autowire="byType" is flawed in combination with proxies [SPR-1650]

[spring-projects-issues]

Alef Arendsen opened SPR-1650 and commented

A user on the forum came with an issue where a singleton essentially isn't a singleton.

I've taken a futher look at the issue and it seems that if (in the attached sample), autowiring is turned on for the LocalSessionFactory in combination with two objects being proxied, 17 instances of the SingletonRepository are created whereas with autowiring turned off, only one instance is created.

There are three possible scenarios:

autowiring=autodetect + proxying enabled: test fails
autowiring=off + proxying objects: test passes
autowiring=autodetect + no proxied object: test passes

I don't have time to further investigate the problem at the moment and can't really think of a possible thing causing this ;-).

regards,
Alef

---

Affects: 1.2.6, 2.0 M1

Attachments:

• autowiring.zip (4.83 kB)

[spring-projects-issues]

Andy Shen commented

Also I found that moving SessionFactoryBean to the end of the config file will make the failed test pass.

[spring-projects-issues]

Juergen Hoeller commented

It's not "autodetect", actually... "byType" causes the problem. "autodetect" simply chooses "byType" as default in the present case.

Unfortunately, this issue is hard to address.
The general advice to solve it is: Don't use autowiring by type at all. This avoids such side effects one and for all.
The less general advice is: Don't use autowiring by type for core beans that many other beans depend on.

We will be refining the behavior here, but this will mean throwing exceptions if an overuse of autowiring-by-type has been detected.
We unfortunately can't reliably predict whether autowiring-by-type will work properly upfront.

Juergen

[spring-projects-issues]

Juergen Hoeller commented

I have fixed the central issue behind this: The BeanFactory didn't clean up properly in case of attempted instantiations of a part of the object graph (on type checking). Some instances were able to hang around with old bean references in them. I have fixed this to clean up completely in all cases. This means that the attached test does pass now, since there will only ever be one instance of "SingletonRepository" at application runtime.

However, we can't do much about the temporarily created bean instances during type checking, so there are still multiple SingletonRepository instances created during initialization. There is no reliable to perform a FactoryBean type check without instantiating the backend object graph, unfortunately. So to avoid those unnecessary instantiation attempts, I strongly recommend to avoid autowire-by-type, or at least only use it very selectively (typically just on application services, not on core infrastructure beans).

That said, the result will be correct in all cases now. Using or not using autowire-by-type is just about potentially wasted instantiation effort now - essentially an optimization.

Juergen