From the scale command to a worker polling for jobs, measured on GKE with the cluster autoscaler. Node provisioning and the image pull dominate the new-node path — the worker's own boot is seconds.
0
30s
60s
90s
Warm node
cluster has room, image cached
~5 s
New node
cluster autoscaler adds a VM
Provision node
~60 s
Pull image
~24 s
~87 s
Provision node (VM boot + join)
Pull worker image (126 MB, uncached)
Start container (image cached)
Boot worker + connect (~3 s)
Scaling on a metric through an HPA adds its reaction time (typically 15–60 s) in front of both paths. The first run on a fresh worker adds a one-time ~2–3 s cold start.