Load Balancers

Almost every design has one, and almost every candidate mentions it without saying which kind or how it actually picks a backend. Both details are where the follow-up questions live.

L4 vs. L7

A Layer 4 load balancer routes based on IP and port alone — it doesn't look at the HTTP request, just forwards TCP connections, so it's fast and protocol-agnostic. A Layer 7 load balancer reads the actual HTTP request — path, headers, cookies — and can route /api/* to one backend pool and /static/* to another, or route based on a cookie for sticky sessions. The trade-off is the obvious one: L7 does more work per request (it has to actually parse HTTP) in exchange for routing decisions L4 physically cannot make.

Routing algorithms

• Round robin — cycles through backends in order. Simple, and a reasonable default when every backend and every request costs about the same.
• Least connections — routes to whichever backend currently has the fewest active connections. Better than round robin when requests vary a lot in how long they take, since round robin can pile slow requests onto one unlucky instance.
• Consistent hashing — routes based on a hash of some request property (e.g. a user ID), so the same key reliably lands on the same backend. This is what makes a sharded cache work at all — without it, "which Redis node holds this key" would change every time the backend pool resizes.

Health checks and removal from rotation

A load balancer only helps if it stops routing to a backend that's actually failing — see Reliability for the liveness-vs-readiness distinction. An instance that fails its readiness check gets pulled out of rotation automatically, and put back once it starts passing again — this is also what makes a rolling deploy safe: new instances only receive traffic once they report ready.

L4/L7 in front of each other

In practice these aren't either/or — a common real deployment shape is an L4 load balancer at the network edge (fast, handles raw connection volume) in front of an L7 layer that does the actual path-based routing. Naming that layering, rather than treating "load balancer" as one undifferentiated box, is a reasonable signal at the senior level.