[FEA] Support lazy quantifier and specified group index in regexp_extract function

[nvliyuan]

It would be nice if we could support lazy quantifier and specified group index in regexp_extract function.
For example:

create a hive table and insert some rows:
hive> show create table datavalid3;
OK
CREATE EXTERNAL TABLE `datavalid3`(
  `col1` int,
  `col2` bigint,
  `col3` tinyint,
  `col4` string,
  `col5` date,
  `col6` map<string,string>,
  `col7` map<string,array<string>>,
  `col8` decimal(5,3))
PARTITIONED BY (
  `col9` string)
ROW FORMAT SERDE
  'org.apache.hadoop.hive.ql.io.orc.OrcSerde'
STORED AS INPUTFORMAT
  'org.apache.hadoop.hive.ql.io.orc.OrcInputFormat'
OUTPUTFORMAT
  'org.apache.hadoop.hive.ql.io.orc.OrcOutputFormat'

run below SQL:
spark.sql("select regexp_extract(col4,'\"xst\":\"(.*?)\"') from datavalid3").show

fallback info:

!Exec <ProjectExec> cannot run on GPU because not all expressions can be replaced
    @Expression <Alias> regexp_extract(col4#3, "xst":"(.*?)", 1) AS regexp_extract(col4, "xst":"(.*?)", 1)#21 could run on GPU
      !Expression <RegExpExtract> regexp_extract(col4#3, "xst":"(.*?)", 1) cannot run on GPU because regex group count is 0, but the specified group index is 1; Lazy quantifier *? not supported near index 10

[revans2]

The specified group count issue I think is a side effect of not being able to support the lazy quantifier. I think we stopped parsing the regexp and didn't see the capture group around the lazy quantifier.

[revans2]

It looks like CUDF does support this feature already.

>>> import cudf
>>> s = cudf.Series(['a:"hello"', 'a:"there"'])
>>> s
0    a:"hello"
1    a:"there"
dtype: object
>>> s.str.extract('a:"(.*?)"')
       0
0  hello
1  there

So this is probably mostly a feature for us to figure out how to transpile .*? into something that will match exactly with what java/spark does with the regexp.

[nvliyuan]

It will take more time for the customer to double-confirm the priority.