[ https://issues.jboss.org/browse/ISPN-5016?page=com.atlassian.jira.plugin.... ] Radim Vansa commented on ISPN-5016: ----------------------------------- {quote}Because data is stored in the shared store after the invalidation command was executed on all the nodes, it’s also possible for a node to read the old value from the shared store after it has executed the invalidation.{quote} I don't understand the reason very well, could you elaborate a bit more? {quote}Invalidations triggered by write operations are synchronous and the operation will fail with a SuspectException if another node crashes. The entry will not be updated on the originator, but invalidation will still be performed on the non-crashed nodes.{quote} I thought SuspectExceptions are silently handled, aren't they? {quote}For multi-key operations like putAll(), the operation is sent to all the primary owners of the affected keys. When there is a single primary owner for all the affected keys, the operation succeeds and its effects are visible atomically. When there are multiple primary owners, the operation is not atomic: overlapping read operations will see random subsets of the updated values.{quote} As reads don't acquire any locks, I think that even update on single primary owner can be seen as non-atomic. -- This message was sent by Atlassian JIRA (v6.3.11#6341)