We would love to get our thread pools down to a single CPU thread pool (size = numCores) and a blocking thread pool (arbitrarily large). We may also require a scheduler pool for various options as well (limited size 1-2?).
To do this we need to remove remnants of our blocking code as possible. Possible issues for blocking are mostly around locks and io operations.
The persistence layer was completed with
ISPN-9722 so that is not an issue.
The requirement around locking can be relaxed if the locks are guaranteed to be small in scope and do not wrap other blocking operations. An example would be a lock such as ones in CHM as long as we don't have large blocks for functional argument types.
If code cannot be made non blocking we must offload the operation to the blocking thread pool. Care must be taken to ensure that once the blocking portion of code is completed that we switch back the to CPU thread pool as soon as possible. The listener API for example is violating this and will run code in Infinispan from any thread that completes the listener that could be done from a user.