See User discussion thread: http://apache-ignite-users.70518.x6.nabble.com/CacheStore-being-serialized-to-client-td1931.html
Brief summary: When a grid client joins the grid (clientMode=true) it receives a message from the server node(s) on the grid that contains the serialized CacheStore implementation object. If the client does not have this class on its CLASSPATH (and there is no reason it should, as it is a client) then the de-serialization of this message will fail, causing this exception:
where c.g.r.cachewrapper.ignite.CacheMissHandlerIgnite is the CacheStore implementation.
The ostensible reason for the CacheStore serialization is so that clients of a TRANSACTIONAL cache can begin the transaction on the underlying store.
The only current solution to this is to add the grid node's CacheStore implementation class definition to the CLASSPATH of the client. This creates an undesirable coupling between server and client.
UPDATE (copy-paste from comment below)
This is actually more generic issue. When a new node joins (client or server), all existing cache configurations (which include cache stores) are sent to this node. It then deserializes it during startup which can cause exceptions on clients or servers where cache is not supposed to be deployed as defined by node filter.
As a solution we can do the following:
- During discovery, send node filter and cache store factory in binary format along with the cache configuration, not as its parts.
- On server, check node filter first and deserialize cache configuration and cache store only if it returns true. In case of error, STOP join process (now we just print exception in log and go on, which is very error-prone).
- On client, do not deserialize cache configuration and cache store until user's code tries to actually use the cache (calls Ignite.cache. If cache is ATOMIC, never deserialize the cache store.