Thread and fork safety with mongocxx

TLDR: Always give each thread its own mongocxx::client.

In general each mongocxx::client object AND all of its child objects should be used by a single thread at a time. This is true even for clients acquired from a mongocxx::pool.

Even if you create multiple child objects from a single client, and synchronize them individually, that is unsafe as they will concurrently modify internal structures of the client. The same is true if you copy a child object.

Never do this

mongocxx::instance instance{};
mongocxx::client c{};
auto db1 = c["db1"];
auto db2 = c["db2"];
std::mutex db1_mtx{};
std::mutex db2_mtx{};

auto threadfunc = [](mongocxx::database& db, std::mutex& mtx) {
    std::scoped_lock<std::mutex>(mtx);
    db["col"].insert({});
}

// BAD! these two databases are individually synchronized, but they are derived from the same
// client, so they can only be accessed by one thread at a time
std::thread([]() { threadfunc(db1, db1_mtx); threadfunc(db2, db2_mtx); });
std::thread([]() { threadfunc(db2, db2_mtx); threadfunc(db1, db1_mtx); });

In the above example, even though the two databases are individually synchronized, they are derived from the same client. There is shared state inside the library that is now being modified without synchronization. The same problem occurs if db2 is a copy of db1.

A better version

mongocxx::instance instance{};
mongocxx::client c1{};
mongocxx::client c2{}
std::mutex c1_mtx{};
std::mutex c2_mtx{};

auto threadfunc = [](stdx::string_view dbname, mongocxx::client& client, std::mutex& mtx) {
    std::scoped_lock<std::mutex>(mtx);
    client[dbname]["col"].insert({});
}

// Good! these two clients are individually synchronized, so it is safe to share them between
// threads.
std::thread([]() { threadfunc("db1", c1, c1_mtx); threadfunc("db2", c2, c2_mtx); });
std::thread([]() { threadfunc("db2", c2, c2_mtx); threadfunc("db1", c1, c1_mtx); });

The best version

mongocxx::instance instance{};
auto threadfunc = [](mongocxx::client& client, stdx::string_view dbname) {
    client[dbname]["col"].insert({});
}
// don't even bother sharing clients. Just give each thread its own,
std::thread([]() {
    mongocxx::client c{};
    threadfunc(c, "db1");
    threadfunc(c, "db2");
});

std::thread([]() {
    mongocxx::client c{};
    threadfunc(c, "db2");
    threadfunc(c, "db1");
});

In most programs, clients will be long lived - so its less of a hassle (and more performant) to make one per-thread. Obviously in this contrived example, there’s quite a bit of overhead because we’re doing so little work with each client - but in a real program this is the best solution.

Fork safety

Neither a mongocxx::client or a mongocxx::pool can be safely copied when forking. Because of this, any client or pool must be created after forking, not before.