Why weighted capacity exists

Weighted capacity becomes useful when you set the group’s capacity unit to vCPU instead of instance count.

Instance-based capacity units

When the capacity unit is set to instance, scaling happens in whole instances. The downside is that a large instance type might launch unexpectedly, increasing cost and potentially leaving you with underutilized compute.

vCPU-based capacity units

When you set the capacity unit to vCPU, Elastigroup uses weights to represent the vCPU value of each instance type. This allows Elastigroup to scale based on compute power, not instance count.

Custom instance weights

Custom weighting lets you define a unique weight for each instance type. Scaling decisions consider the total weighted capacity rather than raw instance numbers.

How weighted scaling works

• You set a target capacity in vCPUs.

• Each instance type has a weight that represents its compute power.

• Elastigroup divides the target capacity by the selected instance weight.

• If the result isn’t a whole number, Elastigroup rounds up to maintain enough capacity.

• Elastigroup chooses instance types based on availability, pricing, and durability.

Weighted capacity is ideal for heterogeneous clusters and workloads where raw compute is more important than instance count.

Why weighted capacity shouldn’t be used in stateful Elastigroups

Weighted capacity becomes problematic when persistent state is involved.

Target capacity behaves differently

In stateful Elastigroups, the target capacity represents the number of stateful nodes (SSIs) — not the number of active instances.
This makes weighted calculations unreliable.

Example:
If you set the target to 5, and you have 5 SSIs with weighted instance types, the target is recalculated based on the total weight rather than the number of SSIs.

This causes the group to go out of sync because the capacity no longer matches the actual number of stateful nodes.

Why this becomes an issue

• Stateful groups depend on persistent volumes and network interfaces.

• Weighted capacity recalculates target capacity dynamically.

• Those recalculations don’t align with stateful resource retention rules.

Because of this mismatch, weighted capacity "works" technically but leads to unexpected behavior and is not best practice.

Recommendation

Weighted capacity is supported but not recommended for stateful Elastigroups.

Flexera recommends using weighted capacity only for non-stateful groups, where scaling is based on compute power rather than persistent resources.