Write applicatives for concurrent actions

Alex Kelley
9 min readFeb 17, 2019

--

Note: This is Tutorial 21 in the series Make the leap from JavaScript to PureScript . Be sure to read the series introduction where we cover the goals & outline, and the installation, compilation, & running of PureScript. I’ll be publishing a new tutorial approximately once-per-month. So come back often, there’s a lot more to come!

Index | << Introduction < Tutorial 20 | Tutorial 22 > Tutorial 27 >>

In the last tutorial, we continued our exploration of Applicative Functors with a practical example that captured the pattern of nested loops for constructing foldable (e.g., array) comprehensions. We’ll finish off applicatives in this tutorial with another example that pings a mock database, returning a result from a query. As you’ll see, a query in our example is independent of the result of previous queries. Consequently, instead of performing these queries serially, we can leverage concurrency by employing applicatives to run them simultaneously. We’ll look at both the serial and concurrent approaches so that you can see the difference in coding. We’ll mock a “database find” method using our homegrown Task module from Tutorial 13.

I borrowed this series outline, and the JavaScript code samples with permission from the egghead.io course Professor Frisby Introduces Composable Functional JavaScript by Brian Lonsdorf — thank you, Brian! A fundamental assumption is that you’ve watched his video on the topic before tackling the equivalent PureScript abstraction featured in this tutorial. Brian covers the featured concepts exceptionally well, and I feel it’s better that you understand its implementation in the comfort of JavaScript.

You’ll find the text and code examples for this tutorial on Github. If you read something that you feel could be explained better, or a code example that needs refactoring, then please let me know via a comment or send me a pull request. Also, before leaving, please give it a start to help me publicize these tutorials.

Let’s start with a quick review of TaskE then move onto the example.

Reviewing the Task module

Recall from Tutorial 13 that TaskE runs an asynchronous computation that may fail, and it models any side-effects explicitly. There isn't a specific TaskE module in PureScript, so we created one ourselves to mimic data.task from the JavaScript Folktale library. As you'll see, TaskE is a good structure to help mock a database API query, because it runs asynchronously and allows for effects that may fail. For example, when pinging a database endpoint, there's always the possibility that your endpoint might be down. Also, what if the caller invokes an invalid query? Well, in our example, we don't want to continue with the next query because we're building a report header that's constructed from the aggregate queries. In these and other cases TaskE should fail gracefully by reporting that there's a problem back to the caller.

The type alias for TaskE is:

type TaskE x a = ExceptT x Aff a

With ExceptT x, we're creating a structure that can fail during its computation, and Aff a denotes that it is asynchronous and produces side-effects. Note that ExceptT x is a monad transformer for the Either x Monad, whose type we introduced in Tutorial 3.

We haven’t covered monad transformers in this series, and we may not have the opportunity. However, in simple terms, ExceptT helps us to handle a successful or failed computation within the Aff monad. We wrap the computation with a Right a and Left a from the Either monad, denoting success or failure respectively. So, we've got two monads operating together (i.e., Aff and Either) and ExceptT is the glue that makes this possible.

A simple and successful TaskE is created by wrapping the result using pure from Control.Applicative. In Tutorial 13, we used the taskOf method for this purpose. It’s just syntax sugar for pure since we hadn't covered applicatives at the time. The example below wraps the string "good task" in a Task whose computation is later run in main. It is guaranteed to be successful since we're returning a string, "good task", unfettered by any side-effects.

goodTask :: TaskE String String 
goodTask = pure $ "good task"
main :: Effect Unit
main = do void $ launchAff $
goodTask
# fork (\e → Console.error e) (\r → Console.log r)

The launchAff method launches an asynchronous fiber, which in this case is goodTask. The fork method is there to handle the two possible outcomes - either success or failure. Ultimately, it logs "good task" to the console. In case you're wondering how to create a "bad" task - throw an error! You use throwError with the comfort that ExceptT ensures that Either and Aff play nice together.

badTask :: TaskE String String 
badTask = throwError "bad task"
main :: Effect Unit main = do void $ launchAff $
badTask
# fork (\e → Console.error e) (\r → Console.log r)

These, are perhaps, the simplest tasks to write. In the next section, we’ll use our “database find” example to show how to add callbacks within our tasks and clean up the fiber after finishing the asynchronous computation.

Mocking a database query using Task

With our review of TaskE complete, let's move on to our mock example that queries a database API for project records and later logs the titles of those projects. Here, we have a database of projects; each modeled with a record of type, Project = { id :: Int, title :: String } where id is the record identifier and title is the name of the project.

We perform a query by running an asynchronous task with effects that may fail with an exception of type String or succeed by returning a result of type Project. Also, since we're mocking an API endpoint, the Project record response is a time-dependent value. We can model that behavior using setTimeout within our Task to mimic the delay in the response returned by the API.

type Project = { id :: Int, title :: String } dbFind :: Int → TaskE String Project 
dbFind id =
let
query :: Int → Effect (Either String Project)
query id_ = pure $ do
let validIds = [20, 8]
case (elemIndex id_ validIds) of
Just _ → Right { id: id_, title: "Project: " <> (show id)}
Nothing → Left $ "record id: " <> (show id) <> " not found."
in
newTask $ \callback → do
let requestResponse = query id >>= \r →
callback $ either (\e → rej e) (\s → res s) r
_ <- setTimeout 100 requestResponse
pure $ nonCanceler

Within dbFind, a successful query returns a Project record with side-effects modeled explicitly (i.e., Effect (Right Project)). Alternatively, in the case of failure (i.e., Effect (Left String)), we return an error message that the record id was not found in the database.

The newTask method has been designed to perform this computation by accepting (as arguments) a function that transforms an effectful callback to a canceler. We perform the query and then, using the bind 1 operation (>>=), we pass the return value in sequence to our callback. Depending on success or failure and, with the help of the either function, the callback uses res or rej, respectively to wrap the result in our TaskE type signature.

To simulate a time-dependent value, we call setTimeout to delay the callback by 100 milliseconds. We're not interested in the timeout id, so we ignore it by not binding it to a variable. Finally, we tell makeAff (the function underpinning newTask) how to clean up after our asynchronous computation. In this case, there's no elaborate asynchronous process going on, so providing newTask with nonCanceler (a canceler that doesn't do anything) is good enough.

Sequential dbFinds using Monads

There are two approaches to executing multiple queries on a database. The first approach executes them sequentially, as shown below. Take this approach when the formation of a database query is dependent on the information retrieved from one or more previous queries.

In the code example below, the overall goal is to “console log” a report header constructed from our two queries, dbFind 20 and dbFind 8. We use reportHeader to construct a string from the two project titles. To accomplish this, we call dbFind 20 as our first query, which returns its result wrapped in the TaskE. So alas, we know we're working in the TaskE monad, so bind the first query to the lambda function p1 and compose it with reportHeader and our next query dbFind 8, which is bound to p2.

reportHeader :: Project -> Project -> String 
reportHeader p1 p2 =
"Report: " <> p1.title <> " compared to " <> p2.title
main :: Effect Unit
main = void $ launchAff $
dbFind 20 >>= (\p1 → (\p2 → reportHeader p1 p2) <$> dbFind 8)
# fork (\e → Console.error e) (\p → Console.log p)

Recall that bind, whose infix operator is >==, is equivalent to chain in Brian's javascript examples (see example below). The map function, whose infix operator is <$>, lifts reportHeader into the TaskE monad so that it can operate on the TaskE applicative arguments, p1 and p2. Remember that map is a substitute for pure and one apply operation within an applicative expression.

For a comparison, look at the javascript code adapted from Brian’s example:

const reportHeader = (p1, p2) => 
`Report: ${p1.title} compared to ${p2.title}`
Task.of (Db.find (20).chain (p1 =>
Db.find (8).map (p2 =>
reportHeader (p1, p2))))
.fork (console.error, console.log)

Concurrent dbFinds using Applicative Functors

The queries, dbFind 20 and dbFind 8 can be run independent of each other, so it's not necessary to run them sequentially. When we run them concurrently in an applicative expression, it helps to make our code more readable, and perhaps execute faster.

Compared to our sequential example in the previous section, I believe the applicative expression is much more readable. In the example below, if either of these queries fails, then execute (\e → Console.error e) to call out the offending query. Otherwise reportHeader is cleared to construct our header string which is printed out to the console.

main = void $ launchAff $ 
(\p1 p2 → reportHeader p1 p2) <$> (dbFind 20) <*> (dbFind 8)
# fork (\e → Console.error e) (\p → Console.log p)

Naturally, if your mental model feels more at home with code that appears sequential, then feel free to use applicative do syntax:

main = void $ launchAff $ ado 
p1 <- dbFind 20
p2 <- dbFind 8
in
reportHeader p1 p2
# fork (\e → Console.error e) (\p → Console.log p)

Once again, let’s compare this with the Javascript code, adapted from Brian’s video:

const reportHeader = (p1, p2) => 
`Report: ${p1.title} compared to ${p2.title}`
Task.of (p1 => p2 => reportHeader (p1, p2))
.ap (Db.find (20))
.ap (Db.find (8))
.fork (console.error, console.log)

Both the PureScript Javascript code examples are much easier to read, and again should execute faster than a sequential approach using monads.

Lifting functions into applicative expressions

In the previous section, instead of using pure to lift reportHeader into the TaskE monad, we're using map whose infix operator is <$>. That is, we substituted pure (\a b → f a b) <*> m a <*> m b for (\a b → f a b) <$> m a <*> m b.

If you review the past two tutorials, you'll see that's been our practice. But what’s going on with (\a b → f a b) <$> m a <*> m b and how does it work? Well, it helps to step back and look at the type signatures forpure and the infix operators for apply and map:

pure  ::   (a → b) → m (a → b)
(<*>) :: m (a → b) → m a → m b
(<$>) :: (a → b) → m a → m b

Notice that pure lifts our function a → b into m to get m (a → b). Next is apply whose infix operator is (<*>). This method takes m (a → b) and our input m a , and applies it to return our output m b.

Now, look at map, whose infix operator is <$>. It lifts the function (a → b) into m, then applies it to the input m a and returns our output m b.

Somap, when used in an applicative expression, is a substitution for pure and one apply operation.

In conclusion, I believe (\a b → f a b) <$> m a <*> m b is easier to read than pure (\a b → f a b) <*> m a <*> m band you’ll find that it is the idiomatic syntax. So use this pattern whenever you’re not using lift2, lift3, ... or applicative do sugar.

Summary

In this tutorial, we refactored a sequence of independent queries using monads into an applicative expression that runs concurrently. After refactoring, we found that our code is not only better optimized for performance, but it is also simpler to read and understand.

We have come to the end of our tutorials that explored the practical examples using Applicative Functors. I hope you agree that it is a dominant pattern that you can exploit in whatever functional programming language you choose to program in.

In the next tutorial, we’ll move on to explore another useful construct that allows us to traverse a foldable structure (e.g., array) of applicative functors and applies a function to each. For example, if our foldable structure is an array, then the type signature for traverse is (a → f b) → Array a → f (Array b). This method is useful when our function f produces side-effects. For example, imagine the array contains a list of filename paths that require reading and parsing their text.

That’s all for now, and till next time!

1 In case you’re a little hazy on bind or even monads, then please refer to Tutorial 16 for a refresher.

--

--

Alex Kelley

I work on technologies that simplify the creation and enjoyment of web applications.