Kristian Nielsen added a comment - 2024-12-03 14:07 Patch: https://lists.mariadb.org/hyperkitty/list/commits@lists.mariadb.org/thread/QCTK775G3SFVGFGDZGLYSNNGSCWMTVFG/ The direct cause of this is that the particular REPLACE INTO () VALUES (), () causes a special case where a write rows event is binlogged with a zero-length after-image. At some point in the history, the code was changed so that m_rows_buf is not allocated when not needed, but is_valid() was not updated to reflect this, so is_valid() incorrectly returns false for this write rows event. And this causes the failure when the second () is binlogged as an update rows event which ends up checking is_valid() for the prior write rows event. So this looks like a regression, though I did not spend the effort to research the commit history to discover exactly which version it was introduced in. There are a number of other places in the code where it explicitly handles the case that the after image can be empty like this.

Kristian Nielsen added a comment - 2024-12-03 14:09 Nice find btw, elenst ! It is quite bad that a simple REPLACE statement fails to binlog outright like this, though it is of course a rather special corner case which makes it less of a problem in practice probably.