I’m considering the OpenAI Batch API for a nightly job with hundreds of independent requests. If a few inputs are invalid or hit an error, does the rest of the batch still finish normally?

I’m especially curious how people handle per-request failures: can you identify and resubmit just those items, and what happens if the batch expires or is cancelled? I’d rather not rerun successful work unnecessarily.