Replit· 1 FebKilling Containers at Scale
Replit cut session errors from 3% to under 0.5% by fixing slow Docker container shutdowns on preempt
Replit engineers investigated and fixed slow container shutdowns on preemptible VMs, reducing the session connection error rate from 3% to under 0.5% and the 99th percentile session boot time from 2 minutes to 15 seconds.
The root cause was that Docker kill operations took 20+ seconds when killing 100–200 containers simultaneously, far longer than the expected milliseconds.
