It can be whatever type of client, hence the imprecision ;-)

Here is the use case I have in mind. We have several different
applications using the same DB and so the same cache. However, they
have very different DB/cache usage patterns, and only share a small
amount of data.
With current statistics granularity, it is hard to see if one
application is using significantly more Memcached, or if one other has
a ginificant cache miss ratio ...
Databases usually offer per module statistics, this could somehow be
the equivalent for Memcached.

I guess in a web application context, application segmentation has no
real sense, it would rather be based on the user family.

---
Jean-Charles

On Apr 20, 7:21 pm, Dustin <[email protected]> wrote:
>   What are the types of clients you're talking about?  Can you provide
> a couple of use cases?  Are these for different applications, or just
> different users of the same application?
>
> On Apr 20, 2:08 am, JC <[email protected]> wrote:
>
>
>
> > Hi everyone,
>
> > Currently there are many statistics available in memcached, but all
> > these statistics are aggregated in a global way or based on their
> > hosting slabs.
> > This could be useful to also have aggregation based on client type:
> > this could helps to have a clear view on the impact/load/behaviour of
> > the various clients interacting with the cache.
>
> > It does not seem really difficult to keep in the connection structure
> > a pointer to a "current client type" statistics structure (can still
> > be per thread for performance), that would be updated at the same time
> > the main statistics are updated.
> > The stats command + the identifier of the client type could be used to
> > retrieve these statistics.
> > The trickest point is maybe how to define the type of the current
> > client since there is no handshake like message. Maybe this could be
> > done by an additional client call that would give either a name
> > identifying a type or directly a type id so that internal connection
> > structure is updated accordingly. ("nasty" solution can be to use the
> > noop message for that ...)
>
> > Of course, the number of different possible types would be limited to
> > avoid using too much memory !
>
> > what to you think of such a feature ?
>
> > Jean-Charles

Reply via email to