Use of agents to complete side-effects in STM transactions

Question

I'm aware that it is generally bad practice to put functions with side-effects within STM transactions, as they can potentially be retried and called multiple times.

It occurs to me however that you could use agents to ensure that the side effects get executed only after the transaction successfully completes.

e.g.

(dosync
  // transactional stuff
  (send some-agent #(function-with-side-effects params))
  // more transactional stuff
  )

Is this good practice?

What are the pros/cons/pitfalls?

Answer 1

Original:

Seems like that should work to me. Depending on what your side effects are, you might want to use send-off (for IO-bound ops) instead of send (for cpu-bound ops). The send/send-off will enqueue the task into one of the internal agent executor pools (there is a fixed size pool for cpu and unbounded size pool for io ops). Once the task is enqueued, the work is off the dosync's thread so you're disconnected at that point.

You'll need to capture any values you need from within the transaction into the sent function of course. And you need to deal with that send possibly occurring multiple times due to retries.

Update (see comments):

Agent sends within the ref's transaction are held until the ref transaction successfully completes and are executed once. So in my answer above, the send will NOT occur multiple times, however it won't occur during the ref transaction which may not be what you want (if you expect to log or do side-effecty stuff).

Answer 2

This works and is common practice. However, like Alex rightly pointed out you should consider send-off over send.

There are more ways to capture commited-values and hand them out of the transaction. For example you can return them in a vector (or a map or whatever).

(let [[x y z] (dosync
                ; do stuff
                [@x @y @z])] ; values of interest to sode effects
  (side-effect x y z))

or you can call reset! on a local atom (defined outside the lexical scope of the dosync block of course).

Answer 3

There's nothing wrong with using agents, but simply returning from the transaction values needed for the side-effecting computation is often sufficient.

Refs are probably the cleanest way to do this, but you can even manage it with just atoms!

(def work-queue-size (atom [0]))

(defn add-job [thunk]
  (let [[running accepted?]
        (swap! work-queue-size
               (fn [[active]]
                 (if (< active 3)
                   [(inc active) true]
                   [active false])))]
    (println
     (str "Your job has been "
          (if accepted?
            "queued, and there are "
            "rejected - there are already ")
          running
          " total running jobs"))))

The swap! can retry as many times as needed, but the work queue will never get larger than three, and you will always print exactly once a message that is tied correctly to the acceptance of your work item. The "original design" called for just a single int in the atom, but you can turn it into a pair in order to pass interesting data back out of the computation.