On 11/11/15 4:22 AM, Thomas Munro wrote:
> My thinking was that the reason for wanting to load balance over a set
> of hot standbys is because you have a very read-heavy workload, so it
> makes sense to tax the writers and leave the many dominant readers
> unburdened, so (3) should be better than (2) for the majority of users
> who want such a configuration.
One problem I can see is that even if you have a read-heavy workload,
the writes can still be a bottleneck, since they are necessarily bound
to one node. And so if the feature proposal is, we can make your reads
more consistent but the writes will become slower, then that's not a
good deal.
More generally, no matter whether you pick the writers or the readers to
wait, if you assume that read-only slaves are an application performance
feature, then it's questionable how much better such applications will
perform overall when network-bound waits are introduced in the system.
I think in practice applications that are busy enough to worry about
this don't really work like that anyway. For example, the writes should
go to a message queue and are written out whenever, with a copy kept in
a cache for display in the meantime. Maybe there could be additional
features to make managing this easier.
I think there are a lot of different variations of this in practice, not
only depending on the workload and other measurables, but also
business-dependent decisions on application behavior and degradability.