Improve performance of enqueueing tasks

[pior]

Context

Before enqueueing every task, the name of the queue is published to a Redis Set key (base.AllQueues).
This is only used by the inspector for inspection/monitoring purposes (not for the processing of tasks).

Two consequences:

• Enqueuing one task requires two round-trip to Redis
• The same Redis set key is updated constantly, causing significant Redis load

The issue #549 has more details about the impact on performance.
We've confirmed the figures in production.

Fixes #549
Rework of #550

Changes

• Update the RDB client to only write to base.AllQueues once per worker

Notes:

• This light implementation does not protect against race-conditions, which is not an issue.
• Writing to base.AllQueues will still be retried until it worked.

[kamikazechaser]

If the queue is removed via RemoveQueue, it will not be added back on the next enqueue because it will still exist in the map. We need to also remove it from the map on any SREM("asynq::queues", qname)

[pior]

If the queue is removed via RemoveQueue, it will not be added back on the next enqueue because it will still exist in the map. We need to also remove it from the map on any SREM("asynq::queues", qname)

Right.
Removing from the queue is a nice, simple solution.

But is it enough to remove the queue from the map "locally" (in the same process).
It works if a task is enqueued to the queue in the same process, but won't work for a different process.

Are we ok with this tradeoff?

(Implementation and tests updated)

[kamikazechaser]

It works if a task is enqueued to the queue in the same process, but won't work for a different process.

That's right. I think the tradeoff is fair.