Performance issues when using preload with large number of rows

[Ihor Tsykalo]

Issue

There is a performance issue when using preload for a large number of rows. Even though, the request itself may be very fast there is a very large amount of time spent in the Bullet::Rack part. Example code:

@comments = Comment.preload(:article).first(5_000)

With bullet turned OFF:

Completed 200 OK in 652ms (Views: 409.3ms | ActiveRecord: 60.8ms | Allocations: 807853)

Request completed in ~0.6 seconds

Ruby-Prof flat output:

Measure Mode: wall_time
Thread ID: 14120
Fiber ID: 15100
Total: 1.916769
Sort by: self_time

 %self      total      self      wait     child     calls  name                           location
 16.54      0.317     0.317     0.000     0.000    10020   IO#write                       
  4.06      0.078     0.078     0.000     0.000        9   PG::Result#values              
  3.07      1.916     0.059     0.000     1.857    71567  *Array#each                     
  2.13      0.041     0.041     0.000     0.000        1   PG::Connection#exec_params     
  1.57      0.433     0.030     0.000     0.403    15020   ActiveSupport::LoggerThreadSafeLevel#add /Users/igortsykalo/.rbenv/versions/2.7.5/lib/ruby/gems/2.7.0/gems/activesupport-7.0.1/lib/active_support/logger_thread_safe_level.rb:50
  1.56      0.186     0.030     0.000     0.156    70719  *Class#new                      
  1.50      0.078     0.029     0.000     0.049   125484  *Hash#fetch                     
  1.26      0.070     0.024     0.000     0.046    35000   ActiveRecord::AttributeMethods#respond_to? /Users/igortsykalo/.rbenv/versions/2.7.5/lib/ruby/gems/2.7.0/gems/activerecord-7.0.1/lib/active_record/attribute_methods.rb:207
  1.15      0.022     0.022     0.000     0.000        7   PG::Connection#async_exec      
  1.14      0.022     0.022     0.000     0.000   169832   Module#===                     
  1.08      0.023     0.021     0.000     0.003       43   <Module::Bootsnap::CompileCache::Native>#fetch 

Full profiling information can be found here: https://github.com/iSarCasm/bullet-performance-app/tree/main/rubyprof

With bullet turned ON:

Completed 200 OK in 1095ms (Views: 643.1ms | ActiveRecord: 60.2ms | Allocations: 1857914)

Request completed in ~10 seconds

Ruby-Prof flat output:

Measure Mode: wall_time
Thread ID: 14340
Fiber ID: 15100
Total: 45.175374
Sort by: self_time

 %self      total      self      wait     child     calls  name                           location
 21.73     16.381     9.816     0.000     6.565 25000015   Set#merge                      /Users/igortsykalo/.rbenv/versions/2.7.5/lib/ruby/2.7.0/set.rb:422
 19.01     39.600     8.589     0.000    31.010   276582  *Array#each                     
 11.75      8.288     5.309     0.000     2.980 25005000   <Class::Bullet::Detector::Association>#call_object_associations /Users/igortsykalo/.rbenv/versions/2.7.5/lib/ruby/gems/2.7.0/gems/bullet-7.0.1/lib/bullet/detector/association.rb:62
  7.68      3.469     3.469     0.000     0.000 25005000   Bullet::Registry::Base#[]      /Users/igortsykalo/.rbenv/versions/2.7.5/lib/ruby/gems/2.7.0/gems/bullet-7.0.1/lib/bullet/registry/base.rb:12
  6.97      4.516     3.151     0.000     1.365    25079  *Array#hash                     
  4.70      2.121     2.121     0.000     0.000 25000004   Hash#update                    
  3.71      1.675     1.674     0.000     0.000 25100087   Thread#[]                      
  3.59      1.622     1.622     0.000     0.000 25005027   Kernel#instance_of?            
  3.22      1.453     1.453     0.000     0.000 25010014   Kernel#instance_variable_get   
  3.18      1.434     1.434     0.000     0.000 26205193   Kernel#class                   
  3.01      1.360     1.359     0.000     0.000 25025185   String#hash                    
  2.93      1.324     1.324     0.000     0.000 25170467   <Class::Thread>#current        
  1.56      0.706     0.706     0.000     0.000     5014   Array#flatten   

Full profiling information can be found here: https://github.com/iSarCasm/bullet-performance-app/tree/main/rubyprof

Steps to reproduce

I have created a sample Rails 7 application with 2 scaffolded models and a seed file hosted here: https://github.com/iSarCasm/bullet-performance-app

It also mentions the full list of steps taken to recreate the problem.

Possible problem

This is probably caused by the UnusedEagerLoading detector (https://github.com/flyerhzm/bullet/blob/master/lib/bullet/detector/unused_eager_loading.rb) which seems to have O(n^2) time complexity since there are 25000015 Set#merge invocation which is ~5000*5000 (5000 Comments with 5000 Articles loaded in 2 DB queries).

[David Santos Merino]

Thanks @iSarCasm for sharing this. We are having a similar performance issue for a large amount of records query. In our case it seems like the issue is with the N + 1 query detector.

[Camilo Vargas A]

It could be related to #623 too.