Spark - Non-time-based windows are not supported on streaming DataFrames/Datasets;

Question

I need to write Spark sql query with inner select and partition by. Problem is that I have AnalysisException. I already spend few hours on this but with other approach I have no success.

Exception:

Exception in thread "main" org.apache.spark.sql.AnalysisException: Non-time-based windows are not supported on streaming DataFrames/Datasets;;
Window [sum(cast(_w0#41 as bigint)) windowspecdefinition(deviceId#28, timestamp#30 ASC NULLS FIRST, RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW) AS grp#34L], [deviceId#28], [timestamp#30 ASC NULLS FIRST]
+- Project [currentTemperature#27, deviceId#28, status#29, timestamp#30, wantedTemperature#31, CASE WHEN (status#29 = cast(false as boolean)) THEN 1 ELSE 0 END AS _w0#41]

I assume that this is too complicated query to implement like this. But i don't know to to fix it.

 SparkSession spark = SparkUtils.getSparkSession("RawModel");

 Dataset<RawModel> datasetMap = readFromKafka(spark);

 datasetMap.registerTempTable("test");

 Dataset<Row> res = datasetMap.sqlContext().sql("" +
                " select deviceId, grp, avg(currentTemperature) as averageT, min(timestamp) as minTime ,max(timestamp) as maxTime, count(*) as countFrame " +
                " from (select test.*,  sum(case when status = 'false' then 1 else 0 end) over (partition by deviceId order by timestamp) as grp " +
                "  from test " +
                "  ) test " +
                " group by deviceid, grp ");

Any suggestion would be very appreciated. Thank you.

Answer 1

I believe the issue is in the windowing specification:

over (partition by deviceId order by timestamp) 

The partition would need to be over a time based column - in your case timestamp . The following should work:

over (partition by timestamp order by timestamp) 

That will not of course address the intent of your query. The following might be attempted: but it is unclear whether spark would support it:

over (partition by timestamp, deviceId order by timestamp) 

Even if spark does support that it would still change the semantics of your query.

Update

Here is a definitive source: from Tathagata Das who is a key/core committer on spark streaming: http://apache-spark-user-list.1001560.n3.nabble.com/Does-partition-by-and-order-by-works-only-in-stateful-case-td31816.html