Add broker_default with nil

[skunkworker]

Summary

Adds a first party queue_type setting so queues can be declared as classic, quorum, or stream — or, by default, with no opinion at all so the broker's own default_queue_type applies.

Configurable globally or per route:

::ActionSubscriber.configure do |config|
  config.queue_type = :quorum
end

::ActionSubscriber.draw_routes do
  route UserSubscriber, :created, :queue_type => :quorum
  route AuditSubscriber, :created, :queue_type => :broker_default
end

Value	x-queue-type sent
nil (default), or :broker_default	not sent — broker applies its own default_queue_type
:classic	classic
:quorum	quorum
:stream	stream

nil is the canonical "defer to the broker" value; :broker_default is accepted as a more readable spelling and normalizes to nil. Values are normalized on assignment, so a typo raises an ArgumentError where it was set rather than at route-draw time or inside MessageRetry at runtime. :quorum and :stream force durable => true, since RabbitMQ only supports those as durable queues.

The bug this uncovered

The two drivers have been disagreeing about queue type this whole time.

# march_hare 4.8:  @options.fetch(:type, args.fetch(QUEUE_TYPE, Types::CLASSIC)).to_s
# bunny 2.24:      @options[:type]

fetch only falls back when the key is absent. Since setup_queue never passed :type, JRuby has been declaring every queue with x-queue-type: classic while MRI sent no argument at all. Verified against a real MarchHare::Queue:

{}                            -> args={"x-queue-type" => "classic"}
{type: nil}                   -> args={}

Passing :type => nil explicitly — key present, value nil — is what suppresses the argument. That is the whole mechanism behind :broker_default, and it's why the option has to reach the driver as an explicit nil rather than by omitting it.

Worth noting for reviewers: you cannot send x-queue-type with a null value as a middle ground. amq-protocol will encode it as AMQP void ("\x00\x00\x00\x0E\fx-queue-typeV"), but it can't decode its own output, and RabbitMQ resolves the type via rabbit_queue_type:discover/1, which matches no known type for a void and fails the declare. Omitting the key is the only way to defer to the broker.

⚠️ Breaking on JRuby

MRI behavior is unchanged. On JRuby, newly declared queues change from classic to whatever the broker defaults to.

This is not limited to new queues. Because queue type is fixed at declaration, an existing classic queue is now redeclared without x-queue-type. That's harmless on a vhost whose default_queue_type is classic, since the broker resolves to the same type — but it fails with PRECONDITION_FAILED on a vhost defaulting to quorum or stream.

Before rolling this out to a JRuby deployment, audit every vhost it connects to:

rabbitmqctl list_vhosts name default_queue_type

If any are non-classic, set config.queue_type = :classic before deploying — that restores the previous JRuby behavior exactly.

Known limitation

MessageRetry declares its *.retry_* queues from the global config.queue_type, not the type of the route that produced the message, so a route-level type is not propagated to its retry queue. Documented in the README rather than fixed here — propagating it isn't a straight pass-through, since retry queues carry x-message-ttl and x-dead-letter-exchange and streams support neither. Happy to take it in a follow-up.

Testing

Full suite green on JRuby 10.0.4 against a live broker: 130 examples, 0 failures (1 pending is pre-existing). New coverage in spec/lib/action_subscriber/queue_type_spec.rb includes JRuby-only examples that assert the actual MarchHare::Queue argument table for each option — including one pinning the surprising omitted :type -> classic behavior, so a driver change that alters it fails loudly.

Note on commit scope

This branch carries two commits. d08277b (Appraisal matrix, CI split by Rails version, RabbitMQ spec helper) is largely pre-existing work that was already in the working tree; the only new part is giving Rails 8.0/8.1 their own CI matrix instead of exclude entries against the Ruby 3.1-class images — job coverage is unchanged at 20. 1a348ec is the queue_type feature. Review the second commit for the substance here.

🤖 Generated with Claude Code