On a collection containing 3 million+ records, Paranoia pretty much trashes the normally-fast count() method. This is because, instead of asking Mongo for its count of the collection, we're actually doing a query counting all elements where deleted_at.ne => nil. An index on deleted_at (which ends up being worthless in most cases... our collection has 3 million+ records and maybe a dozen soft-deletions), it speeds it up some, but not nearly as much as would be desirable.

Before index:

Benchmark.measure { Product.count }
 =>   0.000000   0.000000   0.000000 (  218287.7)

After index on deleted_at:

Benchmark.measure { Product.count }
 =>   0.000000   0.000000   0.000000 (  3.594815)

Still not great. So here's what I came up with for our system:

 def self.count
    collection.count - deleted.count
  end

Benchmark on that method:

Benchmark.measure { Product.count }
 =>   0.000000   0.000000   0.000000 (  0.000808)

Huzzah!

So my question is: what do people think about making this the standard behavior of the count method in the Mongoid::Paranoia module? Certainly, if you're doing mostly soft-deletes with few un-deleted records in your collection, this makes less sense, but that seems like the exception rather than the rule. Since Rails is about convention, wondering if this might be a good change? Or is there some fundamental reason I'm missing as to why this is a bad idea?
(I'd be happy to submit a pull request.)