Re: draft RFC: concept for partial, wal-based replication

Boszormenyi Zoltan wrote:

> c. splitting wal into different replication sets

Just a side note: in addition to its use for partial replication, this
might have potential for performance-prioritizing databases or tablespaces.

Being able to separate WAL logging so that different DBs, tablespaces,
etc went to different sets of WAL logs would allow a DBA to give some
databases or tablespaces dedicated WAL logging space on faster storage.
If partial recovery is implemented, it might also permit less important
databases to be logged to fast-but-unsafe storage such as a non-BBU disk
controller with write cache enabled, without putting more important
databases in the same cluster in danger.

More importantly, if the WAL writing was done in different wal writer
backends, the admin could also use nice and ionice to encourage the OS
to favour WAL logging for some DBs over others.

Currently all these things require splitting the install into multiple
clusters, incurring config management and backup overhead and most
importantly partitioning shared memory.

OTOH, even with split WAL logging, you still have the shared bgwriter to
contend with, and the effects of an unimportant query pushing data
related to more performance-critical DBs out of shm or OS cache. So
perhaps splitting the cluster is actually the best answer, and a
complete implementation of DB prioritization would land up looking a lot
like multiple Pg clusters multiplexed on one port anyway...

In any case, I thought it worth mentioning as something that may be
worth keeping in mind - or considering and disregarding - while looking
at the WAL changes involved in partial replication.

--
Craig Ringer