akka-typed(8)

本小妞迷上赌 提交于 2020-08-10 21:16:42

 前面介绍了事件源(EventSource)和集群(cluster),现在到了讨论CQRS的时候了。CQRS即读写分离模式,由独立的写方程序和读方程序组成,具体原理在以前的博客里介绍过了。akka-typed应该自然支持CQRS模式,最起码本身提供了对写方编程的支持,这点从EventSourcedBehavior 可以知道。akka-typed提供了新的EventSourcedBehavior-Actor,极大方便了对persistentActor的应用开发,但同时也给编程者造成了一些限制。如手工改变状态会更困难了、EventSourcedBehavior不支持多层式的persist,也就是说通过persist某些特定的event然后在event-handler程序里进行状态处理是不可能的了。我这里有个例子,是个购物车应用:当完成支付后需要取个快照(snapshot),下面是这个snapshot的代码:

       snapshotWhen {
          (state,evt,seqNr) => CommandHandler.takeSnapshot(state,evt,seqNr)
       }
...
 
  def takeSnapshot(state: Voucher, evt: Events.Action, lstSeqNr: Long)(implicit pid: PID) = {
    if (evt.isInstanceOf[Events.PaymentMade]
        || evt.isInstanceOf[Events.VoidVoucher.type]
        || evt.isInstanceOf[Events.SuspVoucher.type])
      if (state.items.isEmpty) {
        log.step(s"#${state.header.num} taking snapshot at [$lstSeqNr] ...")
        true
      } else
        false
    else
      false


}

判断event类型是没有问题的,因为正是当前的事件,但另一个条件是购物车必须是清空了的。这个有点为难,因为这个状态要依赖这几个event运算的结果才能确定,也就是下一步,但确定结果又需要对购物车内容进行计算,好像是个死循环。在akka-classic里我们可以在判断了event运算结果后,如果需要改变状态就再persist一个特殊的event,然后在这个event的handler进行状态处理。没办法,EventSourcedBehavior不支持多层persist,只有这样做:

 

      case PaymentMade(acct, dpt, num, ref,amount) =>
             ...
              writerInternal.lastVoucher = Voucher(vchs, vItems)
              endVoucher(Voucher(vchs,vItems),TXNTYPE.sales)
              Voucher(vchs.nextVoucher, List())
             ...   

 

我只能先吧当前状态保存下来、进行结单运算、然后清空购物车,这样snapshot就可以顺利进行了。

好了,akka的读方编程是通过PersistentQuery实现的。reader的作用就是把event从数据库读出来后再恢复成具体的数据格式。我们从reader的调用了解一下这个应用里reader的实现细节:

 

    val readerShard = writerInternal.optSharding.get   
    val readerRef = readerShard.entityRefFor(POSReader.EntityKey, s"$pid.shopId:$pid.posId")
    readerRef ! Messages.PerformRead(pid.shopid, pid.posid,writerInternal.lastVoucher.header.num,writerInternal.lastVoucher.header.opr,bseq,eseq,txntype,writerInternal.expurl,writerInternal.expacct,writerInternal.exppass)

可以看到这个reader是一个集群分片,sharding-entity。想法是每单完成购买后发个消息给一个entity、这个entity再完成reader功能后自动终止,立即释放出占用的资源。reader-actor的定义如下:

object POSReader extends LogSupport {
  val EntityKey: EntityTypeKey[Command] = EntityTypeKey[Command]("POSReader")

  def apply(nodeAddress: String, trace: Boolean): Behavior[Command] = {
    log.stepOn = trace
    implicit var pid: PID = PID("","")
    Behaviors.supervise(
      Behaviors.setup[Command] { ctx =>
        Behaviors.withTimers { timer =>
          implicit val ec = ctx.executionContext
          Behaviors.receiveMessage {
            case PerformRead(shopid, posid, vchnum, opr, bseq, eseq, txntype, xurl, xacct, xpass) =>
              pid = PID(shopid, posid)
              log.step(s"POSReader: PerformRead($shopid,$posid,$vchnum,$opr,$bseq,$eseq,$txntype,$xurl,$xacct,$xpass)")(PID(shopid, posid))
              val futReadSaveNExport = for {
                txnitems <- ActionReader.readActions(ctx, vchnum, opr, bseq, eseq, trace, nodeAddress, shopid, posid, txntype)
                _ <- ExportTxns.exportTxns(xurl, xacct, xpass, vchnum, txntype == Events.TXNTYPE.suspend,
                     { if(txntype == Events.TXNTYPE.voidall)
                       txnitems.map (_.copy(txntype=Events.TXNTYPE.voidall))
                     else txnitems },
                     trace)(ctx.system.toClassic, pid)
              } yield ()
              ctx.pipeToSelf(futReadSaveNExport) {
                case Success(_) => {
                  timer.startSingleTimer(ReaderFinish(shopid, posid, vchnum), readInterval.seconds)
                  StopReader
                }
                case Failure(err) =>
                  log.error(s"POSReader:  Error: ${err.getMessage}")
                  timer.startSingleTimer(ReaderFinish(shopid, posid, vchnum), readInterval.seconds)
                  StopReader
              }

            Behaviors.same
            case StopReader =>
              Behaviors.same
            case ReaderFinish(shopid, posid, vchnum) =>
              Behaviors.stopped(
                () => log.step(s"POSReader: {$shopid,$posid} finish reading voucher#$vchnum and stopped")(PID(shopid, posid))
              )
          }
        }
      }
    ).onFailure(SupervisorStrategy.restart)
  }

reader就是一个普通的actor。值得注意的是读方程序可能是一个庞大复杂的程序,肯定需要分割成多个模块,所以我们可以按照流程顺序进行模块功能切分:这样下面的模块可能会需要上面模块产生的结果才能继续。记住,在actor中绝对避免阻塞线程,所有的模块都返回Future, 然后用for-yield串起来。上面我们用了ctx.pipeToSelf 在Future运算完成后发送ReaderFinish消息给自己,通知自己停止。

在这个例子里我们把reader任务分成:

1、从数据库读取事件

2、事件重演一次产生状态数据(购物车内容)

3、将形成的购物车内容作为交易单据项目存入数据库

4、向用户提供的restapi输出交易数据

event读取是通过cassandra-persistence-plugin实现的:

    val query =
    PersistenceQuery(classicSystem).readJournalFor[CassandraReadJournal](CassandraReadJournal.Identifier)

    // issue query to journal
    val source: Source[EventEnvelope, NotUsed] =
      query.currentEventsByPersistenceId(s"${pid.shopid}:${pid.posid}", startSeq, endSeq)

    // materialize stream, consuming events
    val readActions: Future[List[Any]] = source.runFold(List[Any]()) { (lstAny, evl) => evl.event :: lstAny }

这部分比较简单:定义一个PersistenceQuery,用它产生一个Source,然后run这个Source获取Future[List[Any]]。

重演事件产生交易数据:

    def buildVoucher(actions: List[Any]): List[TxnItem] = {
      log.step(s"POSReader: read actions: $actions")
      val (voidtxns,onlytxns) = actions.asInstanceOf[Seq[Action]].pickOut(_.isInstanceOf[Voided])
      val listOfActions = onlytxns.reverse zip (LazyList from 1)   //zipWithIndex
      listOfActions.foreach { case (txn,idx) =>
        txn.asInstanceOf[Action] match {
          case Voided(_) =>
          case ti@_ =>
            curTxnItem = EventHandlers.buildTxnItem(ti.asInstanceOf[Action],vchState).copy(opr=cshr)
            if(voidtxns.exists(a => a.asInstanceOf[Voided].seq == idx)) {
              curTxnItem = curTxnItem.copy(txntype = TXNTYPE.voided, opr=cshr)
              log.step(s"POSReader: voided txnitem: $curTxnItem")
            }
            val vch = EventHandlers.updateState(ti.asInstanceOf[Action],vchState,vchItems,curTxnItem,true)
            vchState = vch.header
            vchItems = vch.txnItems
            log.step(s"POSReader: built txnitem: ${vchItems.txnitems.head}")
        }
      }
      log.step(s"POSReader: voucher built with state: $vchState, items: ${vchItems.txnitems}")
      vchItems.txnitems
    }

重演List[Event],产生了List[TxnItem]。

向数据库里写List[TxnItem]:

 

 def writeTxnsToDB(vchnum: Int, txntype: Int, bseq: Long, eseq: Long, txns: List[TxnItem])(
                   implicit system: akka.actor.ActorSystem, session: CassandraSession, pid: PID): Future[Seq[TxnItem]] = ???

注意返回结果类型Future[Seq[TxnItem]]。我们用for-yield把这几个动作串起来:

  val txnitems: Future[List[Events.TxnItem]] = for {
      lst1 <- readActions    //read list from Source
      lstTxns <- if (lst1.length < (endSeq -startSeq))    //if imcomplete list read again
        readActions
        else FastFuture.successful(lst1)
      items <- FastFuture.successful( buildVoucher(lstTxns) )
      _ <- JournalTxns.writeTxnsToDB(vchnum,txntype,startSeq,endSeq,items)
      _ <- session.close(ec)
    } yield items

注意返回结果类型Future[Seq[TxnItem]]。我们用for-yield把这几个动作串起来:

  val txnitems: Future[List[Events.TxnItem]] = for {
      lst1 <- readActions    //read list from Source
      lstTxns <- if (lst1.length < (endSeq -startSeq))    //if imcomplete list read again
        readActions
        else FastFuture.successful(lst1)
      items <- FastFuture.successful( buildVoucher(lstTxns) )
      _ <- JournalTxns.writeTxnsToDB(vchnum,txntype,startSeq,endSeq,items)
      _ <- session.close(ec)
    } yield items

注意:这个for返回的Future[List[TxnItem]],是提供给restapi输出功能的。在那里List[TxnItem]会被转换成json作为post的包嵌数据。

现在所有子任务的返回结果类型都是Future了。我们可以再用for来把它们串起来:

             val futReadSaveNExport = for {
                txnitems <- ActionReader.readActions(ctx, vchnum, opr, bseq, eseq, trace, nodeAddress, shopid, posid, txntype)
                _ <- ExportTxns.exportTxns(xurl, xacct, xpass, vchnum, txntype == Events.TXNTYPE.suspend,
                     { if(txntype == Events.TXNTYPE.voidall)
                       txnitems.map (_.copy(txntype=Events.TXNTYPE.voidall))
                     else txnitems },
                     trace)(ctx.system.toClassic, pid)
              } yield ()

说到EventSourcedBehavior,因为用了cassandra-plugin,忽然想起配置文件里新旧有很大区别。现在这个application.conf是这样的: 

akka {
  loglevel = INFO
  actor {
    provider = cluster
    serialization-bindings {
      "com.datatech.pos.cloud.CborSerializable" = jackson-cbor
    }
  }
  remote {
    artery {
      canonical.hostname = "192.168.11.189"
      canonical.port = 0
    }
  }
  cluster {
    seed-nodes = [
      "akka://cloud-pos-server@192.168.11.189:2551"]
    sharding {
      passivate-idle-entity-after = 5 m
    }
  }
  # use Cassandra to store both snapshots and the events of the persistent actors
  persistence {
    journal.plugin = "akka.persistence.cassandra.journal"
    snapshot-store.plugin = "akka.persistence.cassandra.snapshot"
  }
}
akka.persistence.cassandra {
  # don't use autocreate in production
  journal.keyspace = "poc2g"
  journal.keyspace-autocreate = on
  journal.tables-autocreate = on
  snapshot.keyspace = "poc2g_snapshot"
  snapshot.keyspace-autocreate = on
  snapshot.tables-autocreate = on
}

datastax-java-driver {
  basic.contact-points = ["192.168.11.189:9042"]
  basic.load-balancing-policy.local-datacenter = "datacenter1"
}

akka.persitence.cassandra段落里可以定义keyspace名称,这样新旧版本应用可以共用一个cassandra,同时在线。

 

易学教程内所有资源均来自网络或用户发布的内容,如有违反法律规定的内容欢迎反馈
该文章没有解决你所遇到的问题?点击提问,说说你的问题,让更多的人一起探讨吧!