How does it handle heterogeneous clusters that might have different versions of code running? Or is it like Cloud Haskell where you have to bring everything dow

How does it handle heterogeneous clusters that might have different versions of code running? Or is it like Cloud Haskell where you have to bring everything down when deploying?