Object storage gives you durable files, but a directory of files doesn't by itself define a transaction. A table format uses metadata to describe a committed set of files. Readers choose a snapshot, and a writer publishes a new snapshot only after its files are ready. That makes a failed rewrite recoverable without exposing half the new data.
File layout still needs care. Tiny files increase planning and open-request overhead. Very broad partitions force queries to read unrelated data, while a partition for each user may create millions of small groups. Compaction rewrites files into a more useful size; it should preserve row identities and publish through the table's commit protocol. Updates and deletes require matching keys and version rules. A source replay can otherwise restore a deleted record or apply an older value over a newer one.