Lazy evaluation: good or bad?

[raimohanska]

The fact that bacon.js evaluates event values lazily is a bit confusing. This bit me in my Oredev presentation. So if I couldn't take that into account, how many new Bacon.js users will?

Now, if you do

    fieldValue = $(".foo").asEventStream('keyup').map((e) => $(e.currentTarget).val()).toProperty()
    firstClick = clicks.take(1)
    secondClick = clicks.skip(1).take(2)
    twoValues = Bacon.combineAsArray(
      fieldValue.sampledBy(firstClick),
      fieldValue.sampledBy(secondClick)
    )
    twoValues.log()

What do you get? An array of the values of the .foo textfield at the times of first and second click? Well, not exactly. You'll get the value of the textfield at the time of the second click, twice.

This is because the value of the events isn't evaluated until it's actually used. (see also https://github.com/baconjs/bacon.js/#lazy-evaluation)

Now the question is, should we remove lazy evaluation altogether in Bacon.js 0.8? My suggestion would be to make a spike and see if that brings any serious performance penalty. Then proceed accordingly.

[rpominov]

If an operation always happens eventually, without laziness performance should be better. But if some operations never applied, for example in a.map(heavyWork).throttle(1000), it depends on what user do in heavyWork. So I think it would be tricky to estimate performance penalty here.

[phadej]

@pozadi yet your example could be done with few helpers:

lazy = (f) -> (x) -> () -> f(x)
force = (thunk) -> thunk()

$out = $in.map(lazy(heavywork)).throttle(1000).map(force)

[phadej]

I.e. If I have to choose I'd pick eager evaluation, because as a user I can add lazyness, but cannot remove it.

[lautis]

I'd place my bet on eager being faster, too. When you run the performance benchmarks with CPU profiler, ~35% of time is spent in GC. For every event Bacon wraps the value in either Bacon._.always or Bacon._.cached and both create a new function for every value. But it wouldn't be the first time I'm wrong.

[rpominov]

Yes removing laziness might bring quite significant performance boost, btw!

Also take a look at this perf test https://github.com/pozadi/kefir/blob/master/test/perf/perf-specs/deoptimization.coffee

not so random object
----------------------------------------------------------------
Kefir x 2,894,800 ops/sec ±1.94% (66 runs sampled)
Bacon x 613,077 ops/sec ±1.93% (66 runs sampled)
RxJS x 1,479,617 ops/sec ±1.99% (65 runs sampled)
-----------------------
Kefir 1.00   Bacon 0.21   RxJS 0.51


random object
----------------------------------------------------------------
Kefir x 533,397 ops/sec ±6.09% (60 runs sampled)
Bacon x 290,787 ops/sec ±5.29% (61 runs sampled)
RxJS x 451,112 ops/sec ±4.05% (62 runs sampled)
-----------------------
Kefir 1.00   Bacon 0.55   RxJS 0.85

It shows that Kefir works slower when a lot of objects of different type* passing through it, compared to Bacon/Rx. I think the reason that Bacon/Rx suffers less under that condicions is because in those libs values travel wrapped in closures, don't sure thought. But maybe it worth to consider when removing laziness, I'd make a similar perf test at least after removing.

* I mean object structure e.g., {a: 1} and {a: 2} have same type, and {b: 1} has a different one.

[skozin]

I'd prefer removing lazy evaluation. It can be implemented manually when needed, without penalizing the performance in cases when laziness is of no use, which are far more common in UI applications.

Maybe it would make sense to ship these simple helpers mentioned by @phadej with Bacon. Something like this:

Bacon.lazify = (f) -> (x) -> () -> f(x)
_.apply = (f) -> if 'function' is typeof f then f() else f
Observable::unlazify = -> @map _.apply

$out = $in.map(Bacon.lazify(heavywork)).throttle(1000).unlazify()

They won't work in case of combinators with array and object output (combineAsArray, combineTemplate), but that doesn't seem like a huge issue to me.

[raimohanska]

Thanks guys! Who wants to make a PR?

[algesten]

+1 be gone. it cleans up the bacon code as well.

[AlexGalays]

By the way, are we talking about both kinds of laziness here? ([1] stream not starting until it has subscribers, and [2] lazy evaluation of some mapping functions)

[raimohanska]

1 is a cornerstone and will remain
2 is what we're talking about

[AlexGalays]

Perfect!