I'm working on an approach where the decrypted DEK only lives for the lifetime of a transaction, this means hitting the kms on every transaction that uses keys. It will be slower, but the time the decrypted key stays in memory would be minimized.
Watch out for KMS api quotas if you go that route. Their docs don't state what the default quotas are, so you have to go to your quotas page in the console to find out, but they likely aren't very high and might well be exceeded by the transaction rate on even a relatively small db instance.
Thanks for pointing that out, it's true that it's a limited route with cloud KMS. If you control the device like a Zymkey in a secure enclosure, the cost is minimal, although the key derivation rate is very slow.