[FEATURE] Large backlog processing when materialized view cold start

[dai-chen]

Is your feature request related to a problem?

When auto refresh on materialized view starts, an initial source file list is generated by FileStreamSource. By default, it just returns files sorted by modified time. In the case of aggregate MV, if the timestamp of events in these file are not strictly aligned with file modified time, disordered (late coming) data may be dropped due to watermark bumped.

What solution would you like?

Provide source option for user to enable MV cold start processing backlog sorted by event timestamp in data rather than file modified time in file metadata.

What alternatives have you considered?

Guide user to give initial starting point by WHERE clause and meanwhile configure large watermark to avoid disorder data dropped. But this may consume more memory because of aggregate window open waiting for possible late data for long time.