[infinispan-issues] [JBoss JIRA] (ISPN-8396) Add interceptor preventing going out of memory

[ https://issues.jboss.org/browse/ISPN-8396?page=com.atlassian.jira.plugin.... ] William Burns edited comment on ISPN-8396 at 11/13/17 10:30 AM: ---------------------------------------------------------------- So here is my initial thoughts of how to do this in the core, which then could be used further up the stack We would add a new "eviction" type currently we have (COUNT and MEMORY) - I can't think of a good name, maybe MEMORY-EXCEPTION. This would then add a new interceptor to the stack towards the top that only takes a WrappedBytes and checks the length and the OffHeapMemoryAllocator to see if it can allocate that much memory. If not it would throw an exception. This allows off heap to use the non bounded data container which is substantially faster than the bounded one. The only drawback is this wouldn't work with on heap. We could add on heap (OBJECT storage would still not be supported), but it would require an additional wrapper around the DefaultDataContainer to count the byte[] coming in and out when using BINARY. So the only issue I found with this approach is that nodes can vary on whether they think they can insert values or not. So if a node was to write a given entry, a backup might not have room but the primary does. This could cause an issue with the write in nontx mode as it could have been updated on some of the nodes. Maybe we always write if the primary says we should? was (Author: william.burns): So here is my initial thoughts of how to do this in the core, which then could be used further up the stack We would add a new "eviction" type currently we have (COUNT and MEMORY) - I can't think of a good name, maybe MEMORY-EXCEPTION. This would then not actually add a new interceptor to the stack towards the top that only takes a WrappedBytes and checks the length and the OffHeapMemoryAllocator to see if it can allocate that much memory. If not it would throw an exception. This allows off heap to use the non bounded data container which is substantially faster than the bounded one. The only drawback is this wouldn't work with on heap. We could add on heap (OBJECT storage would still not be supported), but it would require an additional wrapper around the DefaultDataContainer to count the byte[] coming in and out when using BINARY. So the only issue I found with this approach is that nodes can vary on whether they think they can insert values or not. So if a node was to write a given entry, a backup might not have room but the primary does. This could cause an issue with the write in nontx mode as it could have been updated on some of the nodes. Maybe we always write if the primary says we should? -- This message was sent by Atlassian JIRA (v7.5.0#75005)