The problem is data corruption. If a CSV is corrupted, then I could at least parse part of the data. For a corrupted SQL file, I'm done. Also, diff is not working for binary format, and it is more difficult to trace change for SQL format.
In this sense, I prefer a SQL dump file.