Evening peak hours in Korea concentrate traffic into a short window. A cold-start regression that looked acceptable in a midday staged rollout can dominate support tickets after 7 p.m. KST.
Budget against perceived wait, not lab ideals
Users forgive a slightly longer splash if the first useful screen arrives predictably. They do not forgive a start that sometimes finishes in two seconds and sometimes in eight. Measure p50 and p90 on the device classes that actually open the app during peak.
Separate network-bound work from local init
If your cold start waits on a remote config call, peak congestion will punish you. Release health reviews flag this early: move non-critical fetches after first paint, and keep a local default that still renders a usable home.
Rehearse the expand at the right clock time
When we support a KR launch, we prefer at least one expand step during a representative evening sample — even at 5% — so cold-start telemetry reflects real radio and CDN conditions.