Previously, if a `@Cacheable` method was accessed with the same key by
multiple threads, the underlying method was invoked several times instead
of blocking the threads while the value is computed. This scenario
typically affects users that enable caching to avoid calling a costly
method too often. When said method can be invoked by an arbitrary number
of clients on startup, caching has close to no effect.
This commit adds a new method on `Cache` that implements the read-through
pattern:
```
<T> T get(Object key, Callable<T> valueLoader);
```
If an entry for a given key is not found, the specified `Callable` is
invoked to "load" the value and cache it before returning it to the
caller. Because the entire operation is managed by the underlying cache
provider, it is much more easier to guarantee that the loader (e.g. the
annotated method) will be called only once in case of concurrent access.
A new `sync` attribute to the `@Cacheable` annotation has been addded.
When this flag is enabled, the caching abstraction invokes the new
`Cache` method define above. This new mode bring a set of limitations:
* It can't be combined with other cache operations
* Only one `@Cacheable` operation can be specified
* Only one cache is allowed
* `condition` and `unless` attribute are not supported
The rationale behind those limitations is that the underlying Cache is
taking care of the actual caching operation so we can't really apply
any SpEL or multiple caches handling there.
Issue: SPR-9254
Previously, if a managed bean had only one non-default constructor, we
should still annotate it with `@Autowired` to properly use constructor
injection. Not doing so resulted in an error as the container was
trying to call the default (non-existing) constructor.
This commit updates this behaviour to automatically applyed the
autowiring semantic to any bean that has only one constructor. As
before, if more than one constructor is defined, `@Autowired` must be
specified to teach the container the constructor it has to use.
Issue: SPR-12278