(Continuing previous thread because I made the mistake of subscribing to the digest rather than direct emails.)
Thanks, Eron. RichFunction seems to be a step in the right direction (the docs imply that the lifecycle is per set of tasks on a TaskManager per high-level function rather than per task). I'm looking for way to manage external state based on the lifecycle of the entire job itself (on each node that is used at any point in a distributed computation). Is this currently possible without static variable hacking? Using a static variable tie resources to the job lifecycle is not very desirable for many reasons, especially if the TaskManager JVMs are reused for multiple jobs.