SolverManager: a high-level wrapper for the low-level Solver to handle thread management, multitenancy, load-balancing, ...

The Solver deliberately doesn't do any thread creation. It just provides the low-level tread-safe methods such as terminateEarly() and addProblemFactChange(pfc).

The SolverManager will be a wrapper that does do thread management to add the following functionality.

Phase 1:

• It manages the thread pool with the internal solver threads.
• It exposes the solver status (not started, solving, terminate early, stopped, etc) to external threads.
• It exposes the best solution (and the best score) to the external threads.
• It support adding problem fact changes
• Thread safe
• Graceful shutdown See PLANNER-534
• Exception handling and reporting: Don't eat an exception (for example thrown by a custom score calculator), but instead relay it to the external threads when they ask for something.
• Pluggable thread management: Reuse managed ExecutorService from wildfly/EAP so the thread pool size can be controlled through JMX.

Phase 2 and later

• Handle a solver will itself spawns multiple threads. For example an 8 pool thread with 2 solvers each doing a 4-threaded Partitioned Search is already full...
• Multitenancy load balancing: Running 10 solver datasets with 8 cores will only run 8 in parallel and round-robin swap out the other 2 (if desired, alternatively it starts 10 threads and let the JVM do the round-robining).
  • it can run 100 solvers simultaneously on an 8 core machine but still only saturate 6 cores (if configured for that behavior),
    by injecting a runnableThreadSemaphore in each solver with a runnableSolverThreadLimit = 6.
    I've already done that trick for PartitionedSearch and that works like a charm
• Fail-over: threads can die
• High Availability? Horizontal scaling across multiple nodes.
• Load balancing?

SolverManager should be reusable by camel-optaplanner, mule-optaplanner, optaplanner-execution-server, ...

Strawman api proposal

interface SolverManager {
   Long addSolver(optional KieContainer, String solverConfigPath); // returns a solverId
   boolean removeSolver(Long solverId);
  
   boolean solveAsync(Long solverId, Solution planningProblem, optional callback);
   Solution getBestSolution(Long solverId);
   SolverStatus getSolverStatus(Long solverId);
   boolean terminateEarly(Long solverId);
   boolean addProblemFactChange(Long solverId, ProblemFactChange);
   ...
}

Discussions document: https://docs.google.com/document/d/16gZZUuxufjG4NLAS-FdLm9Rquhc0TrSespXJeddX5Xk/edit?usp=sharing