Autoscaling is configured and nothing scales

Symptom

Load rises and the replica count does not move — or it never comes back down when the load goes away. No error is reported, because nothing failed: a scaling decision that is not taken is not an error.

Diagnose

1. Is the policy enabled? A scaling block that exists and is not enabled is inert. Read it back from the stored deployment rather than from the document you sent.
2. Do the bounds leave room? A minimum equal to the maximum disables scaling as effectively as switching it off, and looks configured.
3. Is the metric arriving? Scaling reads metrics; a deployment whose metrics endpoint is unreachable gives the scaler nothing to act on. Check the deployment’s metrics in the dashboard.
4. Is the cooldown still running? After each change there is a quiet period, deliberately, so a spike does not produce a sawtooth. A decision withheld during cooldown is the system working.
5. Was a manual scale applied? A hand-set replica count and an automatic policy are two controls on one number.

Resolve

• Not enabled, or bounds equal: correct the block. The fields and their defaults are in the deployment reference.
• No metric: make the metric reachable, or scale on one that is.
• Cooldown: wait it out before changing anything else; two changes inside one cooldown teach you nothing about either.

Prevent

Choose bounds from the traffic you have measured, not from what looks safe, and leave the cooldowns at their defaults unless you have a reason with a number in it.