Watermarks are a way of tracking the progress of time in a streaming pipeline. They are used to determine when a window can be closed and the results emitted. Watermarks can be either event-time based or processing-time based. Event-time watermarks track the progress of time based on the timestamps of the data elements, while processing-time watermarks track the progress of time based on the system clock. Event-time watermarks are more accurate, but they require the data source to provide reliable timestamps. Processing-time watermarks are simpler, but they can be affected by system delays or backlogs.
By using watermarks, you can define the expected data arrival window for each windowing function. You can also specify how to handle late data, which is data that arrives after the watermark has passed. You can either discard late data, or allow late data and update the results as new data arrives. Allowing late data requires you to use triggers to control when the results are emitted.
In this case, using watermarks and allowing late data is the best solution to capture the late data in the appropriate window. Changing the windowing function to session windows or tumbling windows will not solve the problem of late data, as they still rely on watermarks to determine when to close the windows. Expanding the hopping window might reduce the amount of late data, but it will also change the semantics of the windowing function and the results.
Reference:
Streaming pipelines | Cloud Dataflow | Google Cloud
Windowing | Apache Beam

To ensure that the data analytics team does not have access to sensitive columns, you should:
B . Ensure that the data analytics team members do not have the Data Catalog Fine-Grained Reader role for the policy tags. This role allows users to read metadata for data assets that have policy tags applied, which could include sensitive information.
C . Enforce access control in the policy tag taxonomy. By setting access control at the policy tag level, you can restrict access to specific columns within a dataset, ensuring that only authorized users can view sensitive data.

Online prediction
.Optimized to minimize the latency of serving predictions.
.Predictions returned in the response message.
Batch prediction
.Optimized to handle a high volume of instances in a job and to run more complex models.
.Predictions written to output files in a Cloud Storage location that you specify.

Triggers control when the elements for a specific key and window are output. As elements arrive, they are put into one or more windows by a Window transform and its associated WindowFn, and then passed to the associated Trigger to determine if the Windows contents should be output.