Introduce new use_replica method in DB load balancer

Issue https://gitlab.com/gitlab-org/gitlab/-/issues/322133

ee/lib/gitlab/database/load_balancing/connection_proxy.rb

@@ -57,9 +57,21 @@ module Gitlab
           end
         end
 
+        def transaction(*args, &block)
+          if ::Gitlab::Database::LoadBalancing::Session.current.use_replica?
+            read_using_load_balancer(:transaction, args, &block)
+          else
+            write_using_load_balancer(:transaction, args, sticky: true, &block)
+          end
+        end
+
         # Delegates all unknown messages to a read-write connection.
         def method_missing(name, *args, &block)
-          write_using_load_balancer(name, args, &block)
+          if ::Gitlab::Database::LoadBalancing::Session.current.use_replica?
+            read_using_load_balancer(name, args, &block)
+          else
+            write_using_load_balancer(name, args, &block)
+          end
         end
 
         # Performs a read using the load balancer.

ee/lib/gitlab/database/load_balancing/session.rb

@@ -55,6 +55,28 @@ module Gitlab
           @ignore_writes = false
         end
 
+        # Indicate that all the SQL statements from anywhere inside this block
+        # should use a replica. This is a weak enforcement. The queries
+        # prioritize using a primary over a replica in those cases:
+        # - If the queries are about to write
+        # - The current session already performed writes
+        # - It prefers to use primary, aka, use_primary or use_primary! were called
+        def use_replica(&blk)
+          return yield if use_primary? || performed_write?
+
+          begin
+            used_replica = @use_replica
+            @use_replica = true
+            yield
+          ensure
+            @use_replica = used_replica && !use_primary? && !performed_write?
+          end
+        end
+
+        def use_replica?
+          @use_replica && !use_primary? && !performed_write?
+        end
+
         def write!
           @performed_write = true