Introducing the primitive PTransform object – Partition
The GroupByKey transform creates a set of sub-streams based on a dynamic property of the data – the set of keys of a particular window can be modified during the pipeline execution time. New keys can be created and processed at any time. This creates the complexity mentioned in the previous section – we need to store our data in keyed states and flush them on triggers. A question we might have is – would the task be easier if we knew the exact set of keys upfront, during pipeline construction time?
The answer is yes, and that is why we have a PTransform object called Partition.
Important note
A pipeline is generally divided into three phases during its life cycle: pipeline compile time, pipeline construction time, and pipeline execution time. Compile time refers (as usual) to the time we compile the source to bytecode. Construction time is the time when the pipeline's DAG of transformations...
