# Configuring Neon autoscaling correctly

## The trap
Agents set minimum 0.25 CU and maximum 56 CU and assume the database will smoothly use all of it. In reality the autoscaling range cannot exceed 8 CU (max minus min), and scaling follows actual load, not the configured ceiling.

## The rule
1. Open the compute's Edit drawer and set the autoscaling range, e.g. min 1 CU, max 4 CU.
2. The max permitted range is 8 CU: with min 1 CU, max can be at most 9 CU.
3. You can set project-level autoscaling defaults so new computes (new branches, read replicas) inherit sane ranges instead of the default.
4. Scale-to-zero is configured in the same drawer; the two compose (scale to zero when idle, autoscale when busy).

## Checklist
- Min should cover your baseline; max should cover your realistic peak, within 8 CU of min.
- 1 CU = 4 GB RAM; size from memory, not vibes.
- After changing the range, monitor with the `neon_utils.num_cpus()` function to see how the autoscaler actually sizes the compute under load.
- Autoscaling reacts to sustained load; a one-second spike will not move it, which is correct behavior, not a bug.