To fix that, pg_rewind could always start the rewinding process from the last checkpoint before the point that the histories diverge, instead of the exact point of divergence.
Is that something required even if someone plans to use pg_rewind for a cluster with checksums enabled ? I mean since only first update after checkpoint is WAL logged, pg_rewind will break if another update happens after standby forks. Or would the recovery logic apply first WAL without looking at the page lsn ? (Sorry, may be I should read the code instead of asking you)
If we do what you are suggesting, it seems like a single line patch to me. In XLogSaveBufferForHint(), we probably need to look at this additional GUC to decide whether or not to backup the block.
That would make the rewinding more expensive as it needs to read through a lot more WAL, but I think it would still be OK.
Yeah, probably you are right. Though the amount of additional work could be significantly higher and some testing might be warranted.
