Dave Syer opened SPR-17646 and commented

Several opportunities:

  • Pointcut processing is expensive, and involves reflecting on all methods of all beans. Clearly this doesn't scale well and could be indexed away. Could also be used with @Transactional (and @Secured etc.).
  • The cache provider (e.g. EhCache) has expensive internal initialization. This could perhaps be deferred until the cache is actually used.
  • Some other stuff in Spring Boot might help (condition evaluation and import selection)

Some analysis and benchmarks here: https://github.com/spring-projects-experimental/spring-init/issues/5

Affects: 5.1.3

Reference URL: https://github.com/spring-projects-experimental/spring-init/issues/5

Comment From: spring-projects-issues

Stéphane Nicoll commented

The cache provider (e.g. EhCache) has expensive internal initialization. This could perhaps be deferred until the cache is actually used.

Just a note that several things may go wrong (configuration file not found, configuration invalid, network topology that has to be validated on startup) so whatever we do here should be opt-in. It would be nice though if we wouldn't have to do that for every single component that we optimize this way

Comment From: snicoll

Some additional data point here https://github.com/scratches/spring-cache-bench

Comment From: jhoeller

For a start, #22420 covers the pointcut processing part: We are not traversing classes now if they are known to not declare the desired annotation type. This is implemented for our async, caching, and transaction annotations, ignoring java* classes by default but also participating in the single unified hook (AnnotationUtils.isCandidateClass) which we have for our upcoming indexer integration.

Comment From: jhoeller

On review, I'm not really seeing much potential in asynchronous cache provider bootstrapping. Caffeine starts very quickly, not worth doing anything asynchronously. And cache preloading is usually done through early triggering of certain Cacheable methods (e.g. on ContextRefreshedEvent), not through the cache provider itself, even with EhCache these days.

From that perspective, I'll close this ticket in its provider initialization variant. Any further improvements in annotation discovery won't be specific to caching annotations and are therefore better off in a different ticket.