Solver keeps running after optimal solution found

[baklazan]

What version of OR-Tools and what language are you using?
Version: v9.15
Language: C++

Which solver are you using (e.g. CP-SAT, Routing Solver, GLOP, BOP, Gurobi)
CP-SAT

What operating system (Linux, Windows, ...) and version?
Linux

What did you do?

1. Create an operations_research::sat::Model instance
2. Register a feasible solution observer
3. Register an external boolean as limit
4. Run operations_research::sat::SolveCpModel using the Model from step 1

Here's a minimum example:

#include <atomic>
#include <fstream>
#include <iostream>

#include "ortools/sat/cp_model.pb.h"
#include "ortools/sat/cp_model_solver.h"
#include "ortools/util/time_limit.h"

namespace sat = operations_research::sat;

int main(int argc, char** argv) {
  if (argc != 2) {
    std::cerr << argv[0] << " input-proto-message\n";
    return -1;
  }
  sat::CpModelProto cp_model;
  std::ifstream in(argv[1], std::ios::binary);
  if (!cp_model.ParseFromIstream(&in)) {
    std::cerr << "Failed to open or parse " << argv[1] << "\n";
    return -1;
  }

  sat::SatParameters params;
  params.set_num_workers(16);
  params.set_log_search_progress(true);

  std::atomic<bool> stop;
  sat::Model model;

  model.Add(NewSatParameters(params));
  model.Add(sat::NewFeasibleSolutionObserver(
      [](const sat::CpSolverResponse& response) {
        std::cerr << response.best_objective_bound() << " ≤ "
                  << response.objective_value() << "\n";
      }));
  model.GetOrCreate<operations_research::TimeLimit>()
      ->RegisterExternalBooleanAsLimit(&stop);

  sat::CpSolverResponse response = sat::SolveCpModel(cp_model, &model);
  std::cout << "Solver finished, status=" << response.status() << "\n";
}

I ran the above code on this input: model.zip.
It takes about 2 minutes to find an optimal solution.

What did you expect to see

After the solver finds an optimal solution, it terminates.

What did you see instead?

The solver keeps on running.
I first noticed this issue when I "manually" (using the stop variable) stopped the solver after 10 minutes and it reported an optimal solution with a tight lower bound.
I subsequently verified by adding a debug print to operations_research::SharedTimeLimit::Stop.

Anything else we should know about your project / environment

If I reverse the order of model.Add calls (register bool limit first, then the feasible solution observer), the solver terminates as expected.

I think the issue might be caused by NewFeasibleSolutionObserver creating an instance of SharedResponseManager, which forces creation of ModelSharedTimeLimit. When I later register the external boolean limit in TimeLimit, it doesn't correctly propagate to SharedTimeLimit.

[lperron]

is it deterministic ?

[lperron]

Actually, look at:

or-tools/ortools/sat/samples/stop_after_n_solutions_sample_sat.cc

Line 58 in b21a132

StopSearch(&model);

[baklazan]

is it deterministic ?

Don't know if deterministic, but it seems pretty consistent.

Actually, look at:

or-tools/ortools/sat/samples/stop_after_n_solutions_sample_sat.cc

Line 58 in b21a132

StopSearch(&model);

Is StopSearch safe to call from another thread?

For context, I'm trying to stop the solver after the earlier of these events:

1. An optimal solution was found and proven to be optimal
2. A time limit of $L$ has passed after the first feasible solution was found.

I'm currently doing this by spawning a new thread after the first feasible solution is found; the thread sets the atomic bool after the time limit has passed. Would it be safe to call StopSearch instead?

bool timer_started = false;
  model.Add(sat::NewFeasibleSolutionObserver([&](const sat::CpSolverResponse& response) {
        if (!timer_started) {
          timer_started = true;
          std::thread([&]() {
            std::this_thread::sleep_for(std::chrono::seconds(300));
            stop.store(true);
          }).detach();
        }
      }));

[lperron]

yes. It uses the same atomic boolean mechanism. What happens is that there are multiple time limit objects in the solver sharing the atomic_bool, and overriding the atomic bool will sever the connection between these time limits Laurent Perron | Operations Research | ***@***.*** | (33) 1 42 68 53 00 Le mar. 20 janv. 2026 à 12:33, baklazan ***@***.***> a écrit :

…

*baklazan* left a comment (google/or-tools#4997) <#4997 (comment)> is it deterministic ? Don't know if deterministic, but it seems pretty consistent. Actually, look at: or-tools/ortools/sat/samples/stop_after_n_solutions_sample_sat.cc <https://github.com/google/or-tools/blob/b21a1326905a85b55cb52945716e10990a7cb3bf/ortools/sat/samples/stop_after_n_solutions_sample_sat.cc#L58> Line 58 in b21a132 <http:///google/or-tools/commit/b21a1326905a85b55cb52945716e10990a7cb3bf> StopSearch(&model); Is StopSearch safe to call from another thread? For context, I'm trying to stop the solver after the earlier of these events: 1. An optimal solution was found and proven to be optimal 2. A time limit of $L$ has passed after the first feasible solution was found. I'm currently doing this by spawning a new thread after the first feasible solution is found; the thread sets the atomic bool after the time limit has passed. Would it be safe to call StopSearch instead? bool timer_started = false; model.Add(sat::NewFeasibleSolutionObserver([&](const sat::CpSolverResponse& response) { if (!timer_started) { timer_started = true; std::thread([&]() { std::this_thread::sleep_for(std::chrono::seconds(300)); stop.store(true); }).detach(); } })); — Reply to this email directly, view it on GitHub <#4997 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACUPL3PFDLVX6WBVXHNZXJ34HYG7JAVCNFSM6AAAAACSGLTQPKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTONZSGQZDANBZGM> . You are receiving this because you were assigned.Message ID: ***@***.***>