Trill
Trill copied to clipboard
Single invocation of code generation per query
In the current implementation, generated code artifacts are created on demand. That implies a code generation session for each generated operator, a generation session for each generated batch subtype, and a generation session for each generated dictionary subclass.
We speculate that if we can gather all code generation needs into a single code generation session to occur just before data source subscription that startup time for the query should decrease.
Note: If this feature does not result in a significant gain in startup performance, or it causes runtime throughput or latency regression, it should be abandoned.