heroku_delayed_job.rb

## /config/initializers/dynamic_job.rb

require 'heroku'

# base class for all jobs that you wish to automatically scale and go down in Heroku
class DynamicJob
    #set a cap on maximum number of users ever - just in case. 
    MAX_CONCURRENT_WORKERS = 100
    
    def initialize
      #assumption: the job class is new-ed up before enqueueing, which is always the case
      workers_needed = [Delayed::Job.count + 1, MAX_CONCURRENT_WORKERS].min
      if(ENV["RAILS_ENV"] == "production")
          username = ENV['HEROKU_USERNAME']
          password = ENV['HEROKU_PASSWORD']
          app     = ENV['HEROKU_APP']
          client = Heroku::Client.new(username, password)
          client.set_workers(app, workers_needed) 
      end
    end
    
    # call this from the derived class when you are done.
    def clean_up
      #assumption: this method is always called from within a derived job class
      workers_needed = [Delayed::Job.count - 1, 0].max
      if(ENV["RAILS_ENV"] == "production")
        username = ENV['HEROKU_USERNAME']
        password = ENV['HEROKU_PASSWORD']
        app     = ENV['HEROKU_APP']
        client = Heroku::Client.new(username, password)
        client.set_workers(app, workers_needed)
      end
    end
end

#example job class LoadImagesJob
class LoadImagesJob < DynamicJob
  def initialize topic_id
    @topic = Topic.find(topic_id)
    #remember to call super() with parentheses to call superclass initializer without    parameters
    super()
  end
  
  def perform
    @topic.load_images! 
    clean_up
  end
end

#example enqueue
job = LoadImageJob.new topic.id  # => calls some.do_something later
Delayed::Job.enqueue(job)

Yes, the value must be configured to be larger than anything you expect to see in production. The config is provided only for the off chance that a bug causes workers to never be decremented and we want to limit the number of workers so we dont lose our savings:)
Afaik, heroku will try to release workers gracefully, and will then send a force abort after some time. I dont remember the details, but I think an unfinished job whose worker is killed gets picked up by a free worker anyway, so it doesnt matter if an active worker is dropped. Lot of this is based on my testing a few weeks ago, so it may change.if heroku changes.