Github user maropu commented on the issue:
https://github.com/apache/spark/pull/21777
Yea, as you said, the JDK compiler generates different bytecode though, I
couldn't get obvious performance gains for TPCDS as compared to the janino one.
So, I couldn't find the strong reason to implement this in terms of
performance;
https://docs.google.com/spreadsheets/d/1Mgdd9dfFaACXOUHqKfaeKrj09hB3X1j9sKTJlJ6UM6w/edit#gid=1236423798
As another viewpoint, I think it might be useful to check if the generated
Java code of Spark could be compiled by the JDK compiler (JDK8 code
compatibility checks). But, since the compilation of the JDK compiler is too
slow (see the performance values in the google spread sheet above), IMO it is
impractical to check this in Jenkins.... (I found it took 7~8 hours to run the
tests of the `spark/sql` only in aws instances). WDYT?
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
