[Date Prev][Date Next]
Re: Documentation volunteer (was: Cluster replication: multiple masters)
Howard Chu wrote:
When you can solve the problems in
then you might have something worth writing about.
In the general case, there are unsolvable problems with conflict resolution.
There is a specific case where you might profitably use multimastering: in a
situation where you have multiple redundant servers, but only a single one is
ever targeted for write operations (thus avoiding change conflicts). In this
scenario, if the "write server" crashes you can immediately failover to any
one of the other redundant servers; in essence you are still doing
master-slave replication, but with automatic promotion of a slave to master
status. It is not straightforward to enforce the single-writer usage though;
i.e., there's no simple way to ensure that clients only write to the
currently selected write server.
Yes there is.
We've been running the setup Derek mentioned for more than 2 years now.
We have two IP on the load balancer, one for reading and the other one
for writing. The reader one does round-robin but the writer one
implements fail-over, i.e. the second server is only ever used if the
first one is down.
Works fine (and hey - we're a large ISP and I guess we would be broke
by now if it didn't).
Granted, the problem is unsolvable (at least if using replication).
But I'd like to kindly ask you please make multimaster replication an
"official" feature, add a warning about the aforementioned problem in
the docs but leave the decision whether to use it (and accept the
impact) up to the user.