model-writtenrun 9536 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Magnus Hagander opened the thread after Michael Banck questioned the documented superuser requirement, observing that pg_rewind's file-access functions could receive individual EXECUTE grants #1.
- Michael Paquier supplied a documentation patch showing the four grants needed for a non-superuser role and explaining the CHECKPOINT needed on a recently promoted source to refresh its control-file timeline information #2.
- Magnus Hagander requested listing the restricted role before the superuser alternative and supported including the checkpoint guidance; Michael Paquier accepted the wording change and clarified that the permissions documentation would reach PostgreSQL 11 while the checkpoint guidance would reach 9.5 #3 #4 #5.
- Michael Paquier committed the documentation from #2 with that light wording change as bfc80683ce, then confirmed the version-specific backpatches #6.
- Michael Banck questioned why promotion did not itself ensure a completed checkpoint and requested clearer source-server wording; Michael Paquier explained that promotion requests an asynchronous checkpoint which might not even have begun when pg_rewind runs #7 #8 #12 #14.
- Magnus Hagander proposed having pg_rewind issue the checkpoint automatically, but Michael Paquier identified unnecessary I/O and the renewed superuser requirement; discussion then shifted toward detecting and waiting for the source timeline to update, with a timeout and revised same-timeline behavior #9 #13 #15 #16.
- Kyotaro Horiguchi described an unresolved corruption case involving diverged histories on the same timeline, illustrating a further limitation of treating matching timeline IDs as proof that rewind is unnecessary #17.
- Peter Eisentraut requested tests proving non-superuser operation, and Michael Paquier supplied a separate patch that creates a minimally privileged rewind_user and uses it in the existing online-source tests; Paquier subsequently moved discussion of that patch to a new thread #18 #19 #20.
- The thread ended with the documentation committed and automatic checkpoint waiting still under design: Magnus Hagander favored waiting by default with an immediate-error option, while Michael Paquier proposed a timeout that errors if timelines still match and suggested timeout zero for immediate failure #21 #22.