model-writtenrun 2553 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tomas Vondra proposed accounting for fillfactor when estimating rows without statistics after observing that a 10% fillfactor could inflate estimates tenfold and produce oversized hash indexes #1.
- Corey Huinker reported successful tests and accepted the narrow risk of changing fillfactor after loading data, while suggesting a single-division formula; Andres Freund also supported the change #2 #3.
- Peter Eisentraut questioned percentage conversion and rounding, and Tomas ultimately defended two integer divisions to round down usable bytes and then tuples per page #4 #5 #6.
- Tomas committed the original patch unchanged as 29cf61ade3 and confirmed that the two-division formula was retained #1 #7.
- In January 2025, Heikki Linnakangas reported that wide tuples with low fillfactor made density round down to zero, yielding a one-row estimate and sometimes preventing parallel scans; he proposed a minimum density of one because the executor still places a tuple on each page #8.
- Tomas accepted the diagnosis and supplied a follow-up fix using clamp_row_est(), with a regression test covering wide inline attributes and low fillfactor #9 #10.
- Tomas simplified the regression test to one char(1000) column and prepared the commit message, then Heikki approved the final patch before it landed #11 #12.
- The final fix landed unchanged as 9ba7bcc894, with a backpatch to PostgreSQL 17; Tomas confirmed the push, leaving no outstanding patch work in the thread #11 #13.