[Spread-users] Spread for HA rpc services

[Colin Meyer]

On Wed, Aug 08, 2007 at 05:26:36PM -0400, Daniel F. Savarese wrote:
> 
> Since Colin's questions have already been answered, I just want to
> touch on Colin's original description of wanting the "first available
> server" to process a request.  The approaches described amount to
> round-robin load balancing, which may or may not achieve the same
> effect depending on your application.  

Agreed. Van's suggestion of hashing the input would would perhaps lead
to a better distribution of jobs than round robin.

> That's why I said there are
> a number of different ways to go about it (exact answers are always
> application-specific).  If the time to process a request is not a
> constant, either because of heterogeneous hardware or because of the
> nature of the processing, you run the risk of underusing or starving
> your fast workers and piling up unprocessed requests for your slow
> workers.

In my case the hardware will be homogenous, and the jobs will typically
be well within an order of magnitude of each other (with the bulk of them
being nearly identical in difficulty). 

I suppose that one approach could be for a given server to remove itself
from the queue, while it was feeling overworked.

-Colin.