Making CocoaPods Fast

Transcript

1. Making CocoaPods Fast
   with Modern Ruby Tooling
   Samuel Giddins
   1
   

   View Slide

2. @segiddins
   Mobile Developer Experience @ Square
   CocoaPods Core Team
   Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 2
   

   View Slide

3. bundler & rubygems
   Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 3
   

   View Slide

4. What even is CocoaPods?
   Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 4
   

   View Slide

5. ==
   +
   Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 5
   

   View Slide

6. CocoaPods combines a definition of libraries (podspecs)
   with a way to integrate theminto a user's application (podfile)
   Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 6
   

   View Slide

7. RubyGems Bundler CocoaPods
   Gemspec Podspec
   Gemfile Podfile
   Gemfile.lock Podfile.lock
   rubygems.org index.rubygems.org github.com/CocoaPods/
   Specs
   Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 7
   

   View Slide

8. That little difference: Xcode
   Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 8
   

   View Slide

9. Xcode
   ➡ Proprietary toolchain for compiling apps for  platforms
   ➡ Uses its own manifest file format
   ➡ Invokes many compilation tools
   Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 9
   

   View Slide

10. So, CocoaPods has to do a heck of lot
    more than RubyGems or Bundler
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 10
    

    View Slide

11. ➡ Updating local specs repositories
    ➡ Analyzing dependencies
    ➡ Fetching dependencies
    ➡ Installing dependencies
    ➡ Generating Pods project
    ➡ Integrating client project
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 11
    

    View Slide

12. So, CocoaPods exists!
    It was slow?
    So what.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 12
    

    View Slide

13. Why Performance?
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 13
    

    View Slide

14. Why Performance?
    ➡ Optimization is fun!
    ➡ Rapid iteration == happier & more productive developers
    ➡ "Scale"
    ➡ Free performance is hard to find
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 14
    

    View Slide

15. So, I have a slow ruby app.
    How do I make it faster?
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 15
    

    View Slide

16. How do I make it faster?
    ➡ Is it really slow?
    ➡What is it doing?
    ➡How often does it run?
    ➡Would it being faster make a difference?
    ➡ Can it do less work?
    ➡ How can I keep it from getting worse?
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 16
    

    View Slide

17. Is it really slow?
    ➡ How much would I invest to make it 10% faster?
    50%? 90%?
    ➡ How do I know which part of what it's doing
    is slow?
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 17
    

    View Slide

18. ➡
    !
    this feels slow
    ➡
    ⏱
    this took 10 seconds
    ➡
    #
    this segment took 10 seconds
    ➡
    $
    this method was called 97645 times,
    averaging 23ms per call,
    taking up 48% of total runtime
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 18
    

    View Slide

19. Profiling
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 19
    

    View Slide

20. Profiling
    The art of putting numbers to the feeling of
    "this feels slow"
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 20
    

    View Slide

21. What can I profile?
    ➡ Method calls
    ➡ Call Stack Sampling
    ➡ Memory allocations
    ➡ Disk I/O
    ➡ Network I/O
    ➡ Database queries
    ➡ Garbage collection
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 21
    

    View Slide

22. My work on CocoaPods focused on:
    Method Call Profiling
    Call Stack Sampling
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 22
    

    View Slide

23. quick aside:
    One of our contributors got amazing gains
    by profiling allocations
    and memoizing things like #hash & Pathname#join
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 23
    

    View Slide

24. Profilers
    Just like most tools,
    it's important to pick the right one for the job at hand
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 24
    

    View Slide

25. Tracing Profilers
    Tell you how many times something was called,
    and usually how much time is spent in that call.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 25
    

    View Slide

26. Tracing Profilers
    ➡ Can distort relative numbers
    ➡ Can make running a program painfully slow
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 26
    

    View Slide

27. Tracing Profilers
    If it takes 35 minutes with no profiling, it may never finish under
    rubyprof. Sorry, I didn't think of that.
    — CocoaPods/CocoaPods#5180
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 27
    

    View Slide

28. Allocation Profilers
    Tracing profilers that count memory allocations
    instead of method calls.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 28
    

    View Slide

29. Allocation Profilers
    ➡ Suffer from similiar problems,
    but even more so since the typical ruby program
    is incredibly allocation-heavy.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 29
    

    View Slide

30. Sampling Profilers
    Take a peek at you program at regular intervals,
    from the outside.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 30
    

    View Slide

31. Sampling Profilers
    ➡ (Generally) will not interfere at all with your app's normal execution.
    ➡ Possible to miss things,
    it's luck whether a particular stack trace
    gets sampled at the right time.
    ➡ Can’t tell you how many times something has been called.
    ➡ “Here were the stack traces when I looked”
    ➡What was at the top of the stack at the time?
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 31
    

    View Slide

32. Manual Profilers
    Sometimes, you don't want or need a fancy tool.
    The easiest output to understand
    is the one you write yourself.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 32
    

    View Slide

33. Manual Profilers
    ➡ Wrapping method calls in a call to Benchmark.measure
    will tell you how long it took.
    ➡ Can help you focus in on what you already suspect
    to be hotspots
    ➡ More likely to tell you if the distribution of calls is not unimodal
    ➡e.g. calling a memoized method, so only the first call will take a
    significant amount of time
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 33
    

    View Slide

34. Chronometer
    A tool I built to automate this sort of low-fidelity tracing.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 34
    

    View Slide

35. Chronometer
    ➡ Wraps methods with some timing / recording infrastructure
    ➡ Can output results in a format compatible with
    chrome://tracing
    ➡ Happens in-process, so you can write code that uses the
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 35
    

    View Slide

36. Chronofile
    for_class Pod::Installer do
    method :install!, context: event_phase_context[:toplevel]
    methods %i[
    prepare
    resolve_dependencies
    download_dependencies
    validate_targets
    generate_pods_project
    integrate_user_project
    perform_post_install_actions
    run_podfile_post_install_hooks
    ]
    end
    # ...
    for_class Pod::Resolver do
    method :resolve
    end
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 36
    

    View Slide

37. Profilers
    For me, this meant a mix of
    ➡ rubyprof
    ➡Tracing Profiler
    ➡ rbspy
    ➡Sampling Profiler
    ➡ chronometer
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 37
    

    View Slide

38. All 3 are amazing,
    and I used all of them extensively last year.
    But knowing which to reach for at each step of the investigation
    would’ve saved me a bunch of time
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 38
    

    View Slide

39. Chronometer hit the sweet spot
    for the type of software I was developing,
    and also solved a couple other problems I had.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 39
    

    View Slide

40. Profiling revelead a fundemental problem with
    CocoaPods' architecture.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 40
    

    View Slide

41. Graph Traversal
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 41
    

    View Slide

42. Graph Traversal
    The performance bottleneck that all
    build systems end up fighting.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 42
    

    View Slide

43. Graph Traversal
    A
    / \
    B C
    \ /
    D
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 43
    

    View Slide

44. Dependency Graph
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 44
    

    View Slide

45. Graph Traversal
    The base problem we often face is
    How do you find “all the nodes/vertices that come after A”?
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 45
    

    View Slide

46. It's tempting to say:
    “A is followed by B and C, let’s recurse!”
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 46
    

    View Slide

47. It's tempting to say:
    “A is followed by B and C, let’s recurse!”
    But then you end up visiting D twice!
    This might not sound like a big deal,
    but imagine 50 more nodes come after D...
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 47
    

    View Slide

48. It's not just CocoaPods that has this issue!
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 48
    

    View Slide

49. Graph traversal like this happens all over build tools.
    Why?
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 49
    

    View Slide

50. DAGs
    ➡ Directed Acyclic Graphs
    ➡describe dependency structure.
    ➡ Every node (target, library, gem, etc.) has a list of dependencies
    ➡only rule is “you can’t depend upon something that depends on you”
    ➡ Most operations rely on what’s called the “transitive closure” of a node
    ➡“what are all the nodes that come after this one, no matter how far
    away?”
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 50
    

    View Slide

51. def recursive_predecessors
    vertices = predecessors
    vertices += vertices.flat_map(&:recursive_predecessors)
    vertices.uniq!
    end
    to
    def recursive_predecessors
    vertices = Set.new
    visit = ->(vertex) do
    vertex.incoming_edges.each do |edge|
    vertex = edge.origin
    next unless vertices.add?(vertex)
    visit[vertex]
    end
    visit[self]
    vertices
    end
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 51
    

    View Slide

52. Memoization
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 52
    

    View Slide

53. Something that comes along with trying to solve graph traversal
    problems is memoization!
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 53
    

    View Slide

54. From before:
    A
    / \
    B C
    \ /
    D
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 54
    

    View Slide

55. Let’s say some build setting for A
    depends on a build setting of B, C, D.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 55
    

    View Slide

56. This can get slow.
    Really slow.
    Reeeeeeeeeaaaaaaaallllllly slow.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 56
    

    View Slide

57. Memoization
    Calculate the value for the setting once and then store it
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 57
    

    View Slide

58. ➡ This is slightly different from caching
    ➡There’s no “invalidation” that can happen
    ➡ We’re just lazily computing a value, and storing it
    so next time it’s needed,
    we can return the stored value
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 58
    

    View Slide

59. In ruby, we do this a lot! ever seen something like this:
    def expensive_to_calculate
    @expensive_to_calculate ||= ...
    end
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 59
    

    View Slide

60. This works great!
    - But imagine if the return value for expensive_to_calculate can
    be nil, or false.
    - The ||= short-circuiting won’t kick in.
    - We’ll need to recompute the value every time.
    - Not ideal.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 60
    

    View Slide

61. Instead, we need to turn to a more complicated variation on the
    pattern:
    def expensive_to_calculate
    return @expensive_to_calculate if defined?(@expensive_to_calculate)
    @expensive_to_calculate = ...
    end
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 61
    

    View Slide

62. In CocoaPods, when I re-wrote the build settings, I added a tiny
    little DSL to make memoization a bit easier:
    def self.define_build_settings_method(method_name, build_setting: false,
    memoized: false, sorted: false, uniqued: false, compacted: false, frozen: true,
    from_search_paths_aggregate_targets: false, from_pod_targets_to_link: false,
    &implementation)
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 62
    

    View Slide

63. ➡ It uses an ivar hash called @__memoized to store those values,
    calls #fetch on that hash,
    returns the value unless the hash doesnt yet contain the key.
    ➡ One benefit of this is that the memoized state can be easily
    discarded by clearing that one ivar, instead of needing to keep track
    of a bunch of ivars, each only used in a single method
    implementation.
    ➡ By carefully selecting the key, we’re able to memoize both what a
    superclass and subclass return separately, so multiple calls to super
    can also be memoized.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 63
    

    View Slide

64. This DSL (and accompanying refactor)
    allowed me to fix years-old bugs in build settings generation.
    Bug fixes which would've slowed installation down by over 5
    minutes
    under the old implementation.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 64
    

    View Slide

65. Design Lessons Learned
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 65
    

    View Slide

66. Design Lessons Learned
    Even after all this work (and it’s still ongoing!),
    CocoaPods isn't as fast as I’d like.
    And it never will be.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 66
    

    View Slide

67. ➡ some of this comes down to language
    ➡ some of this comes down to the fact that
    the job CocoaPods does is inherently complicated
    ➡ some of it is our own fault,
    for having a system design from 7 years ago
    that wasn’t built to scale this far
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 67
    

    View Slide

68. Some particulars
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 68
    

    View Slide

69. Exposing Mutable Objects
    Makes memoization hard.
    Need to have sidecar objects that hold computed values only when
    its safe to do so.
    (unless you want to get into cache invalidation bugs)
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 69
    

    View Slide

70. Reading & Writing to the file system in-line
    Makes it very hard to parallelize IO.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 70
    

    View Slide

71. Having Consistent CLI Output
    Once again, makes parallelization hard.
    How do you show progress,
    underlying command invocations,
    failures,
    etc?
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 71
    

    View Slide

72. Inefficient Data Structures
    Particularly the podfile & podspec.
    Built using nested hashes, making change-tracking hard.
    Lack of copy-on-write.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 72
    

    View Slide

73. Using Ruby to eval Objects Stored on Disk
    Allows state to leak in, makes computing stable hashes
    complicated because there are multiple valid representations to
    compute them off of.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 73
    

    View Slide

74. Not Storing all Podfile information in the Podfile.lock
    Figuring out when its changed based on only the info checked into
    git is nearly impossible.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 74
    

    View Slide

75. All-or-Nothing Installation Operation
    Needing to do the entire installation every time means all
    invocations are equally slow.
    Fixed now, thanks to @sebastianv1!
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 75
    

    View Slide

76. Pathname
    Nope. Just nope.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 76
    

    View Slide

77. Design Lessons Learned
    Architecture matters just as much as algorithms.
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 77
    

    View Slide

78. Samuel Giddins
    @segiddins
    Making CocoaPods Fast – Samuel Giddins @ Ruby on Ice 2019 78
    

    View Slide