Deathmatch: Bundler vs. Rubygems.org (Scottish Ruby 2013)

The story of the quest to make `bundle install` faster; in which Rubyists around the world inadvertently DDoS rubygems.org, witness its ignominious death, and vow to rebuild it from the ashes stronger than it was before. Then, a tour of the changes; why is Redis so much slower than Postgres? Marvel at the gorgeous metrics and graphs used to measure and optimize; gasp in delight as we track, live, exactly how many Heroku dynos are needed. Finally, a happy ending: today, the server responds to requests TWO ORDERS OF MAGNITUDE faster than it did before.

Transcript

1. Deathmatch:
   vs
   

   View Slide

2. André Arko
   @indirect
    
    
   

   View Slide

3. View Slide

4. View Slide

5. vs
   

   View Slide

6. !
   

   View Slide

7. Bundler 1.0
   

   View Slide

8. Bundler 1.0
   30 Aug 2010
   

   View Slide

9. Bundler 1.0
   30 Aug 2010
   !
   

   View Slide

10. bundle install
    

    View Slide

11. bundle install
    ...and install
    

    View Slide

12. bundle install
    ...and install
    ...and install
    

    View Slide

13. bundle install
    ...and install
    ...and install
    ...and install
    ...and install
    

    View Slide

14. !
    

    View Slide

15. (what seemed like)
    10 minutes later…
    

    View Slide

16. View Slide

17. bundler API
    

    View Slide

18. bundler API
    just the gems you need
    

    View Slide

19. bundler API
    so. much. faster.
    

    View Slide

20. 32s → 8s
    so. much. faster.
    

    View Slide

21. Bundler 1.1
    

    View Slide

22. Bundler 1.1
    07 March 2012
    

    View Slide

23. Bundler 1.1
    07 March 2012
    !
    

    View Slide

24. API requests
    are lots of work
    

    View Slide

25. pride cometh
    etc., etc…
    

    View Slide

26. dramatic
    reenactment
    

    View Slide

27. !
    !
    

    View Slide

28. !
    !
    !
    

    View Slide

29. !
    !
    !
    ! !
    

    View Slide

30. !
    !
    !
    ! !
    ! !
    

    View Slide

31. no more API
    

    View Slide

32. View Slide

33. View Slide

34. View Slide

35. crap.
    

    View Slide

36. replacement
    a new bundler api
    

    View Slide

37. heroku
    sinatra
    postgres
    

    View Slide

38. questions
    

    View Slide

39. questions
    will it work, this time?
    

    View Slide

40. questions
    how is it bad?
    

    View Slide

41. questions
    how can it be good?
    

    View Slide

42. answers
    need measurements
    

    View Slide

43. we meausure
    a lot of things
    

    View Slide

44. librato
    tracks all the metrics
    

    View Slide

45. View Slide

46. papertrail
    logs all the lines
    

    View Slide

47. View Slide

48. measuring
    leads to knowledge
    

    View Slide

49. discovery 1
    small postgres is fast
    

    View Slide

50. discovery 1
    big postgres is faster
    

    View Slide

51. discovery 2
    redis caching was slow
    

    View Slide

52. discovery 3
    usage is hard to measure
    

    View Slide

53. discovery 3
    usage is hard to measure
    

    View Slide

54. discovery 3
    usage is hard to measure
    

    View Slide

55. discovery 3
    usage is hard to measure
    

    View Slide

56. discovery 4
    threading is great!
    

    View Slide

57. discovery 4
    threading is great!
    …but deadlocks aren’t
    

    View Slide

58. discovery 5
    “response time” isn’t
    

    View Slide

59. discovery 5
    “response time” isn’t
    so check from outside
    

    View Slide

60. then the bugs
    came rolling in
    

    View Slide

61. synchronizing
    is really, really hard
    

    View Slide

62. background
    processing is slow
    

    View Slide

63. webooks
    are unreliable
    

    View Slide

64. both is
    actually pretty good
    

    View Slide

65. after all that
    bundling is fast again
    

    View Slide

66. 2s → 20ms
    two entire orders of
    magnitude faster
    

    View Slide

67. it’s okay,
    I guess
    

    View Slide

68. we can do
    way better
    

    View Slide

69. and we
    have a plan
    

    View Slide