Aleksey, hello.
Hi, Sergei,
Thank you for observations! This task is in progress. While doing it I found out that RBR doesn't replicate timestamp fractions. That is: it gets seconds from event timestamp marker and adds fractions from system time. This causes bugs in System Versioning. More on it in Bug 6 here: https://github.com/tempesta-tech/mariadb/issues/578#issuecomment-470533050
I propose to fix it by adding timestamp fractions field to Log_event header after FLAGS_OFFSET. To support backwards compatibility it is needed to increment fdle->binlog_version and probably add new flag LOG_EVENT_HAS_SEC_PART_F (for support sending Log_event without fractions). Do you agree?
Adding a new flag to Log_event could indeed force the binlog_version increment. However as you're concerned with Rows_log_event the new field could be added to this class and its header. You can find how to do that in Rows_log_event::Rows_log_event. Hope this is helpful. Andrei
On Thu, Jan 31, 2019 at 1:10 PM Sergei Golubchik <serg@mariadb.com> wrote:
Hi, Aleksey!
On Jan 31, Aleksey Midenkov wrote: > revision-id: 14c2f90ad08 (versioning-1.0.7-1-g14c2f90ad08) > parent(s): a8efe7ab1f2 > author: Aleksey Midenkov <midenok@gmail.com> > committer: Aleksey Midenkov <midenok@gmail.com> > timestamp: 2018-11-17 16:30:10 +0300 > message: > > Idempotent INSERT events for system versioning > > Case 1: Rows_log_event::write_row() always overwrite historical row. > > Related to MDEV-16370.
I realized that the whole RBR logging of system versioned tables is wrong.
For an update you log update_row event and write_row event separately and execute them separately. It allows trivially to manipulate the history. See a test case attached.
To fix this, the slave should ignore all events that modify the history. And generate them locally. Say, an update on the master generates update_row and write_row events. The slave executes update_row event which updates a current row and this also should generate the historical row. Then the slave ignores write_row event.
Of course, as an optimization, the master should not generate historical write_row events, but as my test case shows, they can be forged, so the slave should ignore them anyway, if they happen to come.
I've created a new https://jira.mariadb.org/browse/MDEV-18432 for that, let's not have it tied to MDEV-16370.
Regards, Sergei Chief Architect MariaDB and security@mariadb.org
-- All the best,
Aleksey Midenkov @midenok
_______________________________________________ Mailing list: https://launchpad.net/~maria-developers Post to : maria-developers@lists.launchpad.net Unsubscribe : https://launchpad.net/~maria-developers More help : https://help.launchpad.net/ListHelp