References to the ServiceDiscovery instance will not stop GC, only references in the other direction would stop GC. The problem here is some cached values that are not cleaned up when a provider is closed, and the caches maintain references that stop the ServiceProvider instances from being cleaned up correctly.
When a provider is closed, the service discovery is notified (see org.apache.curator.x.discovery.details.ServiceProviderImpl#close). This allows the ServiceDiscovery to remove that provider from the list of providers that it will close when the ServiceDiscovery itself is closed. On 29 April 2014 16:14, shuxingzhang1988 <[email protected]> wrote: > Hi > > I think it maybe not a problem. In the test code,the so many > serviceprovders have the reference to the first servicediscovery which is > not released a,so they can not gc . > > > > > > At 2014-04-29 18:55:55,"Joe Littlejohn" <[email protected]> wrote: > > Hi, > > I've observed a memory leak in our production system using Curator service > discovery. I can replicate the problem with 2.4.1 and 2.4.2, though I > haven't tested any older versions. > > This test shows the problem: > > https://www.refheap.com/82891 > > (Hopefully this test is enough to demonstrate the problem. I haven't used > the TestingServer so you'll need a zk instance with a registered service.) > > If you run it and watch the process with jvisualvm you'll see that the > heap grows and grows as the test is running. Taking a heap dump will reveal > thousands of ServiceInstance and ServiceCacheImpl instances that are > retained even though the provider is closed after each usage. The > references appear to be traced back to the PathChildrenCache. I realise > it's possible to retain ServiceProvider instances, but this appears to be a > leak that shouldn't occur if the provider is correctly closed each time. > > Can anyone comment on whether this is indeed a problem? Am I missing > something? > > Cheers > > > >
