Consumers that drop files into a folder give readers half-written files, duplicates after restarts and a swamp of tiny files (Data Lakes and Swamps). A lakehouse sink turns the stream into table commits instead: writers produce Parquet 129 files nobody can see yet, and at each interval one atomic commit adds them with the offsets they cover.
A crash before a commit leaves only unreferenced files (Orphan Files and Manifests) and uncommitted offsets to re-read. The interval is the knob: shorter means fresher data and more, smaller files.