O2Graceful shutdown and draining

The problem

1 s on screen = 100 ms

Why does a routine deploy make some users see errors?

Kill and restart

A deploy restarts each server in turn: it is stopped at once, mid-request, and started again 0.3 s later.

Live metrics for each version of the scene
Metric	Kill	Drain	Change
	—	—	—
	3	3	—
target ≤ 400 ms	—	—	—

O2Graceful shutdown and draining

Killing a server with requests in flight turns a deploy into an outage.