Distributed cache, single node write behind/through, failover if writing node goes down

[andylegear]

Hello

We are looking at possibly using infinispan and were looking at the write behind functionality.   We are hoping to satisfy the following requirements, but are not sure if infinispan is configurable in this way (or if it is event supported)

- Distributed (of course infinispan has this - writes on one node are reflected in the cache on another node, clustered using jgroups)

- Write-through/write behind to backing store only happens on a single node.  While writes to the cache are distributed throughout the cluster,  we would like the write though/behind to write to our backing store only from a single node.   Our backing store is a database table. So we only want only one node writing the updated shared state.  I notice that using the "loader" config you can define a custom loader and store classes to handle this on each node.  I also notice in distributed mode you can set numOwners.  Bug If you have a single owner can you have only that owner write? is the owner effectively the "primary" node? Or do all nodes have to be writer in write behind/though mode?

- Failover.  If our writing node fails, a new node in the cluster takes over and takes the role of writer.

- Collision detection.  I notice that MVCC locking is used in infinispan - As an optimistic locking approach we need to guard against two nodes writing at the same time,  become aware of this and handle the collision (possibly have our application rerun the transaction if it notices a collision).

Some guideance on the these requirements would be a great help. 

thanks

Andy

[galder.zamarreno]

If you only want one of the nodes to write to the backing store, simply enable the singleton store functionality, see https://docs.jboss.org/author/x/DwY5 for more info. I'm pretty sure this works fine with a replicated cache or a totally replicated cached, but with distribution, or partial replication, it might not work as expected because it relies on being the coordinator of the cluster in order to store the data (see https://docs.jboss.org/author/x/GwY5 to find out more about clustering modes). However, with distribution the coordinator might not all the data, so it might only work partially.

[andylegear]

Very good, thanks.  I'll try that out.

And what about collision detection (two transactions on separate nodes writing to the same cache data) - will collisions reveal themselves in the form of an exception? or is it not possible with MVCC?

Andy

[gdanov]

Hi,

I read the documentation you linked to , however it is not fully clear to me - what is the default behaviour of a cacheStore in distributed cache with backups? If node X writes value which is located in the primary partition of node A, and in the backup of B, by default showe CacheStore gets to write it to the database?

Best Regards,

Georgi

[galder.zamarreno]

@Andrew, The reaction to those scenarions depends on whether you use optimistic or pessimistic locking. See https://docs.jboss.org/author/x/FAY5

@Georgi, I don't understand your question. With singleton store, only the coordinator of the cluster will write to the cache store. So, basically, unless the coordinator is an owner or backup for that data, it won't be written. This could be improved of course...