On Sat, 13 May 2000, Greg Cope wrote:
> : Likewise with sessions. Even if you load balance across multiple machines
> : you don't need to access a session database on every request. Most load
> : balancing systems have something so they'll send the seme "session"
> : (typically ip address) to the same backend server for N seconds as long as
> : the backend server is up of course.
> :
> : You can then put a daemon on each backend server as a local buffer and
> : cache between the httpds and the session database. This daemon can even
> : load balance the session information over several databases if it's
> : needed.
> :
> : viola, a lot less "real time" lookups in the database needed and a lot
> : less direct connections.
[...]
> But - is this model not transaction safe ? - It may save alot of DB use and
> hence be alot quicker - but what if the webserver "disappears" halfway
> through an IP based load ballanced "sesssion" - a potential source of
> failure - something which a load ballancer is supposed to help with..
>
> Also if the loadballance goes south, and a secound takes over does it have
> the same IP based session info (probably not - but could be wrong ...) -
> hence another potential source of failure.
You have to do a write-through cache, i.e. you must write to the database
as well as the cache, but you only have to READ the database if you don't
find the session in the cache. If you have a fair amount of read-only use
of the session this can be a big win.
- Perrin
