Let’s say you have a text field, and whenever the user types something into it, you want to make a network request which searches for that query.
Please note that the following examples use Cocoa extensions in ReactiveCocoa for illustration.
The first step is to observe edits to the text field, using a RAC extension to
UITextField
specifically for this purpose:
let searchStrings = textField.reactive.continuousTextValues
This gives us a Signal which sends values of type String?
.
With each string, we want to execute a network request. ReactiveSwift offers an
URLSession
extension for doing exactly that:
let searchResults = searchStrings
.flatMap(.latest) { (query: String?) -> SignalProducer<(Data, URLResponse), AnyError> in
let request = self.makeSearchRequest(escapedQuery: query)
return URLSession.shared.reactive.data(with: request)
}
.map { (data, response) -> [SearchResult] in
let string = String(data: data, encoding: .utf8)!
return self.searchResults(fromJSONString: string)
}
.observe(on: UIScheduler())
This has transformed our producer of String
s into a producer of Array
s
containing the search results, which will be forwarded on the main thread
(using the UIScheduler
).
Additionally, flatMap(.latest)
here ensures that only one search—the
latest—is allowed to be running. If the user types another character while the
network request is still in flight, it will be cancelled before starting a new
one. Just think of how much code that would take to do by hand!
Since the source of search strings is a Signal
which has a hot signal semantic,
the transformations we applied are automatically evaluated whenever new values are
emitted from searchStrings
.
Therefore, we can simply observe the signal using Signal.observe(_:)
:
searchResults.observe { event in
switch event {
case let .value(results):
print("Search results: \(results)")
case let .failed(error):
print("Search error: \(error)")
case .completed, .interrupted:
break
}
}
Here, we watch for the Value
event, which contains our results, and
just log them to the console. This could easily do something else instead, like
update a table view or a label on screen.
In this example so far, any network error will generate a Failed
event, which will terminate the event stream. Unfortunately, this
means that future queries won’t even be attempted.
To remedy this, we need to decide what to do with failures that occur. The quickest solution would be to log them, then ignore them:
.flatMap(.latest) { (query: String) -> SignalProducer<(Data, URLResponse), AnyError> in
let request = self.makeSearchRequest(escapedQuery: query)
return URLSession.shared.reactive
.data(with: request)
.flatMapError { error in
print("Network error occurred: \(error)")
return SignalProducer.empty
}
}
By replacing failures with the empty
event stream, we’re able to effectively
ignore them.
However, it’s probably more appropriate to retry at least a couple of times
before giving up. Conveniently, there’s a retry
operator to do exactly that!
Our improved searchResults
producer might look like this:
let searchResults = searchStrings
.flatMap(.latest) { (query: String) -> SignalProducer<(Data, URLResponse), AnyError> in
let request = self.makeSearchRequest(escapedQuery: query)
return URLSession.shared.reactive
.data(with: request)
.retry(upTo: 2)
.flatMapError { error in
print("Network error occurred: \(error)")
return SignalProducer.empty
}
}
.map { (data, response) -> [SearchResult] in
let string = String(data: data, encoding: .utf8)!
return self.searchResults(fromJSONString: string)
}
.observe(on: UIScheduler())
Now, let’s say you only want to actually perform the search periodically, to minimize traffic.
ReactiveCocoa has a declarative throttle
operator that we can apply to our
search strings:
let searchStrings = textField.reactive.continuousTextValues
.throttle(0.5, on: QueueScheduler.main)
This prevents values from being sent less than 0.5 seconds apart.
To do this manually would require significant state, and end up much harder to read! With ReactiveCocoa, we can use just one operator to incorporate time into our event stream.
Due to its nature, a stream's stack trace might have dozens of frames, which, more often than not, can make debugging a very frustrating activity. A naive way of debugging, is by injecting side effects into the stream, like so:
let searchString = textField.reactive.continuousTextValues
.throttle(0.5, on: QueueScheduler.main)
.on(event: { print ($0) }) // the side effect
This will print the stream's events, while preserving the original stream behaviour. Both SignalProducer
and Signal
provide the logEvents
operator, that will do this automatically for you:
let searchString = textField.reactive.continuousTextValues
.throttle(0.5, on: QueueScheduler.main)
.logEvents()
For more information and advance usage, check the Debugging Techniques document.