Ruby concurrency is hard: how I became a Ruby on Rails contributor

For the past several weeks, I've been trying to fix a cranky spec in Karafka integrations suite, which in the end, lead me to become a Ruby on Rails micro-contributor and submitting similar fix to several other high-popularity projects from the Ruby ecosystem. Here's my story of trying to make sense of my specs and Ruby concurrency.

Ephemeral bug from a test-suite

Karafka is a Ruby and Rails multi-threaded efficient Kafka processing framework. To provide reliable OSS that is multi-threaded, I had to have the option to run my test suite concurrently to simulate how Karafka operates. Since it was a specific use case, I created my micro-framework.

Long story short: It runs end-to-end integration specs by running them in separate Ruby processes. Each starts Karafka, runs all the code in various configurations, connects to Kafka, checks assertions, and at the end, shuts down.

Such an approach allowed me to ensure that the process's whole lifecycle and its components work as expected. Specs are started with supervision, so in case of any hang, it will be killed after 5 minutes.

Karafka itself also has an internal shutdown supervisor. In case of a user shutdown request, if the shutdown takes longer than the defined expected time, Karafka will stop despite having things running. And this is what was happening with this single spec:

E, [2022-11-19T16:47:49.602718 #14843] ERROR -- : Forceful Karafka server stop
F, [2022-11-19T16:47:49.602825 #14843] FATAL -- : #<Karafka::Core::Monitoring::Event:0x0000562932d752b0 @id="error.occurred", @payload={:caller=>Karafka::Server, :error=>#<Karafka::Errors::ForcefulShutdownError: Karafka::Errors::ForcefulShutdownError>, :type=>"app.stopping.error"}>

This damn spec did not want to stop!

Many things are working under the hood:

  • workers that process jobs that could hang and force the process to wait
  • jobs queue that is also connected to the polling thread (to poll more data when no work is to be done)
  • listeners that poll data from Kafka that could hang
  • consumer groups with several threads polling Kafka data that might get stuck because of some underlying error
  • Other bugs in the coordination of work and states.

One thing that certainly worked was the process supervision that would forcefully kill it after 30 seconds.

Process shutdown coordination

The graceful shutdown of such a process takes work. When you have many connections to Kafka, upon a poorly organized shutdown, you may trigger several rebalances that may cause short-lived topics assignments causing nothing except friction and potentially blocking the whole process.

To mitigate this, Karafka shuts down actively and gracefully. That is, until the absolute end, it claims the ownership of given topics and partitions, actively waiting for all the current work to be finished. This looks more or less like so:

Note: Consumer groups internally in Karafka are a bit different than Kafka consumer groups. Here we focus on internal Karafka concepts.

Pinpointing the issue

After several failed attempts and fixing other bugs, I added a lot of extra instrumentation to check what Karafka hangs on. It was hanging because there were hanging listener threads!

As stated above, to close Karafka gracefully, all work from the jobs queue needs to be finished, and listeners that poll data from Kafka need to be able to exit the polling loops. It's all coordinated using a job queue. The job queue we're using is pretty complex with some blocking capabilities, and you can read about it here, but the interesting part of the code can be reduced to this:

@semaphores = { |h, k| h[k] = }

Those queues are used as semaphores in the polling loops until all the current work is done. Since each Queue is assigned to a different subscription group within its thread and hidden behind a concurrent map, there should be no problem. Right?


Once I had my crazy suspicion, I decided to reduce it down to a proof of concept:

require 'concurrent-ruby'

100.times do
  ids =
  semaphores = { |h, k| h[k] = } do do
      ids << semaphores['test'].object_id

  raise "I expected 1 semaphore but got #{ids.size}" if ids.size != 1

once executed, boom:

poc.rb:13:in `<main>': I expected 1 semaphore but got 2 (RuntimeError)

There is more than one semaphore for one listener! This caused Karafka to wait until forced to stop because the worker thread would use a different semaphore than the listener thread.

But how is that even possible?

Well, Concurrent::Hash and Concurrent::Map initialization is indeed thread-safe but not precisely as you would expect them to be. The docs state that:

This version locks against the object itself for every method call, ensuring only one thread can be reading or writing at a time. This includes iteration methods like #each, which takes the lock repeatedly when reading an item.

"only one thread can be reading or writing at a time". However, we are doing both at different times. Our code:

semaphores = { |h, k| h[k] = }

is actually equivalent to:

semaphores = do |h, k|
  queue =
  h[k] = queue

and the block content is not locked fully. One threads queue can overwrite the other if the Ruby scheduler stops the execution in the middle. Here's the flow of things happening in the form of a diagram:

Once in a while listener would receive a dangling queue object, effectively blocking the polling process.

Fixing the issue

This can be fixed either by replacing the Concurrent::Hash with Concurrent::Map and using the #compute_if_absent method or by introducing a lock inside of the Concurrent::Hash initialization block: do |k, v|
  k.compute_if_absent(v) { [] }

mutex = do |k, v|
  mutex.synchronize do
    break k[v] if k.key?(v)
    k[v] = []

Okay, but what does Ruby on Rails and other projects do with all of this?

Fixing the world

I figured out that if I made this mistake, maybe others did. I decided to check my local gems to find occurrences quickly. Inside my local gem cache, I executed the following code:

fgrep -R ' {' ./
fgrep -R ' do' ./
fgrep -R ' {' ./
fgrep -R ' do' ./

and validated that I'm not an isolated case. I wasn't alone!

Then using Sourcegraph I pinpointed a few projects that had the potential for fixes:

  • rails (activesupport and actionview)
  • i18n
  • dry-schema
  • finite_machine
  • graphql-ruby
  • rom-factory
  • apache whimsy
  • krane
  • puppet

I am not a domain expert in any of those, and understanding the severity of each was beyond my time constraints, but I decided to give it a shot.

Rails (ActiveSupport and ActionView)

Within Rails, this "pattern" was used twice: in ActiveSupport and ActionView.

In ActionView, it was used within a cache:

  h[k] =

and assuming that the cached result is stateless (same result each time for the same key), the issue could only cause an extra computation upon first parallel requests to this cache.

In the case of ActiveSupport, none of the concurrency code was needed, so I just replaced it with a simple Hash.

Both, luckily, were not that severe, though worth fixing nonetheless.


Both were merged, and this is how I became a Ruby on Rails contributor :)


This case was slightly more interesting because the concurrent cache stores all translations. In theory, this could cause similar leakage as in Karafka, effectively losing a language by loading it to a different Concurrent::Hash: do do
    I18n.backend.store_translations(rand.to_s, :foo => { :bar => 'bar', :baz => 'baz' })

I18n.available_locales.count #=> 1

This could lead to hard-to-debug problems. Once in a while, your system could raise something like this:

:en is not a valid locale (I18n::InvalidLocale)

without an apparent reason, and this problem would go away after a restart.



Another cache case where the risk would revolve around double-computing.



This one is interesting! Let's reduce the code to a smaller POC first and see what will happen under heavy threading:

require 'singleton'
require 'concurrent-ruby'

class Sequences
  include Singleton

  attr_reader :registry

  def initialize

  def next(key)
    registry[key] += 1

  def reset
    @registry = { |h, k| h[k] = 0 }

seq = Sequences.instance

loop do do {'boom') }

  size = seq.registry['boom']

  raise "Wanted 100 but got #{size}" unless size == 100

poc.rb:37:in `block in <main>': Wanted 100 but got 1 (RuntimeError)

The counter value gets biased. What is even more interesting is that making the map safe won't be enough:

@registry = { |h, k| h.compute_if_absent(k) { 0 } }
poc.rb:36:in `block in <main>': Wanted 100 but got 55 (RuntimeError)

there is one more "unsafe" method:

def next(key)
  registry[key] += 1

this operation also is not atomic, thus needs to be wrapped with a mutex:

def initialize
  @mutex =

def next(key)
  @mutex.synchronize do
    registry[key] += 1

Only then is this code safe to be used.

Other repositories


In my opinion, there are a few outcomes of this story:

  • Karafka has a solid test-suite!
  • If you are doing concurrency-related work, you better test it in a multi-threaded environment and test it well.
  • Concurrency is hard to many of us (maybe that's because we are special ;) ).
  • RTFM and read it well :)
  • Do not be afraid to help others by submitting pull requests!

On the other hand, looking at the frequency of this issue, it may be worth opening a discussion about changing this behavior and making the initialization fully locked.


Concurrent::Hash under cRuby is just a Hash. You can check it out here.

Cover photo by James Broad on Attribution-NonCommercial-ShareAlike 2.0 Generic (CC BY-NC-SA 2.0). Image has been cropped.

RSpec story about disappearing classes

ActiveSupport#descendants can be slow. In a bigger system with layers of descendants, finding all of them can be time-consuming:

puts Benchmark.measure do
  100.times { Dispatchers::Base.descendants }

# 5.235370   0.015754   5.251124 (  5.251069)

In the code I've been working on, it meant that a single lookup was taking around 50ms. That is a lot, especially if used extensively.

To mitigate this, I've implemented a simple caching layer on top of the lookup that would make things fast:

module Mixins
  module CachedDescendants
    extend ActiveSupport::Concern

    cattr_accessor :descendants_map

    self.descendants_map =

    class << self
      # Clears the descendants map cache - can be hooked to Rails reloader
      def reload!

    included do
      class << self
        # @return [Array<Class>] array with descendants classes
        def cached_descendants
          ::Mixins::CachedDescendants.descendants_map[self] ||= descendants

When included and used, it would give great results:

puts Benchmark.measure do
  100.times { Dispatchers::Base.cached_descendants }

# 0.000023   0.000001   0.000024 (  0.000024)

99,99956% faster!

Such code, like any other, deserves to be tested. I wrote some specs for it, including a relatively simple one:

  context 'when there are two independent bases' do
    let(:base1) do do
        include ::Mixins::CachedDescendants

    let(:base2) do do
        include ::Mixins::CachedDescendants

    before do { } { }

    it 'expect for them not to interact' do
      expect(base1.cached_descendants.size).to eq(5)
      expect(base2.cached_descendants.size).to eq(5)
      expect(base1.cached_descendants & base2.cached_descendants).to be_empty

It would just ensure that the way we cache does not create collisions for independent descendants trees.

But once in a while, this code would randomly fail:

  1) Mixins::CachedDescendants when there are two independent bases expect for them not to interact
     Failure/Error: expect(base1.cached_descendants.size).to eq(4)
       expected: 5
            got: 4
       (compared using ==)

How can I create five classes and suddenly have only 4? I initially thought something was wrong with the descendants lookup for anonymous classes. However, this functionality is heavily used by many, including me, and it never created any problems. On top of that, why would it fail only once in a while?

When something fails randomly, it usually means that there's an external factor to it. One that operates under the hood. It wasn't different in this case.

After some investigation, I was able to reproduce it:


puts "Total classes before: #{ObjectSpace.count_objects[:T_CLASS]}" 
puts "String subclasses count before: #{String.subclasses.count}" 

100.times { }

puts "Total classes after defining: #{ObjectSpace.count_objects[:T_CLASS]}" 
puts "String subclasses count after defining: #{String.subclasses.count}" 


puts "Total classes after GC: #{ObjectSpace.count_objects[:T_CLASS]}" 
puts "String subclasses count after GC: #{String.subclasses.count}" 

# Total classes after defining: 1324
# String subclasses count after defining: 102
# Running GC...
# Total classes after GC: 1124
# String subclasses count after GC: 2

Boom! Anonymous classes are being garbage collected! Classes that are not referenced anywhere are subject to garbage collection like other objects, and this code was not memoizing them:

before do { } { }

Hence, the spec would fail if GC kicked in exactly between the classes definitions and the spec execution. This is why it would only fail once in a while.

Fixing such an issue required only minimal changes to the spec:

    let(:descendants) do
      # This needs to be memorized, otherwise Ruby GC may remove those in between assertions
      [ { }, { }

    before { descendants }

That way, the anonymous classes would be referenced throughout the lifetime of this spec.

Summary (TL;DR)

Anonymous classes and modules are a subject of garbage collection like any other object. Unless you reference them, they may be gone before wanting to use them via #descendants or similar lookups. Always reference them in some way or face the unexpected.


It was pointed to me by Jean Boussier, that there's an ActiveSupport::DescendantsTracker that can also be used to improve the lookup performance and that Ruby 3.1 has a #subclasses that also is much faster than iterating over ObjectSpace. While their performance is several times faster than the "old" lookup, it is still a magnitude slower than caching the descendants in dev but has similar performance in production.

0.002020 vs. 0.000023 in dev and 0.000039 vs. 0.000042 in production.

The best code is no code, so now I can deprecate my code. The strong class references, however, are still valid and worth keeping in mind.

Cover photo by Alper Orus on Attribution-NonCommercial-ShareAlike 2.0 Generic (CC BY-NC-SA 2.0). Image has been cropped.

Copyright © 2023 Closer to Code

Theme by Anders NorenUp ↑