TaskExecutor has complex logic to process data, additional output, side inputs and so on. This is because the TaskExecutor manages data flow within a Task. The downside of this approach is that it increases complexity of TaskExecutor if we add other feature of data (e.g., watermark) and it transfers data between Transform by pull-based approach, which may occur additional enqueue/dequeue overhead for each element processing.
We can simplify the TaskExecutor's data processing logic and implement push-based element processing by delegating the data flow in a Task to OutputCollector. We can implement various OutputCollectors such as push-based output collector that emits data to downstream operator directly and sink output collector that writes data to Writer.