If running the import twice creates two receipts, the script is not done. Uniqueness belongs in the database, not in someone's memory.
Tags: Python, Receiving, Reliability, Data Quality, unlisted
I key receipts on a natural id: tracking number plus vendor plus date, or the vendor's ASN line. The insert is upsert. The second run updates timestamps.
If there is no natural id, I generate one from a hash of the file plus row number and I still store the file hash so the same upload cannot fork reality.
Tests are two runs
My basic test is run, run again, count rows. If the count changes, I failed. This has caught more bugs than clever unit tests on parsers.
Tell the operator it was a repeat
"0 new, 14 already seen" is a success message. Silence makes them run it a third time.
