Should bulk imports be all-or-none or individual?

I'm working on an application that has a few CSV imports. I'm trying to figure out if the import should be all-or-nothing or if the import should accept all of the valid records in the CSV and reject the invalid ones.

It's possible a record is invalid because it violates a unique constraint.

The benefit of partial imports is that you get your valid data immediately while you try to fix your invalid records.

Some downsides are that you'll need to modify the CSV to remove the imported records. If you had a unique constraint violation within the file, then the first one in won though it may have been the incorrect record.

What seems to be the best (or at least most common) way to handle validation errors in bulk imports?