首页 > 代码库 > Scalaz(59)- scalaz-stream: fs2-程序并行运算,fs2 running effects in parallel
Scalaz(59)- scalaz-stream: fs2-程序并行运算,fs2 running effects in parallel
scalaz-stream-fs2是一种函数式的数据流编程工具。fs2的类型款式是:Stream[F[_],O],F[_]代表一种运算模式,O代表Stream数据元素的类型。实际上F就是一种延迟运算机制:F中间包含的类型如F[A]的A是一个可能会产生副作用不纯代码(impure code)的运算结果类型,我们必须用F对A运算的延迟机制才能实现编程过程中的函数组合(compositionality),这是函数式编程的标准做法。如果为一个Stream装备了F[A],就代表这个Stream会在处理数据元素O的过程中对O施用运算A,如果这个运算A会与外界交互(interact with outside world)如:文件、数据库、网络等的读写操作,那么这个Stream有数据元素I/O功能的需求。我们可以通过fs2 Stream的状态机器特性(state machine)及F[A]与外界交互功能来编写完整的数据处理(data processing)程序。如果能够在数据库程序编程中善用fs2的多线程运算模式来实现对数据库存取的并行运算,将会大大提高数据处理的效率。我们将在本篇着重讨论fs2在实现I/O程序中的有关方式方法。
首先,我们需要以整体Stream为程序运算框架,把与外界交互的运算A串联起来,然后通过Stream的节点来代表程序状态。我们首先需要某种方式把F[A]与Stream[F,A]关联起来,也就是我们所说的把一个F[A]升格成Stream[F,A]。fs2提供了Stream.eval函数,我们看看它的类型款式:
def eval[F[_], A](fa: F[A]): Stream[F, A] = attemptEval(fa) flatMap { _ fold(fail, emit) }
很明显,提供一个F[A],eval返回Stream[F,A]。这个返回结果Stream[F,A]的元素A是通过运算F[A]获取的:在一个数据库程序应用场景里这个A可能是个数据库连接(connection),那么F[A]就是一个连接数据库的操作函数,返回的A是个连接connection。这次我们来模拟一个对数据库表进行新纪录存储的场景。一般来说我们会按以下几个固定步骤进行:
1、连接数据库,获取connection连接
2、产生新数据(在其它场景里可能是读取数据然后更新)。这可能是一个循环的操作
3、将数据写入数据库
这三个步骤可以用Stream的三种状态来表示:一个源头(source)、传转(pipe transducer)、终点(sink)。
我们先示范如何构建源头:这是一种占用资源的操作,会产生副作用,所以我们必须用延迟运算方式来编程:
1 //用Map模拟数据库表2 import scala.collection.mutable.Map3 type DataStore = Map[Long, String]4 val dataStore: DataStore = Map() ?//> dataStore : fs2Eval.DataStore = Map()5 case class Connection(id: String, store: DataStore)6 def src(producer: String): Stream[Task,Connection] =7 Stream.eval(Task.delay { Connection(producer,dataStore)})8 ? //> src: (producer: String)fs2.Stream[fs2.Task,fs2Eval.Connection]
这个示范用了一个mutable map类型来模拟会产生副作用的数据库表。我们把具体产生数据的源头用Connection.id传下去便于在并行运算示范里进行跟踪。在这个环节里我们模拟了连接数据库dataStore操作。
产生数据是在内存里进行的,不会使用到connection,但我们依然需要把这个connection传递到下个环节:
1 case class Row(conn: Connection, key: Long, value: String)2 val recId = new java.util.concurrent.atomic.AtomicLong(1)3 ?//> recId : java.util.concurrent.atomic.AtomicLong = 14 def createData(conn: Connection): Row =5 Row(conn, recId.incrementAndGet, s"Producer $conn.id: at ${System.currentTimeMillis}")6 ?//> createData: (conn: fs2Eval.Connection)fs2Eval.Row7 val trans: Pipe[Task,Connection,Row] = _.map {conn => createData(conn)}8 ? //> trans : fs2.Pipe[fs2.Task,fs2Eval.Connection,fs2Eval.Row] = <function1>
trans是个Pipe。我们可以用through把它连接到src。
向数据库读写都会产生副作用。下一个环节我们模拟把trans传递过来的Row写入数据库。这里我们需要用延迟运算机制:
1 def log: Pipe[Task, Row, Row] = _.evalMap { r =>2 Task.delay {println(s"saving row pid:${r.conn.id}, rid:${r.key}"); r}}3 def saveRow(row: Row) = row.conn.store += (row.key -> row.value)4 5 val snk: Sink[Task,Row] = _.evalMap { r =>6 Task.delay { saveRow(r); () } }
增加了个跟踪函数log。从上面的代码可以看出:实际上Sink就是Pipe,只不过返回了()。
我们试试把这几个步骤连接起来运算一下:
1 val sprg = src("001").through(trans).repeat.take(3).through(log).to(snk)2 ?//> sprg : fs2.Stream[fs2.Task,Unit] = evalScope(Scope(Bind(Eval(Snapshot),<function1>))).flatMap(<function1>).flatMap(<function1>).flatMap(<function1>).flatMap(<function1>)3 sprg.run.unsafeRun //> saving row pid:001, rid:24 //| saving row pid:001, rid:35 //| saving row pid:001, rid:46 println(dataStore) //> Map(2 -> Connection(001,Map()).id: at 1472605736214, 4 -> Connection(001,Map(2 -> Connection(001,Map()).id: at 1472605736214, 3 -> Connection(001,Map(2 -> Connection(001,Map()).id: at 1472605736214)).id: at 1472605736245)).id : at 1472605736248, 3 -> Connection(001,Map(2 -> Connection(001,Map()).id: at 1472605736214)).id: at 1472605736245)
我们看到mutable map dataStore内容有变化了。
如果我们把以上的例子用并行运算方式来实现的话,应该如何调整?为方便观察结果,我们先在几个环节增加一些时间延迟:
1 implicit val strategy = Strategy.fromFixedDaemonPool(4)2 implicit val scheduler = Scheduler.fromFixedDaemonPool(2)3 def src(producer: String): Stream[Task,Connection] =4 Stream.eval(Task.delay { Connection(producer,dataStore)}5 .schedule(3.seconds))6 7 val trans: Pipe[Task,Connection,Row] = _.evalMap {conn => 8 Task.delay{createData(conn)}.schedule(1.second)}
下面我们把一些类型调整成Stream[Task,Stream[Row]],然后把concurrent.join函数掺进去:
1 val srcs = concurrent.join(3)(Stream(src("001"),src("002"),src("003"),src("004"))) 2 //> srcs : fs2.Stream[fs2.Task,fs2Eval.Connection] = attemptEval(Task).flatMap 3 <function1>).flatMap(<function1>) 4 val recs: Pipe[Task,Connection,Row] = src =http://www.mamicode.com/> { 5 concurrent.join(4)(src.map { conn => 6 Stream.repeatEval(Task {createData(conn)}.schedule(1.second)) }) 7 } ? //> recs : fs2.Pipe[fs2.Task,fs2Eval.Connection,fs2Eval.Row] = <function1> 8 9 def saveRows(row: Row) = { row.conn.store += (row.key -> row.value); row}10 ? //> saveRows: (row: fs2Eval.Row)fs2Eval.Row11 val snks: Pipe[Task,Row,Row] = rs => {12 concurrent.join(4)(rs.map { r =>13 Stream.eval(Task {saveRows(r)}.schedule(1.second)) })14 } ?//> snks : fs2.Pipe[fs2.Task,fs2Eval.Row,fs2Eval.Row] = <function1>
我们试着把它们连接起来进行运算:
1 val par = srcs.through(recs).take(10).through(log("before")).through(chnn).through(log("after")) 2 ? //> par : fs2.Stream[fs2.Task,fs2Eval.Row] = attemptEval(Task).flatMap(<function1>).flatMap(<function1>).flatMap(<function1>) 3 par.run.unsafeRun //> before saving pid:001, rid:3 4 //| before saving pid:003, rid:2 5 //| before saving pid:002, rid:4 6 //| before saving pid:001, rid:5 7 //| after saving pid:001, rid:3 8 //| after saving pid:003, rid:2 9 //| before saving pid:003, rid:610 //| after saving pid:002, rid:411 //| before saving pid:002, rid:712 //| after saving pid:001, rid:513 //| before saving pid:001, rid:814 //| before saving pid:003, rid:915 //| after saving pid:003, rid:616 //| after saving pid:002, rid:717 //| before saving pid:002, rid:1018 //| before saving pid:004, rid:1119 //| after saving pid:001, rid:820 //| after saving pid:003, rid:921 //| after saving pid:002, rid:1022 //| after saving pid:004, rid:11
?从跟踪函数显示可以看出before,after是交叉发生的,这就代表已经实现了并行运算。
下面是本篇示范源代码:
1 import fs2._ 2 import scala.concurrent.duration._ 3 object fs2Eval { 4 5 //用Map模拟数据库表 6 import scala.collection.mutable.Map 7 type DataStore = Map[Long, String] 8 val dataStore: DataStore = Map() 9 case class Connection(id: String, store: DataStore)10 implicit val strategy = Strategy.fromFixedDaemonPool(4)11 implicit val scheduler = Scheduler.fromFixedDaemonPool(2)12 def src(producer: String): Stream[Task,Connection] =13 Stream.eval(Task.delay { Connection(producer,dataStore)}14 .schedule(3.seconds))15 case class Row(conn: Connection, key: Long, value: String)16 val recId = new java.util.concurrent.atomic.AtomicLong(1)17 def createData(conn: Connection): Row =18 Row(conn, recId.incrementAndGet, s"$conn.id: at ${System.currentTimeMillis}")19 val trans: Pipe[Task,Connection,Row] = _.evalMap {conn =>20 Task.delay{createData(conn)}.schedule(1.second)}21 22 def log(pfx: String): Pipe[Task, Row, Row] = _.evalMap { r =>23 Task.delay {println(s"$pfx saving pid:${r.conn.id}, rid:${r.key}"); r}}24 def saveRow(row: Row) = row.conn.store += (row.key -> row.value)25 26 val snk: Sink[Task,Row] = _.evalMap { r =>27 Task.delay { saveRow(r); () } }28 29 val sprg = src("001").through(trans).repeat.take(3).through(log("")).to(snk)30 //sprg.run.unsafeRun31 //println(dataStore)32 33 val srcs = concurrent.join(3)(Stream(src("001"),src("002"),src("003"),src("004")))34 val recs: Pipe[Task,Connection,Row] = src =http://www.mamicode.com/> {35 concurrent.join(4)(src.map { conn =>36 Stream.repeatEval(Task {createData(conn)}.schedule(1.second)) })37 }38 39 def saveRows(row: Row) = { row.conn.store += (row.key -> row.value); row}40 val chnn: Pipe[Task,Row,Row] = rs => {41 concurrent.join(4)(rs.map { r =>42 Stream.eval(Task {saveRows(r)}.schedule(1.second)) })43 }44 45 46 val par = srcs.through(recs).repeat.take(10).through(log("before")).through(chnn).through(log("after"))47 par.run.unsafeRun
Scalaz(59)- scalaz-stream: fs2-程序并行运算,fs2 running effects in parallel