model-writtenrun 7688 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Kyotaro Horiguchi reported intermittent failures because recovery could reuse an XID whose WAL had never reached disk, and posted a checkpoint workaround that stabilized his test #1.
- Tom Lane identified a race in placing CHECKPOINT after the SELECT whose output advances the test, questioned whether checkpointing concealed a broken guarantee, and proposed either removing the test or making transaction-ID reporting flush WAL #2 #4 #5.
- Investigation tied the failures to WAL buffering: Kyotaro reproduced them with larger WAL blocks, Tom reproduced them without allows_streaming, and Andrey Borodin reported failures with WAL compression; Kyotaro explained how buffer sizes and smaller full-page images prevented incidental WAL writes #7 #9 #11 #12 #14.
- Andrey Borodin suggested testing that transaction effects remained invisible, but Kyotaro argued that this would miss the intended ability to determine transaction outcome without inspecting its data #15 #16.
- Thomas Munro proposed making XIDs durable when clients observe them, replacing didLogXid with an LSN and flushing through an existing or newly generated transaction WAL record #8 #17.
- Tom Lane opposed changing shared_buffers defaults across TAP tests and recommended removing allows_streaming only from this test; Michael Paquier noted possible effects on external tests, and Kyotaro accepted the narrower change #19 #20 #21.
- Kyotaro implemented Thomas Munro’s design in a two-patch series that changed the test configuration and flushed WAL before returning published XIDs; tracking the highest already-flushed XID was omitted because coordinating that optimization appeared too complex #22 #23 #24 #25.
- Nearly two years later, Thomas Munro revived the thread after a tanager buildfarm failure, highlighting the flushing proposal’s expense for small sequential transactions and the unresolved reliability of the documented use case #26.
- The closing discussion investigated a possible separate shared-memory cleanup problem, but Michael Paquier ultimately confirmed that tanager frequently reproduced this thread’s test failure; neither attached proposal had landed by the last message #27 #28 #30 #31 #32.
- Michael Paquier subsequently committed removal of the test and its Meson registration, with backports through PostgreSQL 11, explaining that checkpointing would conceal the unsupported guarantee and mandatory WAL flushing would impose unwanted overhead; neither attached implementation was adopted 8c1cd726c5.