Puppet Scalability - any good references

[matan]

Hi, 

Could anyone point out a good existing discussion of Puppet scalability?

I'm relating to the Puppet master and ecosystem parts that are not the actual agents sitting on the managed servers.

In particular anything that would shed insight upon:

1. Does a Puppet master gracefully degrade when overwhelmed?
   Or does stuff start failing, instead of just performing more slowly.
2. How does changing the Puppet polling interval (runinterval etc.) factor in?
   
   Does Puppet make it safe to increase workload and polling frequency knowing that at worst there'll be slowness, or does it leave it to the operator's gut and trial and error figuring how much load is fine, requiring of them to throttle workloads and hold their breaths when rolling out changes. Can workload be cancelled in case of excessive loads same as handling parallel FTP jobs, or this approach is not a design tenet at present?
   I'm assuming the answers in this area are not 0 nor 1's, so a balanced discussion of how close are things to 0 or to 1 is of interest.

Thanks,

matan