4

这篇笔记主要是基于文档展开一下 core.async 在 ClojureScript 当中的基本用法.
具体的内容可以看原文章, 已经比较详细了, 很多在 API 文档的 demo 当中.
关于基础知识跟 cljs 跟 clj 的区别, 这篇文章就不涉及了.

之前用到 core.async , 发现自己中间很多理解缺失了, 趁有时间赶紧看一下.
从 API 文档可以看到 core.async 的函数语义是比较丰富的, 几十个函数,
我顺着看了一圈, 整理下来大致是几个功能, 大致分成几块.

  • 多对一

    • 类似 Promise.all
    • 多个 channel 的数据进行 map
    • 通过 merge 合并多个 channel 到一个
    • 通过 mix 控制多个 channel 具体合并/解开的情况
  • 多选一

    • alts! 的多选一, 对应 Promise.race
    • alt! 的语法套路
  • 一拆二/过滤

    • split 直接拆成两个
    • 用 pipeline 搭配 filter 进行过滤
    • 用 transducer 写法进行过滤
  • 一对多

    • 通过 mult 发送给多个接收端

本来我的触发点是想看 core.async 是都能对应到 Promise 常用的功能,
这样看下来, core.async 功能是过于丰富了, 反而有些用不上.
由于我对 Go 熟悉度有限, 不好跟 Go Channel 做对比了.

后面逐个看一下示例. 为了方便演示, 我增加了两个辅助函数,

  • fake-task-chan, 生成一个随机定时任务, 返回一个 channel
  • display-all, 打印一个 channel 所有返回的非 nil 数据, nil 表示结束.

多对一

类似 Promise.all

(defn demo-all []
  (go
         ; 首先 tasks 得到向量, 内部函数多个 channel
   (let [tasks (->> (range 10)
                    (map (fn [x] (fake-task-chan (str "rand task " x) 10 x))))]
     (println "result"
        ; loop 函数逐个取 tasks 的值, 从头取, 一次次 recur, 直到取完, 结束
        (loop [acc [], xs tasks]
          (if (empty? xs)
              acc
                               ; <! 表示从 channel 取数据, 在 go block 内阻塞逻辑
              (recur (conj acc (<! (first xs)))
                     (rest xs))))))))

由于任务在 loop 之前已经开始了, 类似 Promise.all 的效果.
一个个等待结果, 最终就是全部的值, 耗时就是最长的那个等待的时间.

=>> node target/server.js
rand task 0 will take 0 secs
rand task 1 will take 1 secs
rand task 2 will take 2 secs
rand task 3 will take 3 secs
rand task 4 will take 9 secs
rand task 5 will take 6 secs
rand task 6 will take 7 secs
rand task 7 will take 1 secs
rand task 8 will take 2 secs
rand task 9 will take 9 secs
rand task 0 finished
rand task 1 finished
rand task 7 finished
rand task 2 finished
rand task 8 finished
rand task 3 finished
rand task 5 finished
rand task 6 finished
rand task 4 finished
rand task 9 finished
result [0 1 2 3 4 5 6 7 8 9]

可以看到最终以数组形式返回了每个 channel 返回的数据了.

多个 channel 的数据进行 map

我其实不大清楚这个 map 用在什么样的场景, 就是取两个 channel 计算得到新的数字.

(defn demo-map []
  (let [<c1 (to-chan! (range 10))
        <c2 (to-chan! (range 100 120))
        <c3 (async/map + [<c1 <c2])]
    (display-all <c3)))

所以就是 0 + 100 1 + 101... 得到 10 个数据

=>> node target/server.js
100
102
104
106
108
110
112
114
116
118
nil

总体上还是多个 channel 合并成一个了.

通过 merge 合并多个 channel 到一个

merge 就是把多个 channel 的数据合并到一个, 字面意义的意思.
从得到的新的 channel, 可以获取到原来 channel 的数据.

(defn demo-merge []
  (let [<c1 (chan),
        <c2 (chan),
        <c3 (async/merge [<c1 <c2])]
    (go (>! <c1 "a") (>! <c2 "b"))
    (display-all <c3)))

所以从 c3 就能拿到写到原来的两个 channel 的数据了,

=>> node target/server.js
a
b

通过 mix 控制多个 channel 具体合并/解开的情况

mix 跟 merge 很相似, 区别是中间多了一个控制层, 定义成 mix-out,
通过 admix unmix 两个函数可以调整 mix-out 上的关系,
这个例子当中

(defn demo-mix []
  (let [<c0 (chan)
        <c1 (async/to-chan! (range 40))
        <c2 (async/to-chan! (range 100 140))
        mix-out (async/mix <c0)]
    ; mix 过来两个 channel
    (async/admix mix-out <c1)
    (async/admix mix-out <c2)
    (go
               ; 先取 20 个数据打印
     (doseq [x (range 20)] (println "loop1" (<! <c0)))
     (println "removing c2")
     ; 去掉那个数字特别大的 channel
     (async/unmix mix-out <c2)
               ; 再取 20 个数据打印
     (doseq [x (range 20)] (println "loop2" (<! <c0))))))

得到结果,

=>> node target/server.js
loop1 0
loop1 100
loop1 1
loop1 101
loop1 2
loop1 102
loop1 3
loop1 103
loop1 104
loop1 4
loop1 105
loop1 5
loop1 106
loop1 6
loop1 107
loop1 108
loop1 109
loop1 110
loop1 7
loop1 8
removing c2
loop2 111
loop2 9
loop2 10
loop2 11
loop2 12
loop2 13
loop2 14
loop2 15
loop2 16
loop2 17
loop2 18
loop2 19
loop2 20
loop2 21
loop2 22
loop2 23
loop2 24
loop2 25
loop2 26
loop2 27

可以看到刚开始的时候, 从返回的 channel 可以获取到两个来源 channel 的数据,
进行一次 unmix 之后, 大数的来源不见了, 后面基本上是小的数字.

这个顺序看上去是有一些随机性的, 甚至 unmix 还有一次大数的打印, 后面稳定了.
注意 mix-out 只是用于控制, 获取数据在代码里还是要通过 c0 获取的.

多选一

alts! 的多选一, 对应 Promise.race

这个比较清晰的

(defn demo-alts []
  (go
   (let [<search (fake-task-chan "searching" 20 "searched x")
         <cache (fake-task-chan "looking cache" 15 "cached y")
         <wait (fake-task-chan "timeout" 15 nil)
                       ; 数组里边三个可选的 channel
         [v ch] (alts! [<cache <search <wait])]
     (if (= ch <wait ) (println "final: timeout")
                       (println "get result:" v)))))

就是随机的时间, 取返回最快的结果. 我多跑几次

=>> node target/server.js
searching will take 3 secs
looking cache will take 14 secs
timeout will take 9 secs
searching finished
get result: searched x
^C
=>> node target/server.js
searching will take 10 secs
looking cache will take 1 secs
timeout will take 4 secs
looking cache finished
get result: cached y
timeout finished
searching finished
^C
=>> node target/server.js
searching will take 19 secs
looking cache will take 4 secs
timeout will take 1 secs
timeout finished
final: timeout
looking cache finished
^C
=>> node target/server.js
searching will take 0 secs
looking cache will take 6 secs
timeout will take 1 secs
searching finished
get result: searched x
timeout finished
looking cache finished
^C

可以看到打印的结果都是最短时间结束的任务对应的返回值.
timeout 是这种情况当中比较常用的一个定时器, 控制超时.

alt! 的语法套路

alt! 跟 alts! 就是类似了, 主要是语法比较丰富一点,

(defn demo-alt-syntax []
  (let [<search1 (fake-task-chan "search1" 10 "search1 found x1")
        <search2 (fake-task-chan "search2" 10 "search2 found x2")
        <log (chan)
        <wait (fake-task-chan "timeout" 10 nil)]
    (go
     (loop []
       (let [t (rand-int 10)]
         (println "read log waits" t)
         (<! (timeout (* 1000 t)))
         (println "got log" (<! <log))
         (recur))))
    (go
     (println "result"
       (async/alt!
         ; 匹配单个 channel 的写法
         <wait :timeout
         ; 这个是往 channel 发送消息的写法, 发送也是等待对方读取, 也受时间影响
         ;  这个两层数组是挺邪乎的写法...
         [[<log :message]] :sent-log
         ; 这个匹配向量包裹的多个 channel, 后面通过 ch 可以区分命中的 channel
         [<search1 <search2] ([v ch] (do (println "got" v "from" ch)
                                         :hit-search)))))))

直接多跑几次了, 效果跟上边一个差不多的,

=>> node target/server.js
search1 will take 3 secs
search2 will take 7 secs
timeout will take 3 secs
read log waits 8
search1 finished
timeout finished
got search1 found x1 from #object[cljs.core.async.impl.channels.ManyToManyChannel]
result :hit-search
search2 finished
^C
=>> node target/server.js
search1 will take 2 secs
search2 will take 0 secs
timeout will take 4 secs
read log waits 2
search2 finished
got search2 found x2 from #object[cljs.core.async.impl.channels.ManyToManyChannel]
result :hit-search
search1 finished
timeout finished
^C
=>> node target/server.js
search1 will take 9 secs
search2 will take 0 secs
timeout will take 9 secs
read log waits 6
search2 finished
got search2 found x2 from #object[cljs.core.async.impl.channels.ManyToManyChannel]
result :hit-search
search1 finished
timeout finished
^C
=>> node target/server.js
search1 will take 2 secs
search2 will take 2 secs
timeout will take 2 secs
read log waits 9
search1 finished
search2 finished
timeout finished
got search1 found x1 from #object[cljs.core.async.impl.channels.ManyToManyChannel]
result :hit-search
^C
=>> node target/server.js
search1 will take 6 secs
search2 will take 3 secs
timeout will take 1 secs
read log waits 6
timeout finished
result :timeout
search2 finished

一拆二/过滤

split 直接拆成两个

看文档好像就是直接这样拆成两个的, 对应 true/false,

(defn demo-split []
  (let [<c0 (to-chan! (range 20))]
    (let [[<c1 <c2] (async/split odd? <c0)]
      (go (display-all <c2 "from c2"))
      (go (display-all <c1 "from c1")))))

然后得到数据就是分别从不同的 channel 才能得到了, 奇书和偶数,

=>> node target/server.js
from c2 0
from c1 1
from c1 3
from c2 2
from c2 4
from c2 6
from c1 5
from c1 7
from c1 9
from c2 8
from c2 10
from c2 12
from c1 11
from c1 13
from c1 15
from c2 14
from c2 16
from c2 18
from c1 17
from c1 19
from c1 nil
from c2 nil

用 pipeline 搭配 filter 进行过滤

这个 pipeline 就是中间插入一个函数, 例子里是 filter, 直接进行过滤.

(defn demo-pipeline-filter []
  (let [<c1 (to-chan! (range 20)),
        <c2 (chan)]
    (async/pipeline 1 <c2 (filter even?) <c1)
    (display-all <c2)))

效果就是从 c2 取数据时, 只剩下偶数的值了,

=>> node target/server.js
0
2
4
6
8
10
12
14
16
18
nil

用 transducer 写法进行过滤

transducer 比较高级一点, 用到高阶函数跟比较复杂的抽象,
但是简单的功能写出来, 主要发挥作用的函数那个 (filter even?),
柯理化的用法, 返回函数, 然后被 comp 拿去组合,

(defn demo-transduce-filter []
  (let [<c1 (to-chan! (range 20)),
        <c2 (chan 1 (comp (filter even?)))]
    (async/pipe <c1 <c2)
    (display-all <c2)))

得到的结果跟上边是一样的, 都是过滤出偶数,

=>> node target/server.js
0
2
4
6
8
10
12
14
16
18
nil

一对多

通过 mult 发送给多个接收端

就是把一个数据变成多份, 供多个 channel 过来取数据,

(defn demo-mult []
  (let [<c0 (async/to-chan! (range 10)),
        <c1 (chan),
        <c2 (chan),
        ; mult-in 也是一个控制, 而不是一个 channel
        mult-in (async/mult <c0)]
    (async/tap mult-in <c1)
    (async/tap mult-in <c2)
    (display-all <c1 "from c1")
    (comment "need to take from c2, otherwise c0 is blocked")
    (display-all <c2 "from c2")))

可以看到运行以后就是 c1 c2 分别拿到一份一样的数据了,

=>> node target/server.js
from c1 0
from c2 0
from c1 1
from c1 2
from c2 1
from c2 2
from c1 3
from c1 4
from c2 3
from c2 4
from c1 5
from c1 6
from c2 5
from c2 6
from c1 7
from c1 8
from c2 7
from c2 8
from c1 9
from c1 nil
from c2 9
from c2 nil

大概的场景应该就是一个数据发布到多个 channel 去吧.
不过这个跟监听还有点不一样, 监听广播时发送者是非阻塞的, 这边是 channel 是阻塞的.

结尾

代码后续会同步到 github 去 https://github.com/worktools/... .

这边主要还是 API 的用法, 业务场景当中使用 core.async 会复杂一些,
比如 debounce 的逻辑用 timeout 搭配 loop 处理就比较顺,
具体代码参考, https://zhuanlan.zhihu.com/p/...
但一般都是会搅尾递归在里边, 进行逻辑控制和状态的传递.
网上别人的例子加上业务逻辑还会复杂很多很多...

但总的说, Clojure 提供的 API, 还有抽象能力, 应该是可以用来应对很多场景的.


题叶
17.3k 声望2.6k 粉丝

Calcit 语言作者


引用和评论

0 条评论