* Elielson Fontanezi <ElielsonF@prodam.sp.gov.br> [2004-03-02 15:47:13 -0300]:
Hi,
> I am in the need of using some multi-master replication for
> postgres.
I've implemented something in php. All my applications do their database
access layer, where they only touch whole objects (the structure of
each object class is modeled in class descriptors) - so it is quite
easily solvable.
But this works only under certain conditions:
* all objects have an inode_id (as cluster-wide oid) and mtime field.
* on create (in sql speaking: insert) a new inode_id is allocated
from the inode_id sequence. this sequences has to be adjusted
carefully so each cluster node gets its own id space.
* on each write access, the mtime has to be set to the current time.
* from time to time (i.e. once per minute) a pollout script runs
over the object classes and fetches out the new records (by timestamp).
they are the posted to the other nodes.
(not necessarily to all - you can define sync neighbours on each node)
Well, this is no general solution, but under some conditions
(which are btw design rules for all my database applications),
this works quite well.
regards,
--
---------------------------------------------------------------------
Enrico Weigelt == metux IT services
phone: +49 36207 519931 www: http://www.metux.de/
fax: +49 36207 519932 email: contact@metux.de
cellphone: +49 174 7066481
---------------------------------------------------------------------
-- DSL-Zugang ab 0 Euro. -- statische IP -- UUCP -- Hosting --
---------------------------------------------------------------------