Consider the following Ruby code:

require 'mongo'
DB = Mongo::Connection.new('localhost', 27017)['test_db']
@collection = DB['test_collection']

code = -> { @collection.find(

).each

}
high = 0
pid = 0
size = 0

loop do
th = Thread.new do
code.call
pid, size = `ps ax -o pid,rss | grep -E "^[[:space:]]*#

"`.chomp.split(/\s+/).map

if size > high
puts "New high: #

"
high = size
p "Threads: #{ObjectSpace.each_object(Thread) {}}"
p "Strings: #{ObjectSpace.each_object(String) {}}"
end
end
th.join
end

If you look at the output of this code, Thread objects are never garbage collected after the thread does a query using the MongoDB driver. If you comment out "code.call", thread objects are garbage collected and usage stays down. Try inserting a few objects with string values into the test db/collection and running the above code, and you'll see a steady increase in string objects.

This pretty much makes Mongo + EventMachine.defer (or any thread pool) a no go.

Also a note, this does not happen with Fiber's – only Threads.