Maybe write the many to many relations in a background job/queue, and push the result back to the user when they're synced. Granted, long polling will take a number of connections, but the socket server won't take down your front facing service.

Great to see that MongoDB turned out to be a quick solution, but it seems that the writes could have just been timed more effectively.

Thanks for the insight!