Instance Templates, Managed Instance Groups and Autoscaling

GCP Basics · Compute Engine – Machine Types, VMs, Disks and Managed Instance Groups · Lesson 4 of 5

An instance template describes a VM; a regional managed instance group runs identical VMs across zones with autoscaling, autohealing and rolling updates.

For anything that should scale or heal itself, do not create VMs one by one. Describe them once in an instance template and let a managed instance group (MIG) run them.

flowchart LR
  T["Instance template web-tpl-v2<br/>e2-medium, image family web"] --> MIG["Regional MIG mig-web<br/>asia-southeast1"]
  MIG --> A["VM in zone a"]
  MIG --> B["VM in zone b"]
  MIG --> C["VM in zone c"]
  AS["Autoscaler: CPU 60%, min 3, max 10"] --> MIG
  HC["Health check /healthz"] --> MIG
  %% step T: The template holds machine type, image, disks, network, service account, tags and metadata; it is immutable.
  %% step MIG: A regional MIG spreads identical VMs across the zones of the region.
  %% step A: If a zone fails, the MIG recreates capacity in the remaining zones.
  %% step AS: The autoscaler adds VMs when average CPU is above the 60% target and removes them when it drops.
  %% step HC: Autohealing recreates VMs that fail the application health check.

Instance templates

Immutable: to change anything, create a new template (web-tpl-v2) and roll it out.
Global (default) or regional templates.
Reference an image family so each new VM gets the latest image.

MIG features

Feature	What it does
Zonal vs regional	Regional MIGs spread over up to three zones: prefer them for production
Autoscaling	On CPU, load-balancer serving capacity, Cloud Monitoring metrics or schedules; with min/max replicas and a cool-down period
Autohealing	Health check with an initial delay, so slow-booting VMs are not killed during startup
Rolling updates	Replace VMs with a new template using `maxSurge` and `maxUnavailable`
Canary	Run two versions, e.g. 10% on the new template
Stateful MIG	Keep disks and names for stateful apps
Load balancing	A MIG is the usual backend of a backend service

Unmanaged instance groups only group different VMs for load balancing; they do not autoscale or autoheal.

gcloud compute instance-templates create web-tpl-v2 --machine-type=e2-medium   --image-family=web --image-project=dl-lab-dev --tags=web --no-address   --service-account=web-sa@dl-lab-dev.iam.gserviceaccount.com --scopes=cloud-platform
gcloud compute instance-groups managed create mig-web --region=asia-southeast1   --template=web-tpl-v2 --size=3 --health-check=hc-web --initial-delay=120
gcloud compute instance-groups managed set-autoscaling mig-web --region=asia-southeast1   --min-num-replicas=3 --max-num-replicas=10 --target-cpu-utilization=0.6
gcloud compute instance-groups managed rolling-action start-update mig-web --region=asia-southeast1   --version=template=web-tpl-v3 --max-surge=3 --max-unavailable=0

Command	Meaning
`gcloud compute instance-groups managed list-instances mig-web --region=…`	VMs and their status
`gcloud compute instance-groups managed resize mig-web --size=5 --region=…`	Fixed size (no autoscaler)
`gcloud compute instance-groups managed rolling-action replace mig-web --region=…`	Recreate all VMs

Traps: an initial delay that is too short causes endless recreate loops; editing a VM inside a MIG is lost when the MIG recreates it; maxUnavailable=0 with maxSurge=0 cannot make progress.