# Replica autoscaling: min, max, trigger

The maximum is a ceiling. Replica autoscaling bills the copies that are running. The minimum sets availability and the cost floor.

Canonical URL: https://unbroker.cloud/blog/replica-autoscaling
Published: 2026-11-02
Reviewed: 2026-11-02
Author: Unbroker

## The short answer

Replica autoscaling adds and removes copies of a service as load moves. The maximum is a safety ceiling. You pay for the replicas that are actually running. The minimum is how many stay up, and it is the floor of the bill.

The common mistake is a high maximum with a minimum of 1. The ceiling does not keep a second copy running. [Pricing](https://unbroker.cloud/pricing) is where the rates live, and [how capacity-based billing works](https://unbroker.cloud/blog/usage-vs-capacity-billing) is the longer note on the meter.

## 1. Minimum: availability and the cost floor

The minimum is the fewest replicas the service keeps. Billing follows the autoscaler minute by minute, and the count never drops below that minimum, so this number is also what you pay for when the service is quiet.

Several replicas are spread out, so one failure does not take the app down while more than one is running. That is horizontal capacity you already have, not a copy that appears after a spike. A minimum of 1 means the quiet state is a single replica. A high maximum does not change that.

An extra replica counts once it is ready to serve. A short utilization spike does not raise the price of a replica that was already running. [Set replicas and resources from the Scale tab](https://unbroker.cloud/changelog/2026-06-24-scale-replicas-and-resources).

## 2. Maximum: a ceiling, not a bill

The maximum is the most replicas autoscaling will start. You are not billed for the gap between what is running and that ceiling. The Scale tab shows a monthly estimate as a range while autoscaling is on.

Your plan sets how far a service can scale, and the dashboard shows that limit next to the maximum. On the shared cell a service can run up to 3 replicas. Dedicated cells allow more. There is no separate published ceiling for those cells on this page.

Scheduled windows that hold extra replicas ready at a known time exist on supported dedicated cells. They are not available on the shared cell. [Autoscale on CPU, memory, or both](https://unbroker.cloud/changelog/2026-10-06-autoscale-cpu-memory).

## 3. Trigger: CPU, memory, or both

Each trigger is a share of that replica's limit, averaged across the replicas that are running. You can scale on CPU, on memory, or on both. At least one trigger stays on. When both are on, the one that asks for more replicas wins.

CPU is on by default, with a 70% target. Memory is off by default. When you turn memory on, the target starts at 80%. Both targets can be set from 10% to 95%. Memory that the app does not release after a spike, such as a cache or a heap that only grows, keeps the extra replicas up. Leave memory off for that kind of app.

New replicas start within about 30 seconds of sustained load. Extra replicas are removed after about 5 minutes of low load, so a short dip does not drop capacity. Those two timings are not a single rule of the form "CPU above a percent for five minutes." Alert windows on the same CPU and memory charts are a different control. [Logs, metrics and alerts](https://unbroker.cloud/blog/logs-metrics-alerts) is where those charts and alerts are described.

The [services guide](https://docs.unbroker.cloud/apps/services#scaling-replicas) is the source for the defaults. [Plans](https://docs.unbroker.cloud/billing/plans) say replica autoscaling is not on Free.

## Horizontal and vertical scaling

Horizontal scaling changes the replica count. Vertical scaling changes the CPU and memory each replica gets. Replica autoscaling is the horizontal one. The size of each replica stays the size you set on the Scale tab.

A busier replica does not cost more while its configuration and the running count stay the same. Starting another replica adds that replica's configured CPU and memory for the time it runs. The rates are on [pricing](https://unbroker.cloud/pricing).

## What cannot scale this way

Repository and upload services can run several replicas. Image services, workers, and any service with a persistent volume run one replica.

Free runs one replica, and replica autoscaling is not available. Pro scales up to the shared-cell limit. Replicas do not share memory. Put sessions and other shared state in a database or another shared store before you run more than one.

A scale change is saved immediately and takes effect on the next deploy.

## Three decisions before you turn it on

1. Set the minimum to the number of replicas you want up when the service is quiet. That number is the availability floor and the cost floor.
2. Set the maximum as a ceiling you are willing to reach, inside the limit the dashboard shows next to it.
3. Leave CPU on at 70% unless you have a reason to change it. Turn memory on only if the app releases memory after a spike. The starting target is 80%.

## FAQ

### What is replica autoscaling?

It adds and removes copies of a service as load moves, between a minimum and a maximum you set. You are billed for the replicas that are running, not for the maximum. It requires Pro, and it is not available on Free.

### What is the difference between horizontal and vertical scaling?

Horizontal scaling changes how many replicas run. Vertical scaling changes the CPU and memory of each replica. Replica autoscaling is horizontal. The size of each replica stays what you configured, and a utilization spike does not change that line.

### Do I pay for the maximum number of replicas?

No. You pay for the replicas that are actually running, between the minimum and the maximum. An extra replica counts once it is ready to serve, and the count never drops below the minimum. The Scale tab shows the monthly estimate as a range. Rates are on [pricing](https://unbroker.cloud/pricing).

### What does the minimum replica count do?

It is the fewest copies that stay up, and the floor of the bill. Several replicas are spread out, so one failure does not take the app down while more than one is running. A minimum of 1 can leave you on a single copy.

### Should replica autoscaling follow CPU or memory?

CPU is on by default, at 70% of the per-replica limit, averaged across replicas. Memory is off until you turn it on, and the target then starts at 80%. When both are on, the one that asks for more replicas wins. Leave memory off when the app's memory only grows.

### Does a high maximum with a minimum of 1 keep the app available?

No. The maximum is only a ceiling. With a minimum of 1 the service can sit on one replica until load rises. The spare copies are not running, and you are not paying for them, until the trigger adds them.

### Where are the replica autoscaling limits written?

On the shared cell a service can run up to 3 replicas. Dedicated cells allow more. The dashboard shows your plan's limit next to the maximum. The [services guide](https://docs.unbroker.cloud/apps/services#scaling-replicas) is the source for the timing and the default targets.

## Related reading

- [What Free and Pro include](https://unbroker.cloud/pricing)
- [How capacity-based billing works](https://unbroker.cloud/blog/usage-vs-capacity-billing)
- [Autoscale on CPU, memory, or both](https://unbroker.cloud/changelog/2026-10-06-autoscale-cpu-memory)
- [Replicas and resources on the Scale tab](https://unbroker.cloud/changelog/2026-06-24-scale-replicas-and-resources)
- [VPS vs PaaS](https://unbroker.cloud/blog/vps-vs-paas)
