I do not plan to use slave GTID negotiation by default as "classic" negotiation (file and position) allow me to do everything I need (and is more simple to understand, so to maintain in production). For me, classic negotiation must keep working, be supported and actively developed (including adding crash safe replication). Please keep test cases for both and do not plan to deprecate classic negotiation. Jean-François Gagné On 2015-03-06 14:27, Jonas Oreland wrote:

what do you think ?

/Jonas

Jonas Oreland <jonaso@google.com> writes:

what do you think ?

Interesting... It's nice that a number of tests pass by default with GTID. You used MASTER_USE_GTID=current_pos. This means that both replicated transactions and direct transactions on the server contribute to the GTID position. But for old-style replication, only replicated transactions update the position. So perhaps MASTER_USE_GTID=slave_pos should be the default? I would guess that this would cause even more tests to pass directly. It would also be possible to add an extra variant/combination (to run replication tests both with and without GTID). We really need this for the checksum feature as well. But I suppose it's also a question of test resources, replication tests already take a long time... I guess one would need to analyse the test failures before implementing this? A test that is ignored when it fails isn't much of a test? - Kristian.