> -----Original Message----- > From: [email protected] [mailto:baruwa- > [email protected]] On Behalf Of Noah Mehl > Sent: 10. august 2013 19:17 > To: Baruwa users list > Subject: Re: [Baruwa] Contemplating replacing mailwatch with baruwa > > Jonas, > > > > > Hi Noah > > > > I'm not sure where you have your information from, but it's wrong. I do not > even know what Galera is, nor do I have 3 servers. > > Then you should go research Galera http://codership.com/content/using- > galera-cluster and http://www.percona.com/software/percona-xtradb- > cluster because it's has "mostly-true" synchronous replication for MySQL. In > this mode, you are "semi-garuanteed" that your write has been completed > on all servers.
Umm I think you're missing the whole point. I do not want or need synchronouse replication. It's not at all necessary in a mailscanner cluster (and have'nt been for the several year's we've been running our mailscanner cluster) Having synchronous replication would only slow things down. > > MySQL master-master replication have worked fine out of the box at least > since 5.x > > Yes, there is a replication mode in MySQL that allows for a two-server > Master/Master configuration. However it's NOT synchronous. Therefore > you have no guarantee that the write has happened on the other server. > There are other HA configurations for MySQL: Master/Slave with > asynchronous replication, and Master/Slave with DRBD, but they both have > their own issues. > > That being said, you either have replication lag with a two server MySQL > setup, or you can have lock contention in Galera if you're hitting the same > table, which is probably what would happen with Baruwa. > Ohh it allows for far more than 2 master's Its normal to have 4 or more as well, if needed. Again see above for why synchronous replication isnt needed. > > The reason why you need more than 1 master is to be able to still write to > the DB when the primary master goes down, as it would in a normal > standalone setup or master-slave setup. > > With a plethora of HA setups, you can automatically switch your slave to > master to be able to write within a very small window (transparent to the > end client). Especially with PostgreSQL synchronous replication, when you > can be assured that your DB will stay consistent and up to date during a > failure scenario. Most applications (including Baruwa, since it uses > SQLAlchemy) should have no problem reconnecting to your DB layer after a > failure scenario. I'm not sure why you think that a 30-60 second failover > scenario would not be appropriate for an application like Baruwa? I would > always prefer 30-60 seconds of failover time and DB consistency. I would always prefer NO failover time. Running in active-active or master-master mode is simply superior as you know everything works. Not to mention you actually utilize your resources fully instead of a typical slave or similar setup where you write to 1 server and in the best case read from many. > > > > Not to mention its better performance to write on many, and on top of > that introducing anything with a failover is a risk in itself, since you never > know if it works when you need it. > > Here your information is just plainly incorrect. There is not "better > performance" by writing to many. First of all you can lose consistency very > quickly. Secondly, you're doubling your writes on each server in a two-server > scenario. Performance have nothing to do with consistency, when all daemons/scripts communicate with a local server the performance is obviously better than a single, potentially overloaded DB server, not to mention the network latency it introduces. I hope you have a local DNS server on your nodes as well, the principles are the same, local beats network. > > With true master-master replication you know everything works, because > you use both masters everyday, unlike a passive/cold/slave server. > > This again is where you are incorrect. I urge you to look into consulting > from > experts in this field, such as: http://www.hastexo.com and > http://www.percona.com > > Hope that explains it :) > > You're not explaining anything. And you're giving incorrect information to > the list. It sounds to me like you haven't worked with mysql replication on a larger scale if at all, so please don't try to divert the issue at hand. Bottom line is, many many corporations are running mysql multi-master setup's for mailscanner, since it's the best architecture for the task. Currently, see Andrew's reply, the baruwa architecture doesn't offer anything with quite the same characteristics and advantages/disadvantages. _______________________________________________ http://pledgie.com/campaigns/12056
