TurboTest

Proposal

Discourse contains two tools (highly recommend you try them out for context)

• bin/turbo_rspec a parallel runner with non interleaving forking model. (leverages some of parallel_test … rake parallel:create / migrate to prep the dbs.
• bin/rake autospec an automatic spec runner, which focuses on failed specs, like guard except that unlike guard it is interruptible.

Enter turbo_test (name TBD, suggestions welcome).

turbo_test will be a slot in replacement for the 2 tools Discourse use in a dedicated gem.

turbo_test is to first be fully functional with rspec runners but longer term should also work with minitest.

Features of the turbo_test gem:

• MIT license, standard Rails code of conduct
• Pull based model for forked test runners. Master process manages a queue of tests, forked processes pull from the queue. Transport long term is agnostic though I would recommend initial implementation uses pipes.
• Pull model ensures that all the workers are running tests at all times. The parallel_test model of splitting up the tests upfront means that workers are often idle for long periods of time.
• Rake tasks for administration of partitioned test environments turbo_test:create / migrate / drop
• Non interleaved results (like bin/turbo_rspec)
• While turbo_rspec is running, if you hit a specific key you can see right away information about the current tests that failed without halting the test process.
• Documentation about how to handle custom sharding (memcached / redis) and so on.
• Minimal changes required to Rails projects that decide to use this.
• A key goal is deprecation of bin/turbo_rspec in Discourse.
• Minimal dependencies (no explicit rspec, minitest, redis, pg dependencies)
• Stretch goal, once this is all done … extract bin/rake autospec into this gem as well. (I will do a mini specification if we get there)
• Extra long term stretch goal, pull these concepts back into Rails proper.

Why not guard?

Guard at the moment does not support interruptible tests, this is a must have feature for turbo_test.

Why not parallel_test?

It does not support a pull model so it would be close to a ground up re-write.