# Scheduling Azure Virtual Machine Scale Set

> Scale set schedules deallocate all instances rather than deleting them. Each VM keeps its identity, OS disk and data disks, and compute billing stops for the whole set. Restart reallocates capacity for N machines simultaneously, so a 20-instance set in a constrained SKU is a real 8am risk.

Source: https://zop.dev/integrations/azure/vmss/schedule
Updated: 2026-08-19

---

## Deallocated, every one of them

The schedule powers the set down by deallocating each instance in place. Instance IDs, OS
disks, data disks and configuration all persist; compute billing stops across the set. This is
the whole-set equivalent of stopping a single VM properly (released allocation, preserved
machine), applied N times at once.

## Not the same as scaling in

Azure offers another way to make a scale set cheap at night: drop the capacity so instances
are deleted. The two differ exactly where it matters. Scale-in destroys the instances and
their disks: no disk bill overnight, but a morning of fresh VMs, fresh extensions and cold
caches. Deallocation keeps the machines. Disks bill through the night, but the morning is a
resume, not a rebuild. Stateful or slow-to-provision workloads want deallocation; fully
stateless fleets may prefer scale-in economics.

## N simultaneous capacity requests at start

Starting the set asks the region for capacity for every instance at once. General-purpose
sizes absorb that quietly; constrained SKUs such as GPU sizes and large memory-optimized machines
can fail partially, leaving the set below strength at 8am. If mornings must be certain on a
scarce SKU, that is the argument for capacity reservations, priced accordingly, or for not
scheduling that particular set.

## Autoscale keeps its own counsel

Autoscale rules attached to the set do not know about the calendar. Metric-based rules can
add instances overnight in response to noise, and those additions bill. A scheduled set
should have its autoscale profile aligned, with a night profile whose floor is zero, or the two
systems will trade instances back and forth in the dark.

## The overnight residual

Every deallocated instance's disks keep billing per GB, multiplied by the instance count, plus
any public load balancer and IP charges in front of the set. Per-instance disk cost times N is
the floor the schedule cannot cross.