Seems the speed it really declines with the size of the inserted documents. My production data is a hash with 14-15 keys and string/int values of maybe 10 characters each

When I loop and keep generating these documents:

    to_insert = {}
    10.times {|i| to_insert["key#{i}"] = rand(33333).to_s * rand(6)  }

I get around 140/s. Without the insert() call, I reach 50.000+, so it doesn't seem to be the overhead.

With simple documents (3 keys, int values) I reach 350-400.

p.s. script @ https://gist.github.com/rb2k/5777997