Distributed in-memory only cache with eviction

[evereststeakhouse]

Everyone,

First, let me say Hi to the whole Infinispan community as this is my first post.

Let me describe my case:

The plan is:

An in-memory only cache distributed on a cluster of 3 nodes with 2 owners per entry. Using the index size with a LRU eviction policy to keep a max count of objects we keep in cache hence in memory. As items are added in the cache, olders are kick-out and this is the intent. All entries are mortal but when the load gets significant, alive entries will also be evicted. So far so good but ..

The issue:

Let's assume a cluser of 3 nodes A, B & C with 2 owners per entries. Thus node A has item X and node B has its replicated copy. When A's index fills-up and eviction kicks-in and choose to evict X we observe the following:

• Node A does not have X element anymore (OK)
• Node B still has X copy (OK)

We understand this and this is well documented. Now if we try to access element X on each node we have:

• Node A returns not found
• Node B returns X
• Node C returns only X key without the data

This is where our issue starts.

Reading through the doc leads me to conclude eviction will only work in this mode if there is a cache store (with or without passivation) to back the cache data container.

So the questions:

• Is eviction requires a cache store in a distributed mode to work?
• Have we felt into a permitted but not supported config?
• Any future plans to make it work as we intended to?
• Any other way out of this?

Thanks,

Alain.

[brenuart]

AFAIK, yes, you would need a CacheLoader in order to regain access to the cache entries you have previously evicted.

It seems that the ClusterCacheLoader would fit your needs.

[evereststeakhouse]

Thanks Bertrand, sounds like a fair deal. I will give it a try and posts the results. If this works, the only thing I will have to worry about is the additional traffic on the cluster.

[evereststeakhouse]

Works like a charm.

[gaprashanth]

I have a similar question. If we configure eviction with a shared class loader and if there are 'n' nodes in the cluster,  consider a situation where the same object 'A' was evicted and passivated from node 1 but it is still in-memory in node 2. It is not evicted from node 2 but it is already there in in the disc as it was evicted from node 1. What happens while node 2 also tries to evict the same object later?

Also what happens if the same object A was requested at the same time from node 1 while it was being evicted from node 2 ? Here while it was being removed from the disc it was requested in node 1, it will be attempted to update from node 2 as it is evicted. How does Infiinispan handles the lock situation due to eviction with passivation turned on.

[galder.zamarreno]

1st q: Node 2 will override whatever node 1 evicted.

2nd q: It depends on whether the request is transactional or not, whether the underlying cache store is transactional and whether the nodes are clustered or not (cluster wide locks required or not)