Codegen class compilation takes tens to hundreds of milliseconds, a class cache is hit when generifiedCode of code generator is exactly the same.
It works fine when UDF only takes columns or symbols, but not efficient when one or more parameters in UDF is always distinct from the other.
Take face recognition for example, the face images are almost distinct from each other according to lighting, facial expressions and details.
It is important to reduce redundant class compilation especially for those low latency queries.
Cache miss rate and metaspace gc can also be reduced by eliminating the redundant classes.
Here is the query to get the persons whose last name is Brunner and hire from 1st Jan 1990:
SELECT full_name, hire_date FROM cp.`employee.json` where last_name = 'Brunner' and hire_date >= '1990-01-01 00:00:00.0';
Now get the persons whose last name is Bernard and hire from 1st Jan 1990.
SELECT full_name, hire_date FROM cp.`employee.json` where last_name = 'Bernard' and hire_date >= '1990-01-01 00:00:00.0';
Figure shows the compilation time of the generated code by the above query in FilterRecordBatch on my laptop
Figure shows the only difference of the generated code from the attachments is the last_name value at line 156.
It is straightforward that the redundant class compilation can be eliminated by making the string12 as a member of the class and set the value when the instance is created