We need a methodology for providing partitioning hints.
A current proposal uses message headers to provide:
- partition_key – the item to partition on
- destination_region – what to target
In the proposal the developer used "partition_key" to route the stream message to the node where data was stored in process. This was done so downstream stream operations could work on the data with out suffering any network IO.
The "destination_region" was used to target the type of data the downstream streams were going to use in their stream processing.