invokeOnCompletion

is mostly designed for internal machinery. Franky, I think it should be somehow “hidden” from API to discourage its use from a general-purpose application code, but I don’t know what would be best way to do it. You should define

then

to get this “chained” style of programming like this:

fun <T, U> Deferred<T>.then(next: suspend (T) -> U): Deferred<U> = async { next(await()) }

We might include it in the future updates of

kotlinx.coroutines

for convenience.

Just curious, but how much overhead would there be in chaining builders like this, @elizarov?

Compared to what?

Compared to directly running

next

within the coroutine preceding it?

A few more objects. I think < 100 bytes in total