immediate switch

[leonoel]

The current behavior of ?< is to cancel the current subprocess when the input flow is ready to transfer, then wait for the subprocess to terminate, then transfer value from input flow and create a new subprocess. Consequently :

• any value produced by the subprocess after cancellation is passed to the main process. Especially, if the subprocess crashes on cancellation, the exception must be caught by the user otherwise the whole process crashes, and a meaningful value has to be produced instead which is not always possible (this problem can be mitigated with amb in ap but not in cp)
• the actual switch should be expected to be delayed if the subprocess doesn't terminate immediately.

Alternative design : ?< spawns a new subprocess immediately when the input flow is ready to transfer. The previous subprocess is detached from the main process, it is cancelled and flushed (ie consumed as fast as possible, values are discarded and exceptions ignored). This implies the switch parent process must keep track of all of its concurrent stale subprocesses and wait for all of them to terminate before terminating itself.

Prior art : unknown. Rx has switchmap, but this subtlety is irrelevant here due to lack of graceful shutdown - each process is considered terminated immediately after cancellation.

[dustingetz]

(ns dustin.y2022.missionary-switch
  (:require [missionary.core :as m]
            [hyperfiddle.rcf :as rcf :refer [tests ! %]])
  (:import [missionary Cancelled]))

(hyperfiddle.rcf/enable!)

(tests

  (def >x (m/ap
            (let [>a (m/?> (m/seed (range 5)))]
              (m/?> (m/seed (range >a))))))

  (def cancel ((m/reduce conj >x) ! !))
  % := [0
        0 1
        0 1 2
        0 1 2 3])

(tests

  (def >x (m/ap
            (try
              (let [>a (m/?< (m/seed (range 5)))]
                (m/?< (m/seed (range >a))))
              (catch Cancelled _))))

  (def cancel ((m/reduce conj >x) ! !))
  ; in current design of switch, we can see the terminal emission during cancellation
  % := [nil 
        nil 
        nil
        0
        1
        2
        3])

[dustingetz]

TLDR: two problems with current design:

• userland has to handle the terminal emission during cancellation
• if the cancellation is not immediate, the switch is delayed

[leonoel]

Released in b.27