Skip to main content

Control over evaluation

Using Eval you can control evaluation of a value or a computation that produces a value. This is useful to delay potentially costly computations, and to prevent start overflows by carefully choosing when computations should take place.

You construct Eval instance by providing a function that computes a value, together with an evaluation strategy. There are three basic strategies:

  • evaluates the function immediately.
  • Eval.later waits until the first time the value is requested. Once computed, the result is saved, so subsequent calls return immediately.
  • Eval.always evaluates the function every time we need its value. If you ask for the value more than once, the function is executed again.

We say that Now is an eager evaluation strategy, and Later and Always are lazy evaluation strategies.

One of the main use cases for Eval is stack safety, that is, preventing stack overflows for operations with deep recursion. For example, here is a (overly complicated) way to compute whether a number is even or odd, by jumping between even and odd until we reach 0.

This approach would lead to stack overflow for big numbers, but we can prevent this using Eval. Using Eval.always { n == 0} we indicate that we want the evaluation to be performed when we need the answer; using later or always does not make a big difference here, since we only evaluate once per n in any case. We indicate the next operation by using flatMap.

import arrow.core.Eval

fun even(n: Int): Eval<Boolean> =
Eval.always { n == 0 }.flatMap {
if(it == true)
else odd(n - 1)

fun odd(n: Int): Eval<Boolean> =
Eval.always { n == 0 }.flatMap {
if(it == true)
else even(n - 1)

fun main() {

One difference between Eval and DeepRecursiveFunction is that with Eval we can call functions directly instead of using callRecursive. However, the latter is more performant in general, so we advise using it unless you require the additional control provided by Eval.

Use the provided API

You should not use when with Eval instances. Rather, use map and flatMap to chain computations and value to get the result when needed.

You should not create Eval instances that call value on other Eval instances. This defeats the barriers in place against stack overflows.