I learned that ActiveRecord would roll back transactions when inside an aborted thread. It’s implemented right here, in ActiveRecord’s connection_adapters/abstract/transaction.rb:

def within_new_transaction
  # ...
ensure
  # ...
  elsif Thread.current.status == "aborting" || (!completed && transaction.written)
    # The transaction is still open but the block returned earlier.
    #
    # The block could return early because of a timeout or because the thread is aborting,
    # so we are rolling back to make sure the timeout didn't caused the transaction to be
    # committed incompletely.
    rollback_transaction
  # ...
end

How did I end up here? I recently implemented a feature in GoodJob to track active processes in the database using ActiveRecord. These database queries occur in a background thread where I also use the same database connection for GoodJob’s Postgres LISTENing. It looks something like this (pseudocode):

Concurrent::Future.execute do
  process = Process.create
  loop { listen_for_notify }
ensure
  process&.destroy!
end

While working on one of my Rails projects that use GoodJob (https://dayoftheshirt.com), I noticed that Process records were not cleaned up as I expected running when GoodJob async inside of Rails Server/Puma locally and exiting. Inspecting the logs, I saw this:

TRANSACTION (0.1ms)  BEGIN
GoodJob::Process Destroy (0.3ms)  DELETE FROM "good_job_processes" WHERE "good_job_processes"."id" = $1  [["id", "320de861-4c84-4f8c-ba3e-2e08a8ef0469"]]
TRANSACTION (0.1ms)  ROLLBACK

Huh. I did some Googling and found some references to rollback behavior, and asked in the Rails Link Slack, but nobody knew. This caused me to dig in ActiveRecord, where I found the behavior implemented.

I was still confused because I had never seen this behavior in GoodJob when building out the feature. But after much head-scratching, I realized that I hadn’t followed GoodJob’s README instructions for integrating with Puma, which ensures that GoodJob is gracefully shutdown before Ruby aborts threads at exit:

# config/puma.rb

on_worker_shutdown do
  GoodJob.shutdown
end

MAIN_PID = Process.pid
at_exit do
  GoodJob.shutdown if Process.pid == MAIN_PID
end

Mystery solved.