Are Amazon's Provisioned IOPS a hard limit or guaranteed minimum?

Question

Are Provisioned IOPS a hard limit? Could IO spikes exceed the number of provisioned IOPS?

I couldn't easily find a definitive answer on the question. While I pretty much assume that it's enforced as a hard limit the documentation isn't very conclusive.

Answer 1

According to published benchmarking results and some own testing it isn't possible to exceed the number of provisioned IOPS. This very much suggests a hard limit.

UPDATE:

I've just seen an RDS instance with 4000 provisioned IOPS getting around 4400 IOPS for about 20 minutes while concurrently importing dumps (few GB). Thus provisioned IOPS are not a hard limit. But still, although it's possible to exceed the provisioned limit it's probably not a good idea to expect that.