I am using Flink SQL to define a streaming job that reads from an unbounded source (google PubSub) and uses the HBase sink to write to Google BigTable. The query is windowing the stream and applying some aggregation function on those windows before is written to BigTable.
The job ends up consisting of 2 subtasks. The first one is the source + some map functions + local window aggregate. The second one applies the global aggregate + a map + the sink function.
We will like to be able to define different parallelism for each subtask, under the assumption that the second one could benefit from more parallelism than what the source subtask needs.