For workloads that intersperse Bigslice computations with other long-running computations, there could be significant cost savings by shrinking or eliminating worker machines in between the Bigslice computations.
It's possible to workaround this today by spawning different Bigslice computations, but that is clunky. You have to deal with spawning processes and possibly serializing results.
This may be made much simpler depending on an implementation of #99. If use of a backing store were a prerequisite for shrinking, we could shut down machines with impunity.
For workloads that intersperse Bigslice computations with other long-running computations, there could be significant cost savings by shrinking or eliminating worker machines in between the Bigslice computations.
It's possible to workaround this today by spawning different Bigslice computations, but that is clunky. You have to deal with spawning processes and possibly serializing results.
This may be made much simpler depending on an implementation of #99. If use of a backing store were a prerequisite for shrinking, we could shut down machines with impunity.