I've worked on a couple systems that achieve exactly-once processing and one took a at-most-once writing approach that backfilled missing events after the fact

I've worked on a couple systems that achieve exactly-once processing and one took a at-most-once writing approach that backfilled missing events after the fact (6 nines of completeness was sufficient for ~hours), and another that took the approach of file-chunk ownership with retries by another worker if the destination files (distributed file system in another region) were too short once ownership expired, before chunks were marked fully complete and the next chunk was added to the work queue (transactionally).