nvie / new_workers

Be warned: this is very experimental!

Contributing

The code needs some cleanup, and might not be obvious at all. All of this is very much the result of myself hacking around in it for way too long. Please do add GitHub comments for lines that are unclear, and I'll explain they're in there. This bit of communication will be vital in developing a common understanding of the new model.

Any help with this is really appreciated!

To run the workers (all executing fake work for now):

$ python main.py forking 4  # will start 4 pre-forked child processes
$ python main.py gevent 10  # will use 10 greenlets

The New and Shiny RQ Concurrent Workers

This is a temporary project that I use to flesh out the bodies of the main new worker structure for concurrent RQ workers. The main reason for putting this in a separate repository is to fight complexity without the burden of the existing RQ code base. The new concurrent worker structure is complex enough on its own, without any RQ details.

This project contains no notion of "Redis", "jobs", "queues", or whatever RQ's jargon. What it does concern are "workers" and "work horses", or as they're called in the new structure "children".

The new worker structure differs a lot from the current, non-concurrent, worker structure, which I'll explain below.

The Status Quo

This is how RQ versions before 0.4 work.

The work loop:

• At the top, there's the worker process (a single Unix process)
• This worker connects to Redis and listens for incoming work on queues using the blocking BLPOP call
• If such a job is received, it forks and lets its forked child (called the "work horse") execute the work
• It waits until the work horse is done
• It loops

Hence, at most 2 Unix processes will exist at the same time. There is no concurrency control in RQ workers whatsoever. Processing more jobs concurrently is possible only by starting more workers.

The termination logic is as follow:

• Either SIGINT ("Ctrl+C") or SIGTERM terminates the loop gracefully. The main worker catches these and activates abortion: it sets a flag to stop the loop after the current work horse is done
• The work horse ignores any signal and always continues its work
• If a second SIGINT or SIGTERM is fired during this waiting period, cold shutdown kicks in, meaning the main worker will forcefully kill the work horse
• A cold shutdown means work is brutefully terminated and work is lost

New worker structure

New terminology: In the new situation, the term "work horse" will be avoided. Instead, we'll refer to the unit of concurrency that performs the actual work as the "worker child", or simply "child", whether that'd be a process, a thread, or a greenlet.

Changes from 10,000 ft.:

• RQ workers are concurrent, meaning a single worker can run multiple jobs at the same time, without help from another worker.

• Multiple concurrency mechanics can be used, like the good old forking, multiprocessing, threading and gevent-style cooperative multitasking (pick whatever suits you best)

• The forking worker will still be the default, and still be 1 child by default (same as pre-0.4). Concurrency should be enabled explicitly.

• Abortion semantics and behaviour should not change from a user's perspective:

  • A single SIGINT or SIGTERM should activate warm shutdown (i.e. wait indefinitely for all current work to finish, but then stop)
  • While that waiting period is active, a second SIGINT or SIGTERM will have the main worker brutally kill every child, losing work
  • This behaviour is predictable and exactly the same for all concurrency mechanisms, no matter if forking, multiprocessing, threading, or gevent is used.

Under the hood

Most of the complexity is introduced by the very last requirement. I've often stopped and asked myself whether this is the way to go, as each concurrency mechanism has roughly different semantics that might not fit this use case very well. Nevertheless, I could not find piece of mind with workers behaving very differently depending on the concurrency backend that was used---it had weird behavioural differences. This is the main reason why the GitHub issue for concurrency has been open for over 10 months now.

Let's take a look at the main worker loop. Compared to the pre-0.4 state of affairs, the main worker loop has become simpler from a high-level perspective, as most of the job's details will now be delegated to the children.

The main worker loops indefinitely, spawning children (of the currently used backend: so either processes, threads or greenlets) from a "pool-like" object that either (1) returns a new spawned child, or (2) blocks until there's room for such a child in the pool. Since this is, by definition, an eventually blocking call, this can be called in an endless loop.

This behaviour is exactly modeled after gevent.pool.Pool's behaviour, which is super handy. Starting a worker now means children are spawned upfront and are able to immediately execute work. The number of spawned children is maximal right after the main worker is started, all waiting for work individually. In other words: it is not the case that children are spawned as there is work for them.

Each spawned child then does the following. They connect to Redis and each perform a BLPOP call. When they receive work, they execute it (nothing differs from the pre-0.4 situation from here forward).

This, too, is a simplification, but one thing becomes more complex: handling worker termination with warm and cold shutdowns. Here's why.

Basically, the problem is to know for sure when a worker is safe to terminate. Before 0.4, the main worker BLPOP'ed work from the queue first, then spawned a work horse. It's easy then: the work horse should never be terminated: it is always doing important stuff.

In the new situation, though, the child has two "stages": (1) it waits (blocks) in BLPOP until work is retrieved from Redis, and (2) it executes work. It's clear that children in stage (1) should be terminated when warm shutdown is requested, while children in stage (2) must not be terminated. Making this distinction means children need to communicate to the main worker about their idle state. (By the way, note that this is different from the job states---don't confuse them.)

To enable this nuance, we need a few extra data structures to create an "idle state" administration. The exact data structures to use depend heavily on the concurrency implementation used.