJS library.

There’re things that are common for both and make them quite problematic to use for our goal:

  • values are mutable;
  • semantics are different from JVM in many places. For example, you can call date.setMonth(15) , and it will give you a same date in March next year!

There’s also a JS-Joda project, which is not so popular in JS world, but has much more value to JVM developers, because it brings Java8 time semantics to Javascript.

JS Date

JS Date is defined by ECMA Script standard and is available out of the box in any JS runtime. But it has several weaknesses. Major ones are:

  • quite poor API;
  • time zone support is not universal: behaviour depends on environment time zone, and you can’t ask for a datetime value in an arbitrary zone.

Since JS Date is a part of language standard, ScalaJS bindings for it are provided by ScalaJS standard library.


Despite MomentJS values are mutable and still have minor bugs in calculations, it’s quite a good library, especially, if you need full time zone support.

It also has a ScalaJS facade.


JS-joda is implementation of a nice idea: porting java.time.* functionality to Javascript. Though I’ve never used this project, it looks like an established and well-maintained library.

ScalaJS facade is also in place, so you can definitely give it a try in your Scala project.

The only problem with regard to our goal is it still lacks proper DST support. But it’s already in progress, so you can expect it to be fully functional in observable future.

Cross-platform date/time libraries

After a small research, I found three options. Let’s see them in detail.


This library is the future of cross-platform date/time code. It’s effectively Java 8 time, written from scratch for ScalaJS.

At the time of writing this post, scala-js-java-time already provides LocalTime, LocalDate, Duration , and a handful of other really useful java.time.* classes (full list here).

It means, that you can use these classes in cross-compiled code and you won’t get linking errors: in JVM runtime original java.time.* classes will be used, and JS runtime will be backed by scala-js-java-time implementations.

Problem here, is that we need LocalDateTime and ZonedDateTime in ScalaJS. And they are not there yet.

Spoiler: we’ll be using scala-js-java-time in our final solution for the problem.

Scala Java-Time

Scala Java-Time is a fork of ThreeTen backport project. So it’s main purpose is to provide java.time.* -like functionality on Java 6 & 7.

It is also compiled to ScalaJS, which means we can write cross-platform code with it. And we can even use (to some extent) LocalDateTime!

The only problem is it doesn’t support time zones for ScalaJS yet (providing this support is the main focus of the project now, though).

So this library is close, but still misses our goal by a tiny bit.

Soda time

Soda time is a port of Joda to ScalaJS.

It’s in early development stages and also doesn’t have time zones in ScalaJS, but I still added it to the list, because developers took an interesting approach: they are converting original Joda code with ScalaGen.

So the resulting code really smells Java, but I’m still curious of what this can develop into.

Idea? No, the only option

The reason I’ve given the overview of currently available libraries is simple: it makes clear that there’s only one possible solution to the problem.

There’s no cross-platform library with full time zone support. And for JavaScript runtime there’s only MomentJS, that really fits our requirements. All this leaves us with nothing, except following approach:

  1. We define some type class, that provides rich date/time API. It’s a glue that will allow us to write cross-platform code.
  2. All code, that needs to operate date/time values, becomes polymorphic, like this:
  3. We provide platform-dependent type class instances: java.time.* -based for JVM and MomentJS-based for ScalaJS.
  4. We define common behaviour laws to test the instances against. This will guarantee, that behaviour is consistent across platforms.
  5. MomentJS API is powerful, but it has to be sandboxed and shaped to:
    • provide immutable values;
    • provide JVM-like behaviour;
  6. There’s a limitation, that we can’t overcome without some manual implementation: both JS libraries don’t support nano seconds. So we’ll have to live with milliseconds precision.

We won’t go over all of these points in this article. DTC library does the heavy lifting for all of them. In following sections we’ll just glance over the core implementation and example.

DateTime type class

Let’s just take a small subset of java.time.LocalDateTime API and lift it into a generic type class. We’ll use simulacrum, to avoid common boilerplate:


First of all, a total order for DateTime values is defined. So we can extend cats.kernel.Order and get all it’s goodies out of the box.

Second, thanks to scala-js-java-time, we can use LocalTime and LocalDate to represent parts of the value. Also, we can use Duration for addition operations.

For now, let’s just view it as a glue for LocalDateTime. We’ll get to time zone support a bit later.

Cross-compiled business logic

Having our new type class, let’s define some “complex” logic, that needs to be shared across JVM and browser.


With syntax extensions in place, the code looks quite nice.

More over, you can notice, that nothing here says, if time should be local or zoned. Code is polymorphic, and we can use different kinds of date/time values, depending on the context.

Now let’s get to the flesh and bones: type class instances.

Type class instances

Let’s start with JVM instance, as it’s going to be quite simple. Follow comments in code for details.


With MomentJS it’s going to be much more interesting, because we’ve obliged ourselves to provide values, that are comfortable to work with for a functional programmer.

To enforce immutability, we won’t expose any moment APIs directly. Instead, we’re going to wrap moment values in a simple object, that will be immutable:


Several notable things here:

  1. We make both constructor and underlying value private to make sure there’s no way to modify object internals. We’ll provide a custom constructor later.
  2. Notice month value adjustment to provide JVM-like behaviour. You will see much more of such things in DTC, I even had to write a couple of methods from scratch.
  3. To compare two moment values, we use their raw timestamps.

Now it’s trivial to define DateTime instance for our MomentLocalDateTime:


Now have everything to run our generic domain logic on both platforms. I’ll leave it as an exercise for my dear reader.

Now let’s discuss some aspects of making this thing work for zoned values as well.

Time Zone support

Not much time is needed to realise, that we need separate type classes for local and zoned values. Reasons are:

  • They obey different laws. For example, you can’t expect a zoned value to have same local time after adding 24h to it.
  • They have different constructor APIs. Zoned value needs time zone parameter to be properly constructed.
  • Zoned values should provide additional APIs for zone conversions.

On the other side, most of querying, addition and modification APIs are common to both kinds of date/time values. And we would like to take advantage of that in cases we don’t really care about a kind of the value and wish to allow using both.

This leads us to following simple hierarchy:

  1. LawlessDateTimeTC (which we initially called DateTime) that contains common methods, specific to all date/time values.
  2. LocalDateTimeTC and ZonedDateTimeTC will extend LawlessDateTimeTC and provide kind-specific methods (constructors, for example).

This -TC suffix is ugly, but name clash in JVM code is worse :).

We will also have to provide a cross-compiled wrapper for time zone identifiers, because java.time.ZoneId is not yet provided by scala-js-java-time, and we don’t really want to pass raw strings around.

Everything else is just an evolution of core idea. Full implementation and more examples are available in the DTC repo.

Note on polymorphism

A side-effect of this solution, is that all your code becomes polymorphic over the specific date/time type. While most of the time you’re going to use single kind of time (zoned or local), there are cases when polymorphism becomes useful.

For example, in an event-sourced system, you may require zoned values for most of the calculations within the domain, as well as commands. But, at the same time, it can be a good idea to store events to journal with UTC values instead.
With type class-based approach, you can use same data structures for both purposes, by just converting between type parameters of different kinds.


Though polymorphic code can look scary for some people, described approach give us following advantages:

  1. Truly cross-platform code, that operates on rich date/time APIs with time zone support.
  2. Polymorphism over specific kind of date/time values.

If you’re working with date/time values in Scala on a daily basis, please, give DTC a try and tell me what you think!

Thanks for reading! 🙂

Published slides from my “Recursive implicit proofs with Shapeless HLists” talk

Recently I was giving an introductory talk about constructing implicit proofs with Shapeless at Scala Riga Meetup.

Contents in brief

  • Analogy between Scala implicit resolution and Mathematical proofs.
  • Recursive implicit proofs (why “recursive” is important feature)
  • HList basics
  • Various implicit proofs for HLists with demos. Mathematical induction as a handy tecnique to writing HList proofs.
  • Expanding HList proof to all product-like types with shapeless.Generic
Slides (english, PDF)
Slides (english, Slideshare)

Implementing type-safe request builder: practical use of HList constraints

Hello, Scala developers!

In this post we will develop a simple type-safe request builder. Along the way, we will:

  • encode domain rules and constraints at compile time with implicits;
  • heavily use HLists;
  • use existing and implement a couple of new HList constraints.

The example in this post is a simplified version of a real request builder from my Scalist project. Implementation of intentionally omitted concepts, like response parsing and effect handling, can be found there.

The code from this article can be found on GitHub.


I assume that you, Dear Reader, have a basic understanding of:

  • implicit resolution;
  • typeclasses;
  • what an HList is.

If some of those are missing, I encourage you to come back later: there’s a lot of good articles and talks on these topics out there in the web.

The problem

We will be developing a query request builder for and API that is capable of joining together multiple logical requests and returning all results in one single response. A great example is Todoist API: you are allowed to say “Give me projects, labels and reminders”, and the API will return all of them in a single JSON response.

Such an optimization is great from performance perspective. On the other hand, it imposes difficulties for a type-safe client implementation:

  • you need to somehow track the list of requested resources;
  • return value type must reflect this list and not allow to query for something else.

Those problems are not easy to solve at compile time, but, as you might have heard, almost everything is possible with Shapeless 🙂

The task

The task here is to implement a request builder with following features:

  • allows to request single resource like this:
  • allows to request multiple resources like this:
  • doesn’t allow requesting duplicates. This will not compile:
  • doesn’t allow requesting types that are not resources. This will not compile:
  • execute method will return something that will contain only requested resources and will not allow to even try getting something else out of it.


To focus on the main topic, I will cut off all other aspects of implementing a good HTTP API client, like sending the request, parsing, abstracting over effects and so on.

To keep things simple, let’s make  execute method just ask for an implicit MockResponse typeclass instance, that will supply requested values instead of doing everything mentioned above.

Let’s code!


We’ll start with some simple things, that are required though. First let’s define a model to play with. Just some case classes from task management domain:


What we are actually going to request are lists of domain objects. But get[Projects] looks better than get[List[Project]], so let’s define some type aliases. Also, this is a good place to put an instance of our builder:


Good. We’ll get to the Builder later. Now let’s define

APIResource typeclass

An instance of APIResource[R] is a marker that R can be requested with our builder.


A couple of comments here:

  • In a real case typeclass body will not be empty — it’s a good place to define some specific entity-related properties. A resource identifier, that is used to create a request can be a good example.
  • We mark APIResource as sealed here, because we know all the instances upfront and don’t want library users to create new ones.

Now we’re ready to implement the Builder.

Single resource request

Builder call chain starts with get method — let’s create it:


Quite simple for now: we allow the method to be called only for types that have an implicit APIResource instance in scope. It returns a request definition that can be executed or extended further with and method.

We will solve the execution task first. A RequestDefinition trait defines execute method for all request definitions:


There’s the MockResponse thing I was talking about. It allows us to avoid implementing all the real machinery that is not relevant for the topic of this post.
Almost everything about this typeclass is simple and straightforward, but there’s an interesting thing that will show up later, so I have to put the implementation here to reference it.


Ok, we’re able to execute requests! Let’s see how we can solve the chaining task with and method.

Chaining: 2 resources request

First, we have to decide, value of what type should be returned by an executed multiple resource request. Standard List or Map could do the job in general, but not with our requirements — precious type information will be lost.

This is where HList comes in handy — it’s designed to store several instances of not related types without loss of any information about those types.

Next. When implementing a type-safe API, we have to put all domain constraints into declarations, so that they’re available for the compiler. Let’s write out, what is required to join two resources in one request:

  • both of them must have an APIResource instance in place, and
  • their types must be different

Good, we’re ready to make our first step to a multiple resource request definition:


All our constraints are in place:

  • the context bound ensures that original resource type is valid.
  • AR: APIResource[RR] implicit parameter ensures that newly added resource type is valid.
  • NEQ: RR =:!= R implicit ensures that R and RR are different types.

Also, you can notice the HList in the result type parameter. It ensures the execution result to be precisely what we requested at compile time.

Now we’re all set up to dive into the most interesting and complex problem.

Chaining: multiple resources request

Here the task is to append a new resource request R to a list L of already requested ones. Again, let’s first define required constraints in words:

  1. every element of  L must have an APIResource instance in place;
  2. R must have an APIResource instance too;
  3. L must not contain duplicates. Let’s call such a list “distinct”;
  4. L must not already contain R . In another words, result list must be distinct too.

That’s a lot. Let’s see, what implementation we can come up with here.
We will start from the MultipleRequestDefinition class itself:


Actually, we already have everything for our first requirement. A LiftAll typeclass from shapeless ensures that every element of an HList has a specified typeclass instance.
In our case, implicit allAR: LiftAll[APIResource, L] constraints L to have an APIResource for each element.

Implicit ID: IsDistinctConstraint[L] will ensure that all elements of L are different (requirement #3). There’s no IsDistinctConstraint in shapeless 2.3.0, so we will have to implement it ourselves. We’ll come to that later.

That’s it for the class definition. Let’s move on to the and combinator:


Requirement #2 is trivial here. NotContainsConstraint for requirement #4 will have to be implemented by us too.

All right, so we have two HList constraints to implement. Let’s see how it’s done.

Implementing HList constraint

In general, a constraint is implemented as a typeclass, that provides instances for and only for objects, that meet the constraint.

Most of the time it can be done with a technique similar to mathematical induction. It involves two steps:

  1. Define a base case: implicit constraint instance for HNil or an HList of known length like 1 or 2.
    Base case depends on the nature of the constraint and can involve additional constraints for HList element types.
  2. Define the inductive step: implicit function, that describes how new elements can be added to an arbitrary HList, that already meets the constraint.

We will start with NotContainsConstraint. The typeclass definition is quite straightforward:


U  is the type that L must not contain to meet the constraint.

Let’s define the base case. Here it’s simple:

HNil doesn’t contain anything.

In general we want constraints to stay same under any circumstances, so it’s usual to define implicit rules right in the typeclass companion object:


Seems logical: for any type U we state that HNil doesn’t contain it. Heading over to inductive step, it can be expressed in words this way:

Given an HList that doesn’t contain U, we can add any non-U element to it and get a new HList, that still doesn’t contain U.

Let’s encode it.


Here we require an HList L, that doesn’t contain U (is guaranteed by implicit  ev: L NotContainsConstraint U) and a type T, that is not equal to U( ev2: U =:!= T ). Given those evidences, we can state that L :: T doesn’t contain U. We do it by supplying a new typeclass instance  new NotContainsConstraint[T :: H, U] {} .

Some tests:


Nice, it works!
I hope it’s transparent here how implicit resolution can or can not find a constraint instance for an HList: we start with HNil base case and go to the list head. If implicits chain is not broken along the way by a duplicate element — we get a constraint instance for the whole list.

Now we’re going to implement IsDistinctConstraint in a similar manner. And our fresh NotContainsConstraint is going to help us here!

Base case is quite simple:

HNil is a distinct list.


Inductive step is quite simple too:

If an HList L is distinct and it doesn’t contain type U, than U :: L is a distinct list too.


Tests show that everything works as expected:

Wiring everything up together

Now, when we’ve done all the preparation work, it’s time to get our builder to work.
We’ll try it out in a REPL session. Single request case goes first:


Everything is ok, we’re getting the mocks, defined in MockResponse. But surprise awaits us, if we try to get multiple resources:


There’s no implicit mock response for our HList! We will have to add some implicits into MockResponse companion to help it join our results:


After all those constraint tricks this simple typeclass extension should be transparent to you. We basically supply a MockResponse instance for any combination of MockResponses.

Important note: although the problem we’re solving here looks artificial, it is not — in a real world we will have to propagate requested types through all network & parsing machinery, that obtains the result. It is the only way to keep the compile-time safety.
And, similarly to our example, some tools (probably implicit) will be required for joining several results in an HList.

Finally, we get everything working! Notice the result types and how HList allows to select only requested types.


All safety requirements are also met — there’s no room for programmer errors:

Shapeless 2.3.1

Shapeless 2.3.1 is coming out soon, and it will contain both constraints we implemented here.


Creating a library or an API is a great responsibility. Providing end-user with a type-safe interface, that doesn’t allow to run and deploy malformed definitions is a high priority aspect.

HList constraints are a great tool in a Scala API developer’s toolbox. In this post we’ve seen them in action, applied to a practical example.

Thanks for reading! See you in future posts 🙂

Custom data validation rules with Shapeless tags

Incoming data validation is a problem that every API developer faces at some point in time. In this small article I’ll show, how shapeless tags can be used to express custom validation rules for Play JSON deserializator.

Full sbt project for this article can be found here.


I assume the reader is familiar with basic Play JSON converters and combinators. Though it will help, it’s not a necessary knowledge to get the idea.

Some Play JSON basics can be learned here.

The problem

Let’s say our API accepts credit card payments. We define a simple data model for those cards (expiration date is omitted for brevity):


We’re going to receive this as a JSON field of incoming request and have to validate the credentials against some rules:

  • card number is 16 to 19 digits after removing all whitespace chars (that’s why we made it a String, not a Long)
  • cvv is 3 to 5 digits

How can we implement this? An experienced API designer would shout: “Those are not strings!”, and introduce some types. Completely valid point, a model like this:


would do the job. Single drawback is you’d have to implement custom serialization/deserialization for CVV and CardNumber to stay with simple JSON strings. By default, an instance of this type would serialize like:


Anyway, this is still good design. But what if we want them to be  String’s? For any reason, like we’d have much cleaner code that uses this CreditCard class.
Let’s set this as a requirement and see what we can do.

Simple strings will require defining custom  Reads for all types they belong to. Like if we have card number in some other type T, we’d have to duplicate that rule in Reads[T]. We don’t want that.

Here is where tags come nicely into play.


As a quick intro, a tag is a marker for an existing type that creates a new type with following properties:

  1. Values of the new type can be used as the values of original untagged type.
  2. Values of the original type can’t be treated as tagged ones. Such code won’t compile.

In this article I will use shapeless tags. A simple usage example:


Tags implementation is quite concise, you can look through it in the shapeless repo.

Defining a rule for tagged string

Returning to initial problem, here is how we can use tags to define custom validation rules for those credentials.

First, let’s tag our model fields:


Now we can define rules for tagged types:


Notice, that as we define Reads for a tagged type, we must return the same tagged type. So here is where we tag values.

Doing so allows us to use default Play macros to define Format for CreditCard (and any other type we’d like to put those “custom” strings in):


That’s it. Let’s test:


So it works! 🙂
We left our strings almost untouched while not losing Play Reads  granularity.

Thanks for reading!

UPDATE. Note on Play route url binders.

A nice catch from Doug Clinton in comments: this trick won’t work with Play route parameter.
I’ll quote Doug:

The problem is that the generated routes file uses classOf[T]  when creating its invoker. classOf  expects a class type, and won’t compile when the parameter type is @@[String, IdTag] , which does not have a runtime class.

Thank you for addition, Doug!


I want to say a big “thank you” to Denis Mikhaylov (aka @notxcain) for introducing this concept to me.

How to make an idiomatic Javascript library with Scala.js

UPDATE: Article was updated to Scala.js version 0.6.7, which vastly simplifies Promises related section.

Scala.js opens a big world of frontend development to Scala fans. Most of the time Scala.js project ends up being an independent browser or Node.js application. But there are cases, where you would want to make a library for general frontend developers.

There’re some interesting gotchas in writing Scala.js library such way, that it will be natural to use for an average JS developer. In this article we will develop a simple Scala.js library (code) to work with Github API and will focus on the idiomaticity of it’s JS API.

But first, I’m sure you want to ask

Why would I do that?

Reasonable one.
You should consider developing such a library if:

  1. A client application for your Scala API backend already exists, and it’s native Javascript.

    Sad, you will hardly have a chance to write it from scratch with Scala.js, but at least it makes sense to write a communication / interpretation library for those guys.
    It will simplify interaction between you and frontenders in two ways:

    • You can hide some tricky client-side logic there, and expose much simpler API.
    • Your library can work on model classes, defined in backend project (see Cross-Building). You get typesafe isomorphic code almost for free and can forget about client-server protocol synchronization problems.
  2. You develop a public API for developers, like Facebook’s Parse.

    A perfect solution for a Javascript API SDK. See all the advantages of the previous case.

Recently, I’ve faced the first case. Moreover, our REST-like JSON API has two different browser based clients. So developing an isomorphic library was a logical choice.

Let’s start with our library.


  1. As Scala developers we want to write all business logic in familiar functional style, being able to use all the handy Scala features.
  2. Library API must be natural for JS developers.

Setting up project

Such a project doesn’t differ from a regular Scala.js app. If you are new to Scala.js, you can read this tutorial first.

Folder structure:

resources/index-fastopt.html — a page that will just load our library and  resources/demo.js file, that will test the API.


The purpose of the library is to simplify Github API interaction. For simplicity, we’ll implement only one feature – loading users and their repos by login.

So it’s, basically, a public method and a pair of model classes, that store results (value objects). Model is the place we’ll start writing code.


Let’s define model classes like this:

Everything is easy: User has some repos, a repo is either an origin or a fork. Good old Scala model. How do we export that to JS developers?

For a full reference of exporting features see Export Scala.js APIs to Javascript

Object creation API

Let’s look at, how we should expose such API. It seems an easy solution to expose the constructor:

But this won’t work. You don’t have Option constructor exported, so there’s no way to create homepage  parameter.

Moreover, there are additional limitation for case classes: You can’t export two case constructors that are under inheritance relationship. This code won’t even compile:

So what is the best choice? I found that it’s best to leave constructors alone and just expose JS-friendly factory methods, like this:

Here with the help of js.UndefOr we handle optional parameter JS way: you can pass a String , or don’t pass anything:

Note on caching Scala objects

Making client call  Github() every time is not the best API option. If you don’t need laziness, you can cache it upon startup:

Reading model properties

Seamless types

If we now try to read fork’s name, we’ll get undefined . Fair enough, it’s not exported. Let’s export model properties.

There’re no problem with native types like String , Boolean and Int . They can be exported as is:

A case class field can be exported with @(JSExport@field) annotation. An example for  forks property, that’s not a member of Repo trait:


But as you already can expect, there’s a problem with
homepage: Option[String] . Well, we can export it, but this would be useless – to get the actual string value JS developer would have to call something on an option, and nothing is exported.

On the other side, we’d like to keep Option in place, so that our Scala code, that manipulates value classes, remains powerful and simple.  js.UndefOr[T] API is way less expressive.

A solution here is to export a special JS-friendly getter method:

Let’s try it out, it works:

We retained our beloved Option monad, and exported nice and clean JS API. Great!


User.repos is a List , and has the same problems with being exported. Solution here is the same too: we’ll just export it as a plain JS Array :

Now we can even map them 🙂 :

Sum types

There’s still one problem with  Repo trait. As we’re not exporting constructors, given a  Repo instance, JS developer can’t figure out, what kind of  Repo it is.

In Javascript there’s no pattern matching and using inheritance is not so popular, sometimes even questionable. So we have several options here.

  1. Depending on the context, provide methods like isFork: Boolean or  hasForks: Boolean at the base level. This is perfectly fine, but not general enough.
  2. Add  type: String (or whatever name feels suitable to you) property to all sum types.

I choose the second one, because it can be abstracted and used throughout the whole codebase. Here’s how it can be done. Let’s declare a mixin that exports a type property:

We have to use a different name for scala definition, because it’s a reserved word.
That’s it! We can now mix it in:

… and use it:

To make this a little safer, we can store type names constants, that can be compared with instance type property. This can be done typesafe:

Having this helper class we can define these constants in our Github global for example:

Now we can avoid strings in Javascript! An example:

That’s how we dealt with sum types.

What if I can’t change object, that I want to export?

This is a case if you want to (maybe, partially) export your cross-built model classes or other imported library objects. The solution is the same to Option and List with the only difference: you have to implement JS-friendly replacement classes and conversion yourself.

An important rule here is to use JS replacements only for export ( Scala => JS) and instance creation ( JS => Scala ). All business logic must be implemented with pure Scala classes.

Let’s say you have a Commit class, that you can’t change:

Here what you can do to export it:

Then, for example, a Branch  class, that you own, would look like this:

Since in JS environment commits are represented with CommitJS objects, a factory method for Branch  would be:

Of-course, this workaround is not a beautiful thing, but at least it’s type checked. That’s why I think it’s preferable to view your library not only as a value-classes proxy, but as a facade that hides redundant details and simplifies the API. That way you won’t even need to export the underlying model.

That’s all for exporting model. Let’s move on to the more interesting part – loading the content from Github API.



For the brevity purposes we will use scalajs-dom Ajax extension as a “network” layer. Let’s for some time forget about how we’re going to export things, let’s just implement the API.

For the simplicity, we’ll put everything AJAX-related into API object. It will have two public methods: for loading user and loading repos.

We will also implement a DTO layer, to decouple API from the model. For type-safe error handling we’ll use Xor type from Cats library. The result type of the method call will be Future[String Xor DTO], where DTO is the type of requested data and String will represent error.

I’ve mentioned everything for this listing to be more understandable, here it is:

Deserialization code is hidden, it’s not interesting. The load method returns string error, if response code is not 200, otherwise it converts the response data to JSON and then to DTO’s.

Now we can convert our API results into model classes.

Here we use a monad transformer to combine these “disjunctioned” futures, and then convert DTO’s into model classes.

Well, that is quite idiomatic functional Scala, lots of pleasure. Now let’s think about how we will export  loadUser method to library users.

Share the Future

To follow the article goals we need to answer the question: what is the idiomatic way to handle asynchronous call in Javascript? I already hear experienced frontenders laughing, because there are no such thing. Callbacks, event emitters, promises, fibers, generators, async/await — all of them are somehow valid approaches. So what should we choose?

I think, the closest thing to Scala Future in Javascript are Promises. Promises are very popular and are already native in most modern browsers. So we’ll stick with them.

First, we must let our Scala code know about those promises. Until Scalajs 0.6.7 we would have to use Promise typed facade from scalajs-dom. But with Scalajs 0.6.7 things became much easier, we will just use the “standard” Promises.

All we have to do now is to convert a Future into Promise. Again, since version 0.6.7 this is not more a problem — there’s a toJSPromise converter in JSConverters . We will just need to help it with the left side of our Xor — convert it to a failed Future to get a rejected Promise:

So let’s share the promise with our JS friends! As usual, we put it to Github object, near the original method:

Here in case of failed future we’re rejecting promise with the exception message. That’s all, we can test the whole API now:

Well, we did it! We can use Futures and everything else we are got used to — and still export idiomatic JS API.

For more API usage examples see full demo.js. To play more with the project, just fetch the repo, then build and run it.


Putting it all together, here are some general advice on writing a Javascript library with Scala.js:

  • Cache exported objects on startup.
  • Export seamless types “as is”.
  • Don’t export Options, Lists and other Scala standard. Put a JS-friendly getter nearby, that converts to  js.UndefOr and js.Array. BTW, same with  Map => js.Dictionary.
  • Don’t export constructors. Use a JS-friendly factory method. JS-friendly means it accepts js.* types and converts it to Scala standard types.
  • Mixin a string type property into sum types.
  • Export Future s as js.Promise s
  • Scala first. You are a Scala developer, so don’t limit yourself in any way: use all the power you like. You know now, that you’ll be able to export it.


Solid Type System vs Runtime Checks & Unit Tests (Scala example)

A month ago I made a talk for my colleagues at our private developers meetup called QIWI Conf. It’s about how Scala type system can help you to release safer code.

Talk is mostly for Scala beginners, yet some patterns covered are very powerful and not offered by majority of other languages.

Contents in brief

  • Introduction
    • What are fail-fast options today?
    • Why compile-time checking is the best choice?
  • Patterns that can lift up your code to be checked at compile time
    • Options
    • Either and scalaz.\/
    • Sealed ADT + pattern matching
    • Tagging
    • Phantom types
    • Path dependent types
Slides (english)
Video (russian):

FPConf Notes

Notable talks from the FPConf conference.

Macros in Scala

A good introduction to scala macros from JetBrains scala plugin developer.

Macros is basically an AST transformation.

Simple macros are implementes as a method invocations:

More complex things are achieved with implicit macros.

Macros can help with:

  • Typeclass generation (including generic implicit macros as a fallback)
  • DSLs
  • Typechecked strings (e.g. format-string)
  • Compiler plugins

IDE support is far from ideal, mostly because of todays scala macro implementation limitations. Coming  Scala.meta to the rescue.

Speaker had a custom IDEA build on his laptop, with a super-awesome “expand macros” feature. Just press a magic shortcut and  examine expanded macro code.
Brilliant thing by JetBrains, can’t wait to use it.

Embedding a language in string interpolator

Great example of custom string interpolator use case.

Mikhail Limansky showed an example of creating MongoDB query language interpreter with a string interpolator.

Most of the times you don’t need such things, for example if you use a good ORM. But speaker’s case is valid. He works with two projects each using different and verbose ORM’s to access a single mongo database. So he made a decision to implement a single expressive (and importantly— well-known) language interpreter, that would generate ORM code of choice.

Even though it’s a plain string interpolator, it’s almost compeletely typesafe, thanks to extensive macros usage. Talk covers every step to implement such a thing.

Here’s the library.

Frontend with joy (DataScript database)


Introduction into DataScript database:

In a few words — it’s an immutable in-memory database for browser js applications.
Along with mentioned immutability it has several other advantages:

  • Powerful query language with pattern matching.
  • Well-defined and simple database changes description format. With an ability to query with such diffs it has event sourcing out of the box.
  • It builds indexes.

Scala performance for those who doubt

Slides (from jpoint)

My personal favourite of all the fpconf talks.

You can’t measure performance by hand on JVM, because of
  • Deadcode elimination
  • Constant folding
  • Loop unrolling
Tool to use: JMH
Micro-benchmark runner. Has sbt plugin (sbt-jmh)
To find root of some performance problems it’s useful to look at bytecode and even assembler code.
Tools here: JMH perfasm profiler, javap


1) Pattern matching
Simple ADT match equals in speed to if-clause sequence.
Null-check is much faster then Option-matching. In general, because of type-erasure, parametrized types pattern matching is slower (but it’s a fair price such feature).

2) Tail recursion
Basically is as fast as loops

3) Collections
Fold and map combinators have significant overhead for big arrays, and especially for primitives as elements.
That’s partly because of HotSpot optimization heuristics are shaped for java.

The problem for primitives is boxing. Scala collections are generics, so specialization doesn’t work for them.

There is an alternative collection library, called Debox, which have specialized Buffer, Set  and Map .


  • Scala is slow:
    • it’s easy to write beautiful, but laggy code
    • collections are super-slow with primitives
    • scalac can generate strange code
  • Scala is fast:
    • with good internals knowledge, beautiful code can work as fast as java code
    • with a few hacks you can make collections be friends with primitives
    • JVM can optimize strange scalac generated code