Autoscaling before the spike, not after it
JournalTorus
What a normal Tuesday teaches a model
Written by Ravi Achterberg, Staff Engineer, Coil Scaler
Thursday, 23 October 2026
Classic autoscaling waits for a server to get busy and then adds another. By the time the new machine is warm, the spike has already hurt. Coil Scaler starts from a different question: what usually happens next?
Traffic has a rhythm
Most services follow the same shape every week: a morning climb, a lunch dip, a Monday surge after a release. Once a model has seen a few weeks of that rhythm, the next hour is far less of a surprise than it looks on a dashboard.
Scaling as a forecast
Coil Scaler adds capacity a few minutes before the curve says it will be needed and removes it as soon as the curve turns. When reality breaks the pattern, it falls back to the classic rules, so the worst case is never worse than before.
What to measure first
- Requests per second per service
- Time from scale decision to ready
- Cost of idle capacity per hour
- Error rate during the first minute of a spike
Key takeaways
- Warm-up time is the real enemy
- Patterns beat thresholds for most traffic
- Keep the old rules as a safety net
What a normal Tuesday teaches a model.
