BeanFactoryUtils.transformedBeanName() creates a lot of GC pressure [SPR-17151]

[spring-projects-issues]

Dave Syer opened SPR-17151 and commented

This is something that popped up while analysing a pathological Spring Data app with 2000 repositories, but it points at a potential general optimization that might help in Spring. The problem is that BeanFactoryUtils.transformedBeanName() is called for every bean in the context when someone wants to do DefaultListableBeanFactory.getBeansOfType() which is quite common in Spring Data and Spring Boot (and other frameworks). This in turn creates a new String for every factory bean which in turn shows up in the GC pressure graph in Flight Recorder (25% of GC pressure on startup of this extreme example). If the results were somehow cacheable in DefaultListableBeanFactory that would help.

---

Affects: 5.0.8

Referenced from: commits eddbf13

[spring-projects-issues]

Juergen Hoeller commented

So you mean the substring that temporarily strips the & prefix is causing the GC pressure here? The initial concatenation of prefix + bean name does not show as much?

[spring-projects-issues]

Dave Syer commented

Yes, that was it (the substring).

[spring-projects-issues]

Juergen Hoeller commented

That's a bit surprising... We also concatenate the prefix with the bean name in quite a few places, easily as often as the inverse transformation. I suppose these somehow end up in the JVM's String pool whereas the substring call always creates a new String programmatically.

In any case, I'll try to cache those transformed names in a ConcurrentHashMap. Sounds over the top for a simple prefixed String but it nevertheless seems to matter. It has been raised to me before by a performance engineer, BTW... but not with concrete numbers.