Remove global 'busy' counter, too easy to skew in the face of crashy processes

lib/sidekiq/api.rb

@@ -497,8 +497,20 @@ module Sidekiq
       end
     end
 
+    # Not very efficient if you have lots of Sidekiq
+    # processes but the alternative is a global counter
+    # which can easily get out of sync with crashy processes.
     def size
-      Sidekiq.redis { |conn| conn.get('busy') }.to_i
+      Sidekiq.redis do |conn|
+        procs = conn.smembers('processes')
+        return 0 if procs.empty?
+
+        conn.multi do
+          procs.each do |key|
+            conn.hget(key, 'busy')
+          end
+        end.map(&:to_i).inject(:+)
+      end
     end
   end
 

lib/sidekiq/processor.rb

@@ -80,7 +80,6 @@ module Sidekiq
         hash = Sidekiq.dump_json({:queue => queue, :payload => msg, :run_at => Time.now.to_i })
         Sidekiq.redis do |conn|
           conn.multi do
-            conn.incr('busy')
             conn.hmset("#{identity}:workers", Thread.current.object_id, hash)
             conn.expire("#{identity}:workers", 60*60)
           end
@@ -106,7 +105,6 @@ module Sidekiq
           Sidekiq.redis do |conn|
             processed = "stat:processed:#{Time.now.utc.to_date}"
             result = conn.multi do
-              conn.decr('busy')
               conn.hdel("#{identity}:workers", Thread.current.object_id)
               conn.incrby("stat:processed", 1)
               conn.incrby(processed, 1)

test/test_api.rb

@@ -378,14 +378,6 @@ class TestApi < Sidekiq::Test
         c.hmset(s, '1234', data)
       end
 
-      Sidekiq.redis do |c|
-        assert_equal 0, w.size
-        c.incr("busy")
-        assert_equal 1, w.size
-        c.decr("busy")
-        assert_equal 0, w.size
-      end
-
       w.each do |x, y|
         assert_equal "1234", x
         assert_equal 'default', y['queue']