model-writtenrun 3346 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Richard Guo reported an intermittent prepared_xacts failure containing an unexpected gxid row and explained why his semi-join patch appeared unrelated #1.
- Michael Paquier traced gxid to concurrently running ECPG tests and submitted a patch using regress_ transaction names and filtered, ordered pg_prepared_xacts scans; Alexander Lakhin independently identified the same race #2 #3.
- Tom Lane questioned cross-suite parallelism, but agreed that installcheck should tolerate unrelated database activity and endorsed the query filtering while requesting a less generic ECPG transaction name; Michael accepted that addition #4 #5 #7 #8.
- Richard asked whether recovery and subscription TAP tests had the same exposure, and Michael explained that their private clusters made those scans safe #10 #11.
- Michael committed #2 with an additional ECPG transaction-name change as 7e61e4cc7c, leaving other two-phase tests alone after checking their isolation and identifiers; the commit records that the difficult-to-reproduce race did not warrant backpatching #13.
- The broader test-stability discussion remained open: Richard reiterated concerns about parallelism, and Tom suggested removing a separate unstable test whose output depended on plan-cache invalidation #15 #16.
- Jingxian Li introduced a separate initdb patch correcting reversed pointer subtraction in a strncmp length argument, and Richard agreed with the diagnosis #17 #18.
- Daniel Gustafsson found that the proposed initdb fix also needed to include the separating space in the comparison, then questioned retaining undocumented PAM-option support that had been broken throughout all supported releases #19.
- Peter Eisentraut challenged the premise of the parallelism concerns, explaining that make and Meson differed in scheduling details rather than fundamentally different parallelism models #20.
- After Jingxian noted that other authentication methods also needed options, Daniel replaced the arithmetic repair with removal of initdb auth-option support, leaving any broader implementation for future work #21 #22.
- Aleksander Alekseev and Michael endorsed removal, with Michael suggesting a separate option for any future implementation; Daniel committed #22 unchanged as 83ae824c87 and confirmed the push in the final message, completing both patch efforts while leaving the broader test-parallelism questions unresolved #23 #24 #25.