As it already has been mentioned - Future's poor performance is mostly due redundant context switches on pretty much any operation. If you do `future.map((i: Int) => i + 1)` it can be executed on a different thread, which doesn't make much sense. It was the case in 2.11 stdlib and they're doing lots of improvements in this area, but Future still lags behind all widely-used "IO monads".
But put aside performance, any day of the week I'd choose ZIO (or Cats Effect IO or Monix Task) over Future even if ZIO was 10x slower. Amount of safety and composability "IO monads" give is outstanding.
There's a blooming ecosystem of pure functional programming libraries. Or even two ecosystems: Typelevel and ZIO.
My company uses Scala a lot with with these ecosystems (my team is on Typelevel, another one is on ZIO) and in fact Spark and Akka is what we're trying to avoid as much as possible. Once I got used to the composable nature of FP and static types - both Spark and Akka just feel clumsy and unintuitive, like maintainers of Spark chose Scala because it was a "Java with fancy syntax" back then, ignored most of benefits of static types, HKT, immutability, type classes etc. If there's future behind Scala - it definitely doesn't lie in Apache land. Akka is slightly different and probably has higher potential, but just an overkill for most of use cases it's advertised for.
But put aside performance, any day of the week I'd choose ZIO (or Cats Effect IO or Monix Task) over Future even if ZIO was 10x slower. Amount of safety and composability "IO monads" give is outstanding.