Infinispan 复制缓存中的驱逐是如何工作的？

Question

Infinispan supports a type of cache tier called "Replicated" where the content of multiple caches (in different JVMs or even container instances/VMs/servers) are kept (asynchronously) synchronized. My question is how eviction is handled in this case - if a normal local eviction algorithm is used that only looks at one cache it sounds to me like the result would be suboptimal ie entries that is not used much in one JVM and therefore are evicted could be the most frequently used one in all the others (that, as I understand it, also will see these entries evicted due to the synchronization between them)...

Answer 1

You will want to check out this section of the Infinispan User Guide https://infinispan.org/docs/stable/titles/configuring/configuring.html#eviction-and-expiration_configuring-memory-usage

As you mentioned eviction is done per node and removes entries that were deemed to be used the least most recently. (Note that off JVM heap and on heap have slightly different algorithms). Due to this it is possible to have some nodes with and without data if you have run out of entries. This in turn can cause inconsistencies with what each node can see.

If you need to limit memory usage, which is the main point of eviction, all you need to do to keep consistency between all nodes is to have a persistent storage defined, such as file store which is covered in the next section https://infinispan.org/docs/stable/titles/configuring/configuring.html#persistence .

Note that expiration differs from eviction in that it is done cluster wide. So if you still want data consistency across all nodes, expiration can be used instead if it is sufficient for your use case.