Hi Jim
Could you elaborate a bit on the overhead of indices? A single
index-per-object is gives us a perfect setup for our dataflow, but how
far can this scale reasonably? Though we're starting out at ~20,000
indices, I'd like to be able to plan for 1 million if possible.If this isn't too far fetched, which is the more important scaling
factor, node size or node count?
That is indeed far fetched 
Unless you plan on having hundreds of
thousands of servers...
However, filtered aliases might be what you are after.
I suggest taking a look at Shay's presentation at berlin buzzwords this
year:
clint
Thanks,
-jim