THis bug is reported from a customer (https://na1.salesforce.com/50030000001WQKI). And I have created a Forum issue to discuss it.

Current proposed solution is to allow the cache capacity to grow temporarily. BTW, this bug should not occur often as the condition to happen is:

1. Pool size has been reached immediately (e.g., under highly concurrent load)
2. SFSB is repeatedly accessed concurrently

In most cases, you can set the pool size to be pretty big (since memory is cheap) except where you have big SFSB size.