In the end, a stat delay is what we're going with.  Proportional is the most FAIR distribution - Every valid share is worth the same as every other in the round.



One of the things I did when trying to find out why the load balancer was crashing at long polls was remove the 'hub mode' feature from our bitcoind in case it was trying to spam too many p2p connections with other nodes.  I re-deployed JoelKatz's bitcoind "hub-mode patch" this evening after our invalid.  The two invalids in the last 24 hours were real invalids (the previous 4 from earlier last week were bugged bitcoind/load balancer crashing related).  The new invalids were simply other pools producing a block very close to our own, and the block chain favoring the other block.  Hub mode essentially makes bitcoind aggressively connect to other nodes to maintain a certain number of peers.

With hub mode re-enabled, we should now spread our blocks out among the other nodes on the bitcoin network significantly faster, and we will get the new blocks faster as well to reduce the odds of an invalid being produced.  After a few hours, it has been running strong and not causing any unusual long poll performance issues.