maria-developers team mailing list archive

[Michael Widenius <monty@xxxxxxxxxxxx>]

Hi!

>>>>> "Kristian" == Kristian Nielsen <knielsen@xxxxxxxxxxxxxxx> writes:

<cut>

Kristian> I am not very familiar with timestamps in binlog events. However, all the
Kristian> transactions in the group commit are written out one after the other, in a
Kristian> single thread holding the lock on the binlog. And the GTID events are
Kristian> generated during that loop. So I think that at least the timestamps of the
Kristian> GTID events in the group can never walk backwards, nor can they from one group
Kristian> to the next. But I am not sure if the time stamps of other events in the
Kristian> transaction can be earlier (maybe they were generated when the query was run,
Kristian> not when it was committed?

Time stamps in a statement is based on when the statement started, not
when it committed.  This is needed to ensure that all rows has the
same value f you do:

update big_table set timpestamp=now();

So it's normal that in a group commit, the transactions in the same
commit can have totally different timestamps.  Longer running
transactions will typically have older timestamps.

Regards,
Monty