Look for optimizations in AutoConfigurationImportSelector

[dsyer]

A regression in performance in spring-projects-experimental/spring-init was traced to a hostspot in AutoConfigurationImportSelector.getAutoConfigurationEntry() - once we prevented that method from being called when we already knew the imports, we were able to save nearly 5% of startup time in a micro app.

[wilkinsona]

I'm not sure that there's much to be gained here in a typical app that uses @SpringBootApplication. AutoConfigurationImportSelector.getAutoConfigurationEntry(AutoConfigurationMetadata, AnnotationMetadata) is called once and takes in the region of 10ms on my laptop. 7ms of that is spent in AutoConfigurationImportSelector.filter(List<String>, AutoConfigurationMetadata).

Can you give us some code to run where it's taking 5% of startup and will appear as a hotspot?

[dsyer]

10ms is less than 5% of the app I was running, so probably I overstated it - I made other changes, but this was the biggest one. The hotspot was in getAttributes(). The app was this one: https://github.com/spring-projects-experimental/spring-init/blob/master/tests/benchmarks/src/main/java/com/example (there are 3 apps there and the difference between the "manual" and "init" ones was driving the changes). You'd have to go back a few commits in spring-init to wind back through the optimizations and see the effect.

In a @SpringBootApplication aren't the imports known by reading spring.factories? Why do we even need to read the annotation metadata in that case? It seems over-engineered for the simplest use case, so even if it is only a small gain, it might be worth having.

[wilkinsona]

10ms is less than 5% of the app I was running, so probably I overstated it

That's really frustrating. Precision is vital here otherwise we're going to waste time chasing benefits that don't exist.

The hotspot was in getAttributes()

That method takes < 1ms for me. As I said above, the main cost that I see is in filter. It takes 7ms with 3-4ms of that being the call to load the filters via spring.factories. That 7ms is time well spent as it removes considerably more than 7ms of processing later on in startup. See #7573.

Why do we even need to read the annotation metadata in that case?

To honour the exclude attribute.

[dsyer]

OK. So probably there is nothing further to do here for now. There was a real gain in the prototype, but it can't be translated into a general benefit for Spring Boot.