I should be more clear - cycling the Datomic process is one thing that can cause this issue. Anything that disrupts the Lambda-Datomic Node connection will likely do so. I suspect there are a number of Lambda timeout-related issues that can cause a similar issue.
If you’re unable to implement a retry strategy at the client-side of this interaction, I’d suggest either doing what you did (move to an instance that runs your client) or use the (upcoming) non-lambda connection that will allow API gateway to call directly to the Datomic Node.