Strange to warn of a problem having already dismissed the solution (pooling). Unless your backend actually has fifty thousand cores, making fifty thousand concu

Strange to warn of a problem having already dismissed the solution (pooling). Unless your backend actually has fifty thousand cores, making fifty thousand concurrent connections to it accomplishes nothing more than moving the bottleneck around while defeating TCP flow control. Capacity is always finite, some part of the system is inevitably going to play the role of a bounded queue, and you want that visible and close to (preferably in) the frontend so you can shed load through degraded service.