> On Jul 18, 2022, at 20:03, merryok <merryok@xxxxxxx> wrote: > > I've read the doc, and it doesn't help too much. > Finally I've found START_CRIT_SECTION and END_CRIT_SECTION. It's like mtr.start(), mtr.commit() in mysql. May I ask why many places are wrapped into START_CRIT_SECTION/END_CRIT_SECTION during a single dml operation ? A PostgreSQL critical section is not the equivalent of MySQL InnoDB mini-transaction. A critical section in PostgreSQL is a section of code that needs to run without interruption to avoid corruption of internal in-memory data structures. PostgreSQL doesn't have a direct equivalent of a MySQL mini-transaction. When WAL information is created by a statement, it's stored in the WAL buffers, and then flushed to disk by the WAL writer (to a first approximation). There's no special operation that groups pages together for atomic writes; that's done by the underlying file system flush operation. > And if Assert(CritSectionCount > 0) isn't satisfied (CritSectionCount need't be protected ?), PG server will panic and exit ? If so, what's the probability of that ? If it occurs, it indicates a bug in PostgreSQL. It is *extremely* infrequent (as in, you can go years without seeing one; I can't remember the last time I did).