Unpause FIFO queues on worker completion

[davidrichey]

• Updated Manager#processor_done to check for FIFO queue & unpause
  • Added specs
• Added unpause method on strategies
  • Added specs around #unpause

closes #643

[cjlarose]

Overall approach seems solid! I took another pass after reviewing some of the surrounding context.

[cjlarose]

For symmetry with the method messages_found, let's consider renaming this method message_processed.

The reason is that the interaction between the Manager class and the polling strategy is that the manager is just informing the polling strategy of certain events, allowing the polling strategy to respond as it sees fit. In this case, it's not the responsibility of the Manager to tell the polling strategy to pause a queue, it's just the responsibility for the Manager to let the polling strategy know that it's done processing the message: the polling strategy is welcome to do whatever it wants with that knowledge.

[cjlarose]

Let's also add message_processed to Shoryuken::Polling::BaseStrategy. Its implementation can be empty (for backward compatibility--users might have subclassed the base class themselves).

[cjlarose]

Since users can specify their own polling strategy class, we want to make it optional to implement the new method. Let's invoke this method only if @polling_strategy responds to it. Otherwise, this change would be backwards-incompatible.

[cjlarose]

Thanks for being so responsive to feedback! I'll do some testing on some real queues, but overall, this is looking great!

[cjlarose]

I tested this locally on some FIFO queues. On master, I observed that if concurrency is more than 1 and you have a delay configured, FIFO message throughput is very poor because the "fetcher" pauses every time it gets an empty receive.

On this branch, I observed that the fetcher resumed whenever a job is finished, improving throughput, especially when the number of distinct message_group_ids is small.

Excellent work, @davidrichey!