model-writtenrun 8507 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Robins Tharakan reported that window aggregates with ROWS offset FOLLOWING frames fail with "unexpected end of tuplestore" after spilling at low work_mem, reproducing on master and all supported branches #1.
- Henson Choi traced the failure to an active EOF read pointer following subsequent writes while WindowAgg's seekpos stayed stale, and submitted a resynchronization fix with a regression test; he also proposed a position-management layer and longer-term storage performance improvements #2.
- Tom Lane agreed with the diagnosis but doubted that localized resynchronization covered every active-pointer combination; he advocated absolute-position operations directly in tuplestore, eliminating WindowAgg's duplicate position tracking without an intermediate management layer #3.
- Henson Choi proposed a smaller alternative that prevents the read pointer from reaching EOF by limiting advancement to the last stored row, retained the regression query, and reported successful assertion-enabled checks on master and supported branches while requesting consideration of this immediate fix #4.
- Henson Choi then followed the API approach with a four-patch series introducing absolute row tracking, tell/seek operations, and removal of WindowAgg's seekpos; absolute numbering survives trimming, disk seeks still take time proportional to distance, and the separate tuple-callback cleanup was explicitly optional #5.
- At the thread's end, the new series was proposed for master, with no linked commit or subsequent acceptance shown; Henson Choi requested guidance on whether stable branches should receive that series or the smaller preventive fix #5.