Improves the performance of the Redis set cache

Testing has shown that treating membership optimistically reduces the
number of Redis calls substantially.

`include?` will only perform an EXISTS if SISMEMBER returns negative.

`fetch` will only perform an EXISTS if there are no values.

`read` uses SSCAN to avoid Redis lockup on huge sets

lib/gitlab/repository_cache_adapter.rb

@@ -58,11 +58,16 @@ module Gitlab
         # wrong answer. We handle that by querying the full list - which fills
         # the cache - and using it directly to answer the question.
         define_method("#{name}_include?") do |value|
-          if strong_memoized?(name) || !redis_set_cache.exist?(name)
-            return __send__(name).include?(value) # rubocop:disable GitlabSecurity/PublicSend
-          end
+          return __send__(name).include?(value) if strong_memoized?(name) # rubocop:disable GitlabSecurity/PublicSend
+
+          # If the member exists in the set, return as such early.
+          return true if redis_set_cache.include?(name, value)
+
+          # If it did not, make sure the collection exists.
+          # If the collection exists, then item does not.
+          return false if redis_set_cache.exist?(name)
 
-          redis_set_cache.include?(name, value)
+          __send__(name).include?(value) # rubocop:disable GitlabSecurity/PublicSend
         end
       end
 

lib/gitlab/repository_set_cache.rb

@@ -25,7 +25,7 @@ module Gitlab
     end
 
     def read(key)
-      with { |redis| redis.smembers(cache_key(key)) }
+      with { |redis| redis.sscan_each(cache_key(key)).to_a }
     end
 
     def write(key, value)
@@ -47,11 +47,10 @@ module Gitlab
     end
 
     def fetch(key, &block)
-      if exist?(key)
-        read(key)
-      else
-        write(key, yield)
-      end
+      result = read(key)
+      return result unless result.empty?
+
+      write(key, yield)
     end
 
     def include?(key, value)