Filter partial duplicates with mapWithState Spark

问题:

We have a DStream, such as

val ssc = new StreamingContext(sc, Seconds(1))

val kS = KafkaUtils.createDirectStream[String, TMapRecord](
  ssc,
  PreferConsistent,
  Subscribe[String, TMapRecord](topicsSetT, kafkaParamsInT)).
  mapPartitions(part => {
    part.map(_.value())
  }).
  mapPartitions(part1 => {
    part1.map(c => {
      TMsg(1,
        c.field1,
        c.field2, //And others
        c.startTimeSeconds
      )
    })
  })

So each RDD has a bunch of TMsg objects with some of the (technical) key fields I can use to dediplicate DStream. Basically, if we have two TMsg objects IN ONE OR TWO DISCRETIZED RDDs with same field1 and field2, and they differ by less than 1 second (we look on startTimeSeconds), it's a duplicate.

I looked over mapWithState. Yes I can create K -> V DStream like

val mappedStream = kS.map(m => (m.field1, m.field2) -> m.startTimeSeconds)

So I can use the function but don't understand how I can use it to filter duplicates.

Window function can't help, and I can't use (structured stream).deduplicate function since solution is written in DStreams.

Any solutions? Thanks

P.S. Spark version is 2.2

回答1:

You could use mapWithState. There is a good manual how to use Stateful Streaming. In your case you could:

1.Set checkpoin:

val ssc = new StreamingContext(sc, Seconds(1))
ssc.checkpoint("path/to/persistent/storage")

2.Define update function:

def update(key: (String, String),
           value: Option[Int],
           state: State[Int]): Option[((String, String), Int)] = {
  (value, state.getOption()) match {
    case (Some(_), Some(_)) => None
    case (Some(v), _) =>
      # you can update your state in any value you want
      # it is just a marker that value not new
      state.update(value.get)
      Option((key, v))
    case (_, _) if state.isTimingOut() => None
  }
}

3.Make state spec:

val stateSpec =
  StateSpec
    .function(update _)
    # it is important to define how long 
    # you want to check duplication
    # in this example check interval is 1 second.
    .timeout(Seconds(1))

4.Use it:

ks
  # make key->value pairs
  .map(m => (m.field1, m.field2) -> m.startTimeSeconds)
  .mapWithState(stateSpec)

if you want to take last of values instead, update function may be:

  def update(key: (String, String),
                       value: Option[Int],
                       state: State[Int]): Option[((String, String), Int)] = {
    (value, state.getOption()) match {
      case (Some(_), Some(_)) => None
      case (Some(v), _) =>
        state.update(value.get)
        None
      case (_, _) if state.isTimingOut() => Option((key, value.get))
    }
  }