@Kishore Nallan after we remove the "feature" that was consuming all the resources, our cluster is stable. I'm still going to understand the reason for that, but it's probably because it was an "empty" query which as translated by our middleware to call Typesense with some weird parameters. Several requests in sequence shot the memory to the sky. It's stable now. At the end, this scenario help us to protect better our Typesense against future attacks.