Support calling Java parser in SQLFlow

[tonyyang-svail]

This PR depends on #1170.

This PR is a part of #1126.

[tonyyang-svail]

Calling Java parser takes about 1100ms for HiveQL parser and 800ms for Calcite parser. I am not sure if it's normal for Java command-line programs. If it is not normal, we'd better do some profiling on the Java program. @Yancey1989 @weiguoz Do you have experience in profiling?

You can reproduce the result the following command.

$ echo "select 1" > /tmp/input.sql
$ time java -cp /opt/sqlflow/parser/parser-1.0-SNAPSHOT-jar-with-dependencies.jar org.sqlflow.parser.ParserAdaptorCmd -p calcite -i /tmp/input.sql -o /tmp/output.json

real	0m0.922s
user	0m0.690s
sys	0m0.530s

[typhoonzero]

@tonyyang-svail can you write a test case that starts jvm once and run the parsing for 1000 times to see if the major time consumed is caused by starting the jvm?

[tonyyang-svail]

@typhoonzero I have tested it by changing the command-line argument calcite to i_dont_exists, which would let the program exit before doing anything useful. And it takes about 137ms, which the original time is 900+ms. So should I assume the parsing algorithm is the bottleneck?

$ time java -cp /opt/sqlflow/parser/parser-1.0-SNAPSHOT-jar-with-dependencies.jar org.sqlflow.parser.ParserAdaptorCmd -p i_dont_exist -i /tmp/input.sql -o /tmp/output.json

invalid parser type i_dont_exist
real	0m0.137s
user	0m0.060s
sys	0m0.080s

[typhoonzero]

@tonyyang-svail seems so. Maybe we can merge this first and do more profiling or performance fixes in another PR.

[typhoonzero]

Maybe we can refine this later using the Linux pipe to avoid writing temporary files on the disk.

[weiguoz]

Calling Java parser takes about 1100ms for HiveQL parser and 800ms for Calcite parser. I am not sure if it's normal for Java command-line programs. If it is not normal, we'd better do some profiling on the Java program. @Yancey1989 @weiguoz Do you have experience in profiling?

You can reproduce the result the following command.

$ echo "select 1" > /tmp/input.sql
$ time java -cp /opt/sqlflow/parser/parser-1.0-SNAPSHOT-jar-with-dependencies.jar org.sqlflow.parser.ParserAdaptorCmd -p calcite -i /tmp/input.sql -o /tmp/output.json

real	0m0.922s
user	0m0.690s
sys	0m0.530s

I think the time is mainly costed in the startup phase after running some tests on my mac.
CalciteParserAdaptor.ParseAndSplit(string sql)

# (test cases)	costs: millisecond
30,000	16,253
3,000	4,576

https://github.com/weiguoz/sqlflow/blob/testParsePerf/java/parser/src/test/java/org/sqlflow/parser/CalciteParserAdaptorTest.java#L118-L120

To avoid startup the slow JVM:

1. Calling by gRPC
2. Java Native Interface Golang Interface -- I didn't take a try.

[weiguoz]

LGTM.
Let's improve the calcite calling performance in the next PR

[tonyyang-svail]

I think the time is mainly costed in the startup phase after running some tests on my mac.
CalciteParserAdaptor.ParseAndSplit(string sql)

@weiguoz Wooow, you are awesome!