A Cloud Dataflow job using windowing and transformations on Pub/Sub streaming data fails on streaming insert. What is the most likely cause?

Cloud Dataflow streaming pipelines processing unbounded Pub/Sub data require a windowing strategy to group elements. When a pipeline uses transformations that aggregate data (like GroupByKey or Combine), it must define how these aggregations occur over time. If a non-global (sharded) windowing function is not explicitly applied to unbounded data before such transformations, Dataflow cannot determine how to group the elements, leading to a pipeline creation failure. This is because unbounded data without windows would result in an ever-growing, unmanageable group. Incorrect options: No event timestamps were assigned: While event timestamps are crucial for accurate windowing, their absence typically leads to incorrect windowing results or processing time windows, not a pipeline creation failure for streaming insert itself. Triggers were not configured to handle late data: Triggers define when window results are emitted. A lack of triggers for late data affects result completeness, not pipeline creation. A global windowing function was not applied: A global window processes all data as a single, unbounded window. While it's a valid windowing strategy, it's often not suitable for streaming aggregations that need to emit results periodically. The problem states "non-global (sharded)" which implies a need for bounded windows.