At low load, the DB worker is idle when a request arrives and it can be dispatched immediately. But under load, the DB worker is busy when requests arrive and they build up as a backlog.
When a backlog builds, the DB worker can examine the pending requests and combine those that it can to reduce the total number of requests.
Not all requests are combinable and there can be subtle rules and side-effects. It is likely that a driver with just a modicum of combining ability would be very conservative e.g. simply combining single-key selects that are queued adjacently or such.
Even still, the gains can be massive and performance never worse.
At low load, the DB worker is idle when a request arrives and it can be dispatched immediately. But under load, the DB worker is busy when requests arrive and they build up as a backlog.
When a backlog builds, the DB worker can examine the pending requests and combine those that it can to reduce the total number of requests.
Not all requests are combinable and there can be subtle rules and side-effects. It is likely that a driver with just a modicum of combining ability would be very conservative e.g. simply combining single-key selects that are queued adjacently or such.
Even still, the gains can be massive and performance never worse.
(Oh, equally applicable to NoSQL too.)