PERF: avoid expensive order by random for suggested topics

app/services/random_topic_selector.rb

@@ -0,0 +1,75 @@
+class RandomTopicSelector
+
+  BACKFILL_SIZE = 3000
+  BACKFILL_LOW_WATER_MARK = 500
+
+  def self.backfill(category=nil)
+
+    exclude = category.try(:topic_id)
+
+    # don't leak private categories into the "everything" group
+    user = category ? CategoryFeaturedTopic.fake_admin : nil
+
+    options = {
+      per_page: SiteSetting.category_featured_topics,
+      visible: true,
+      no_definitions: true
+    }
+
+    options[:except_topic_ids] = [category.topic_id] if exclude
+    options[:category] = category.id if category
+
+    query = TopicQuery.new(user, options)
+    results = query.latest_results.order('RANDOM()')
+                   .where(closed: false, archived: false)
+                   .limit(BACKFILL_SIZE)
+                   .reorder('RANDOM()')
+                   .pluck(:id)
+
+    key = cache_key(category)
+    results.each do |id|
+      $redis.rpush(key, id)
+    end
+    $redis.expire(key, 2.days)
+
+    results
+  end
+
+  def self.next(count, category=nil)
+    key = cache_key(category)
+
+    results = []
+
+    left = count
+
+    while left > 0
+      id = $redis.lpop key
+      break unless id
+
+      results << id.to_i
+      left -= 1
+    end
+
+    backfilled = false
+    if left > 0
+      ids = backfill(category)
+      backfilled = true
+      results += ids[0...count]
+      results.uniq!
+      results = results[0...count]
+    end
+
+    if !backfilled && $redis.llen(key) < BACKFILL_LOW_WATER_MARK
+      Scheduler::Defer.later("backfill") do
+        backfill(category)
+      end
+    end
+
+    results
+  end
+
+  def self.cache_key(category=nil)
+    "random_topic_cache_#{category.try(:id)}"
+  end
+
+end

lib/topic_query.rb

@@ -414,7 +414,16 @@ class TopicQuery
         result = result.order("CASE WHEN topics.category_id = #{topic.category_id.to_i} THEN 0 ELSE 1 END")
       end
 
-      result.order("RANDOM()")
+      # Best effort, it over selects, however if you have a high number
+      # of muted categories there is tiny chance we will not select enough
+      # in particular this can happen if current category is empty and tons
+      # of muted, big edge case
+      #
+      # we over select in case cache is stale
+      max = (count*1.3).to_i
+      ids = RandomTopicSelector.next(max) + RandomTopicSelector.next(max, topic.category)
+
+      result.where(id: ids)
     end
 
     def suggested_ordering(result, options)

spec/services/random_topic_selector_spec.rb

@@ -0,0 +1,27 @@
+require 'spec_helper'
+
+describe RandomTopicSelector do
+
+  it 'can correctly use cache' do
+    key = RandomTopicSelector.cache_key
+
+    $redis.del key
+
+    4.times do |t|
+      $redis.rpush key, t
+    end
+
+    RandomTopicSelector.next(2).should == [0,1]
+    RandomTopicSelector.next(2).should == [2,3]
+  end
+
+  it 'can correctly backfill' do
+    category = Fabricate(:category)
+    t1 = Fabricate(:topic, category_id: category.id)
+    _t2 = Fabricate(:topic, category_id: category.id, visible: false)
+    _t3 = Fabricate(:topic, category_id: category.id, deleted_at: 1.minute.ago)
+    t4 = Fabricate(:topic, category_id: category.id)
+
+    RandomTopicSelector.next(5, category).sort.should == [t1.id,t4.id].sort
+  end
+end