Concepts / Building Search UI / Improve performance
Apr. 17, 2019

Improve Performance

Mitigate the impact of slow network in your search application.


“Debouncing” is the process of ignoring too frequent events, keeping only the last one in a series of adjacent events.

The Debouncer class provides a generic way of debouncing calls. It can be useful to avoid triggering too many search requests, for example when a UI widget is continuously firing updates (e.g. a slider).


Throttling works in a similar fashion as debouncing, except it ensures a constant throughput. When throttling, calls are delayed no more than a given amount of time before being fired. In other words, at regular intervals, the throttler will fire the latest call. This way, no matter how many calls are made, exactly one per interval is fired.

Throttling is achieved via the Throttler class.

Request strategy

By default, a Searcher will launch a request every time you call search(...). That’s what you want, right?

Well, not always. When network conditions are bad, for example (high latency, poor bandwidth, packet loss…), the network may not be able to cope with as-you-type search, which could lead to a poor user experience.

That’s why the Searcher class accepts an optional strategy delegate that, when provided, will take care of deciding how to perform searches. This delegate can decide to drop requests, to throttle them, or even alter their metadata.

Adaptive network

The library provides one request strategy implementation: AdaptiveNetworkStrategy. This strategy monitors the response time of every request, and based on the observed response times, switches between various modes: realtime, throttled and manual.

In realtime mode, which is the default, all requests are fired immediately. This is typically what you want in an as-you-type search context, and provides the optimal user experience when the network conditions are good.

As soon as the network starts to degrade, however, things get more complicated: not only may the requests take too long to complete, but if the bandwidth is not sufficient, the response time may even get slower and slower, as requests stack up inside the pipeline. To fight this effect, the throttled mode delays requests, dropping them along the way, to ensure a maximum throughput. Furthermore, the throttling delay is dynamically adjusted so that the search throughput more or less matches the current network’s capabilities.

If the network is very bad, though, as-you-type search stops being the right option altogether. Instead of having users stare at a spinning wheel forever, it’s better to disable as-you-type, and inform the users that they need to explicitly submit their searches. That’s what the manual mode does: all non-final searches are dropped.

Using the adaptive network strategy first requires monitoring response times using a ResponseTimeStats. Then, you create a new AdaptiveNetworkStrategy using those statistics, and assigning it to a Searcher:

let searcher = /* your searcher */
let stats = ResponseTimeStats(searcher: searcher)
let strategy = AdaptiveNetworkStrategy(stats: stats)
searcher.strategy = strategy

Note: As is customary for delegates, a Searcher does not retain its strategy. You must therefore ensure that its lifetime exceeds that of the searcher.

Writing your own strategy

Implementing your own strategy is just a matter of implementing the RequestStrategy protocol, which contains only one method: performSearch(from:userInfo:with:). The Searcher calls the strategy when the search(...) method is invoked, providing the search metadata via the userInfo parameter, and a block that the strategy should call when it decides to perform the search.

Optimize build size

In some cases, you might not need to use our widget system, and just want to use the core parts of InstantSearch. In that case, you can decide to download only specific parts of the library.


pod 'InstantSearch', '~> 2.0'
# pod 'InstantSearch/Widgets' for access to everything
# pod 'InstantSearch/Core' for access to everything except the UI widgets
# pod 'InstantSearch/Client' for access only to the API Client


github "algolia/instantsearch-ios" ~> 2.0 # for access to everything
# github "algolia/instantsearch-core-swift" ~> 3.0 # for access to everything except the UI widgets
# github "algolia/algoliasearch-client-swift" ~> 5.0 # for access only to the API Client


You can easily cache the results of the search queries by enabling the search cache. The results will be cached for a defined amount of time (default: 2 min). There is no pre-caching mechanism but you can simulate it by making a preemptive search query.

By default, the cache is disabled.

// Enable the search cache with default settings.
InstantSearch.shared.searchCacheEnabled = true

// Enable the search cache with a TTL of 5 minutes.
InstantSearch.shared.searchCacheEnabled = true
InstantSearch.shared.searchCacheExpiringTimeInterval = 300

Queries Per Second (QPS)

Search operations are limited by the maximum QPS (the allowed number of queries performed per second) of the plan.

Every time you press a key in InstantSearch using the SearchBox, we count one operation. Then, depending on the widgets you will be adding to your search interface, you may have more operations being counted on each keystroke. For example, if you have a search made out of a SearchBox, a Menu, and a RefinementList, then every keystroke will trigger one operation. But as soon as a user refines the Menu or RefinementList, it will trigger a second operation on each keystroke.

A good rule to keep in mind is that most search interfaces using InstantSearch will trigger one operation per keystroke. Then every refined widget (clicked widget) will add one more operation to the total count.

In case you have issue with the QPS you can consider implement a debounced SearchBox.

Did you find this page helpful?