Why scale changes the problem
A linear search comparing a live face against every enrolled embedding works fine at small scale but gets slower as the enrolled population grows — at real scale, 1:N identification needs indexing and search structures built specifically for high-dimensional comparison, not a brute-force loop.
Speed and accuracy trade-offs
Faster approximate search methods can return results more quickly at large scale, often at a small cost to precision — a trade-off that has to be tuned deliberately rather than assumed away.
Why response time is a real design constraint, not a nice-to-have
For a use case like a merchant-desk checkout, a 1:N identification call needs to return in close to real time — a sub-second design target shapes architecture choices from the ground up, not as an afterthought optimization.