Use parse_url kernel for QUERY parsing

[thirtiseven]

This PR adds plugin support for QUERY parsing in parse_url (returning the entire query).

Contributes to #8963

Depends on: NVIDIA/spark-rapids-jni#1652

All cases from the kaggle dataset passed.

Perf test results

dataset: kaggle dataset repeated 10 times.

spark.time(df.selectExpr("COUNT(parse_url(url, 'QUERY')) as pr1", "COUNT(parse_url(url, 'QUERY')) as pr2", "COUNT(parse_url(url, 'QUERY')) as pr3", "COUNT(parse_url(url, 'QUERY')) as pr4", "COUNT(parse_url(url, 'QUERY')) as pr5", "COUNT(parse_url(url, 'QUERY')) as pr6", "COUNT(parse_url(url, 'QUERY')) as pr7", "COUNT(parse_url(url, 'QUERY')) as pr8", "COUNT(parse_url(url, 'QUERY')) as pr9", "COUNT(parse_url(url, 'QUERY')) as pr0").show())

GPU Time (ms)	CPU Time (ms)	Speed up
966	14,583	15.09x

[thirtiseven]

build

[revans2]

Sorry I did some manual testing and this falls over badly if I include the third parameter for query.

spark.time(spark.range(0, 300000000L, 1, 24).selectExpr("CONCAT('http://a.com?foo=', CAST(id AS STRING)) as h").selectExpr("SUM(length(parse_url(h, 'QUERY', 'foo'))) as q").show())

Caused by: java.lang.UnsupportedOperationException: parse_url(input[0, string, false](h#340), QUERY, foo) is not supported partToExtract=QUERY. Only PROTOCOL and HOST are supported
  at org.apache.spark.sql.rapids.GpuParseUrl.doColumnar(GpuParseUrl.scala:85)
  at org.apache.spark.sql.rapids.GpuParseUrl.$anonfun$columnarEval$5(GpuParseUrl.scala:111)
  at com.nvidia.spark.rapids.Arm$.withResourceIfAllowed(Arm.scala:74)
  at org.apache.spark.sql.rapids.GpuParseUrl.$anonfun$columnarEval$4(GpuParseUrl.scala:108)
  at com.nvidia.spark.rapids.Arm$.withResourceIfAllowed(Arm.scala:74)
  at org.apache.spark.sql.rapids.GpuParseUrl.$anonfun$columnarEval$3(GpuParseUrl.scala:107)
  at com.nvidia.spark.rapids.Arm$.withResourceIfAllowed(Arm.scala:74)

Can we please add tests that verify that we fall back to the CPU if we see the extra parameter for 'QUERY'?

[thirtiseven]

Sorry I did some manual testing and this falls over badly if I include the third parameter for query.

spark.time(spark.range(0, 300000000L, 1, 24).selectExpr("CONCAT('http://a.com?foo=', CAST(id AS STRING)) as h").selectExpr("SUM(length(parse_url(h, 'QUERY', 'foo'))) as q").show())

Caused by: java.lang.UnsupportedOperationException: parse_url(input[0, string, false](h#340), QUERY, foo) is not supported partToExtract=QUERY. Only PROTOCOL and HOST are supported
  at org.apache.spark.sql.rapids.GpuParseUrl.doColumnar(GpuParseUrl.scala:85)
  at org.apache.spark.sql.rapids.GpuParseUrl.$anonfun$columnarEval$5(GpuParseUrl.scala:111)
  at com.nvidia.spark.rapids.Arm$.withResourceIfAllowed(Arm.scala:74)
  at org.apache.spark.sql.rapids.GpuParseUrl.$anonfun$columnarEval$4(GpuParseUrl.scala:108)
  at com.nvidia.spark.rapids.Arm$.withResourceIfAllowed(Arm.scala:74)
  at org.apache.spark.sql.rapids.GpuParseUrl.$anonfun$columnarEval$3(GpuParseUrl.scala:107)
  at com.nvidia.spark.rapids.Arm$.withResourceIfAllowed(Arm.scala:74)

Can we please add tests that verify that we fall back to the CPU if we see the extra parameter for 'QUERY'?

Done, thanks!

[revans2]

build