4 / 5 Load Balancing

The Pool Went Dark

Twenty API servers sit behind an L7 balancer. Their /healthz endpoint runs a full dependency check including a query against the primary database. A four second database hiccup makes all twenty probes fail at once, the balancer ejects the entire pool, and a degraded database becomes a total outage. Which changes reduce this risk? Select all that apply.

A

Require several consecutive failed probes before ejecting a backend, and stagger probe timing so backends are not evaluated in lockstep

B

Enable session affinity so each client stays pinned to the backend it already reached while the shared database is degraded

C

Make the balancer's probe shallow (process liveness only) and keep the deep dependency check on a separate endpoint used by alerting

D

Configure the balancer to fail open, continuing to send traffic to all backends once the healthy fraction falls below a threshold

E

Shorten the probe interval to one second so the balancer detects the failing dependency sooner and ejects unhealthy backends before clients see errors