how to calculate aggregations on a window when sensor readings are not sent if they haven't changed since last event?

Question

How can I calculate aggregations on a window, from a sensor when new events are only sent if the sensor value has changed since the last event? The sensor readings are taken at fixed times, e.g. every 5 seconds, but are only forwarded if the reading changes since the last reading.

So, if I'm would like to create an average of signal_stength for each device:

eventsDF = ... 
avgSignalDF = eventsDF.groupBy("deviceId").avg("signal_strength")

For example, events sent by the device for a one minute window:

event_time  device_id  signal_strength
12:00:00    1          5
12:00:05    1          4
12:00:30    1          5
12:00:45    1          6
12:00:55    1          5

The same dataset with the events that aren't actually sent filled in:

event_time  device_id  signal_strength
12:00:00    1          5
12:00:05    1          4
12:00:10    1          4
12:00:15    1          4
12:00:20    1          4
12:00:25    1          4
12:00:30    1          5
12:00:35    1          5
12:00:40    1          5
12:00:45    1          6
12:00:50    1          6
12:00:55    1          5

The signal_strength sum is 57 and the avg is 57/12

How can this missing data be inferred by spark structured streaming and the average calculated from the inferred values?

Note: I have used average as an example of an aggregation, but the solution needs to work for any aggregation function.

Answer 1

EDITED:

I have modified the logic to compute the average only from the filtered dataframe, so that it addresses the gaps.

//input structure
case class StreamInput(event_time: Long, device_id: Int, signal_strength: Int)
//columns for which we want to maintain state
case class StreamState(prevSum: Int, prevRowCount: Int, prevTime: Long, prevSignalStrength: Int, currentTime: Long, totalRow: Int, totalSum: Int, avg: Double)
//final result structure
case class StreamResult(event_time: Long, device_id: Int, signal_strength: Int, avg: Double)

val filteredDF = ???  //get input(filtered rows only)

val interval = 5  // event_time interval

// using .mapGroupsWithState to maintain state for runningSum & total row count till now

// you need to set the timeout threshold to indicate how long you wish to maintain the state
val avgDF = filteredDF.groupByKey(_.device_id)
  .mapGroupsWithState[StreamState, StreamResult](GroupStateTimeout.NoTimeout()) {

  case (id: Int, eventIter: Iterator[StreamInput], state: GroupState[StreamState]) => {
    val events = eventIter.toSeq

    val updatedSession = if (state.exists) {
      //if state exists update the state with the new values
      val existingState = state.get

      val prevTime = existingState.currentTime
      val currentTime = events.map(x => x.event_time).last
      val currentRowCount = (currentTime - prevTime)/interval
      val rowCount = existingState.rowCount + currentRowCount.toInt
      val currentSignalStength = events.map(x => x.signal_strength).last

      val total_signal_strength = currentSignalStength + 
        (existingState.prevSignalStrength * (currentRowCount -1)) + 
        existingState.total_signal_strength

      StreamState(
        existingState.total_signal_strength,
        existingState.rowCount,
        prevTime,
        currentSignalStength,
        currentTime,
        rowCount,
        total_signal_strength.toInt,
        total_signal_strength/rowCount.toDouble
      )

    } else {
      // if there are no earlier state
      val runningSum = events.map(x => x.signal_strength).sum
      val size = events.size.toDouble
      val currentTime = events.map(x => x.event_time).last
      StreamState(0, 1, 0, runningSum, currentTime, 1, runningSum, runningSum/size)
    }

    //save the updated state
    state.update(updatedSession)
    StreamResult(
      events.map(x => x.event_time).last,
      id,
      events.map(x => x.signal_strength).last,
      updatedSession.avg
    )
  }
}

val result = avgDF
  .writeStream
  .outputMode(OutputMode.Update())
  .format("console")
  .start

The idea is to calculate two new Columns:

1. totalRowCount: the running total of number of rows that are supposed to be present if you have not filtered.
2. total_signal_strength: the running total of signal_strength till now. (this INCLUDES missed row totals too).

Its calculated by:

total_signal_strength = 
  current row's signal_strength  +  
  (total_signal_strength of previous row * (rowCount -1)) + 
  //rowCount is the count of missed rows computed by comparing previous and current event_time.
  previous total_signal_strength

format of the intermediate state:

+----------+---------+---------------+---------------------+--------+
|event_time|device_id|signal_strength|total_signal_strength|rowCount|
+----------+---------+---------------+---------------------+--------+
|         0|        1|              5|                    5|       1|
|         5|        1|              4|                    9|       2|
|        30|        1|              5|                   30|       7|
|        45|        1|              6|                   46|      10|
|        55|        1|              5|                   57|      12|
+----------+---------+---------------+---------------------+--------+

final output:

+----------+---------+---------------+-----------------+
|event_time|device_id|signal_strength|              avg|
+----------+---------+---------------+-----------------+
|         0|        1|              5|              5.0|
|         5|        1|              4|              4.5|
|        30|        1|              5|4.285714285714286|
|        45|        1|              6|              4.6|
|        55|        1|              5|             4.75|
+----------+---------+---------------+-----------------+