We run it on Heroku on production quality dynos. Heroku came out to be a bottle neck. Moving things to lambda is an option we will seriously consider. Checking for, Twitter as well.

What happened then? It was able to serve the site but not do API calls, so no mem-cache in front of your database? The Redis mem-cache didn't flush itself when full?

The site is served via AWS CloudFront. It is SPA app hosted on S3. The API is on Heroku. No mem-cache anywhere. To search we use ElasticSearch hosted on AWS. We bump up the specs to speed it up but that was not the core problem.

I suspect Heroku Postgres DB. We didn't used some Hobby version and I think it is the real bottle neck + some unnecessary middleware for the API that also requires DB.

maybe a mem-cache then?

DB's have limited connections and are often billed as such, the mem-cache might help if there are any repeatable things. It might not be useful for your solution though since everyone searches for their own things all the time.

Increasing DB plan did the job.