A while back, we wrote about How to Run RSpec Test Suites in Parallel with Jenkins CI Build Flow. A version of that still handles our largest test suite, but over time the initial division of specs became unbalanced. We ended up with some tasks that took twice as long as others. Even worse, in an attempt to rebalance task times, we ended up with awkward file patterns like 'spec/api/**/[a-m]*_spec.rb'.

To keep our parallel spec tasks approximately equal in size and to support arbitrary concurrency, we've added a new spec:sliced task:

1
2
3
4
5
6
7
8
9
10
11
12
namespace :spec do
  task :set_up_spec_files do
    spec_files = Dir['spec/**/*_spec.rb']
    @spec_file_digests = Hash[spec_files.map { |f| [f, Zlib.crc32(f)] }]
  end

  RSpec::Core::RakeTask.new(:sliced, [:index, :concurrency] => :set_up_spec_files) do |t, args|
    index = args[:index].to_i
    concurrency = args[:concurrency].to_i
    t.pattern = @spec_file_digests.select { |f, d| d % concurrency == index }.keys
  end
end

As you can see, the set_up_spec_files helper task builds a hash of spec file paths and corresponding checksums. When we invoke the sliced task with index and concurrency values (e.g., 0 and 5), only the spec files with checksums equal to 0 when mod-ed by 5 are run. Thus, the Jenkins build flow would look like:

1
2
3
4
5
6
7
8
parallel (
  {build("master-ci-task", tasks: "spec:sliced[0,5]")},
  {build("master-ci-task", tasks: "spec:sliced[1,5]")},
  {build("master-ci-task", tasks: "spec:sliced[2,5]")},
  {build("master-ci-task", tasks: "spec:sliced[3,5]")},
  {build("master-ci-task", tasks: "spec:sliced[4,5]")}
)
build("master-ci-succeeded")

Now, spec times might continue to be unbalanced despite files being split up approximately evenly. (For a more thorough approach based on recording spec times, see knapsack.) However, this little bit of randomness was a big improvement over our previous approach, and promises to scale in a uniform manner.

Categories: continuous integration, rspec, ruby, testing


Comments