model-writtenrun 7862 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Michael Banck proposed moving -k/--data-checksums into the common options in initdb --help because checksums were sufficiently widely used, attaching the thread's only patchset #1.
- Michael Paquier supported the help change, and Stephen Frost agreed while expanding the discussion to enabling checksums by default #2 #3.
- Michael Banck benchmarked a forced-checkpoint pgbench initialization workload and reported roughly doubled archived WAL and a twentyfold VACUUM slowdown with checksums; WAL compression reduced the extra WAL substantially but left a large runtime penalty, prompting him to suggest enabling compression alongside checksums #4.
- Michael Paquier committed the original help patch unchanged as bc08f7971c, explaining that it stood independently and warning that enabling WAL compression could hurt CPU-bound workloads #5.
- Andres Freund opposed changing the checksum default before reducing its overhead, reporting that checksum computation and PageSetChecksumCopy together halved write throughput in his asynchronous I/O work; Stephen Frost argued that corruption detection justified a safer default despite performance costs #6 #8.
- Magnus Hagander emphasized that disabling checksums was quick whereas enabling them required a slow offline operation, while Andres Freund questioned whether managed-service users could run the necessary local command #11 #13.
- Michael Banck clarified that wal_log_hints produced essentially the same slowdown in his deliberately adverse benchmark, tying much of the observed cost to additional WAL rather than checksum computation alone #18.
- Peter Geoghegan opposed enabling checksums with the existing WAL overhead and distinguished detecting corruption with checksums from preventing crash-related corruption with fsync; Stephen Frost acknowledged the different risks while maintaining that both protections deserved default enablement #21 #27 #35.
- In response to questions about cloud deployments, Michael Banck confirmed that the Azure, Amazon RDS, and Google Cloud SQL instances he checked enabled both data_checksums and full_page_writes; Peter Geoghegan accepted the findings but remained opposed to changing the upstream default #30 #33.
- Amit Kapila also opposed the default change at the observed cost and suggested investigating better WAL compression and reducing PageSetChecksumCopy overhead; Peter Geoghegan later questioned whether the extreme VACUUM slowdown represented real workloads and offered possible help with analysis #34 #37.
- Later discussion weighed protection for minimally tuned installations against additional tuning burdens and upgrade performance surprises, with David Steele pointing out that existing pg_upgrade databases would retain their checksum setting #39 #40 #41 #42.
- The help change remained landed, but enabling checksums or WAL compression by default had no patch or agreement in this thread; discussion ended with Bruce Momjian and Magnus Hagander advocating easier online checksum changes for experimentation and Bruce requesting overhead measurements on a real workload #35 #44 #45 #46.